Eve library

Presentation Center: exhaustive continuation and completion checklist

Tracker

Revision 1

On this page

Historical task source. The Eve task board holds current task state.

Arete humane engagement delivery, 2026-09-09: the 43-slide guide follows one partial day through the canonical check-in contract, three streak vocabularies, recovery, consumer and mobile writes, weekly review, cross-domain links and Living Offering. Four current product crops accompany records, comparisons, maps and a worked trace. The package includes a 43-page PDF, 2,654 words of speaker notes and 1,314.730 seconds (21.9 minutes) of narration. Twenty-two complete V1/features/domain-arete.md units are newly bound through 50 evidence records; with the existing check-in unit, that source now has all 23 meaningful units reviewed and three structural headings. Global coverage is 298 reviewed, 10,018 structural and 103,028 unassigned units. The integrated library has 88 guides, 1,322 unique slides, 1,345 guide-slide instances and 642 cited source entries. The receipt retains the source, product-route, guide, center, PDF, narration and focused owning-test evidence. A canonical consumer write for the recovery evidence spine, shared policy vocabulary, connected mobile POST, deployed-service acceptance, live native-device execution and human listening/comprehension remain open. The full V1 audit remains active. Counts below describe historical checkpoints.

Tara consumer continuity delivery, 2026-09-09: the 34-slide guide follows an interrupted sit across the two Oshun web players, completion projection, local checkpoints, reset and sleep controls, accessibility, breath haptics, favorites, collection, share, download, journal and recovery boundaries. It includes a complete PDF, 3,218 words of speaker notes and 35.2 minutes of narration. Ten complete V1/features/domain-tara.md units are newly bound through 53 evidence records; that source now has all 28 meaningful units reviewed and two structural headings. Global coverage is 171 reviewed, 10,018 structural and 103,155 unassigned units. The integrated library has 85 guides, 1,184 unique slides, 1,207 guide-slide instances and 569 cited source entries. The receipt retains the source, browser, PDF, narration and focused owning-test evidence. Cross-device transport, the phone-to-watch product call, durable reflection writes, live native-device execution and human listening/comprehension remain open. The full V1 audit remains active. Counts below describe historical checkpoints.

Living Scenes integrated delivery, 2026-09-09: the Creation sequence now moves from runtime into composition, the complete technique catalog, compatibility and admission, Compose Assist, governance, and finally the actual consumer surfaces. The combined library has 84 guides, 1,150 unique slides, 1,173 guide-slide instances and 537 cited source paths. The 44-slide technique package and 47-slide consumer package retain their independent source, browser, PDF and narration evidence. Their shared 1,059-slide baseline and narration entries are unchanged; the two addition sets are disjoint. Integration restored the omitted living-consumer-secrets recording and tightened the customer-card headline so its product view fits the compact desktop reader. The integration receipt binds the combined route, preservation and final checks. Full V1, native-device and human acceptance remain open. Counts in the delivery checkpoints below are historical.

Living Scenes consumer delivery checkpoint, 2026-09-09: the 47-slide guide follows the actual customer card and fixture compose workspace into keep, sharing, tenant policy, rendering, durable media, public playback, accessibility and distribution. Five real UI captures accompany source-shaped relationships, gates and disposition matrices. The PDF and 45.6 minutes of new narration are delivered; all 1,059 recordings and all 81 guide HTML files from the integrated remote revision remain exact. At this standalone checkpoint the library had 82 guides, 1,106 unique slides, 1,129 guide-slide instances and 534 cited source entries. Forty complete source units are newly bound through 61 teaching records, bringing global coverage to 150 reviewed, 10,018 structural and 103,176 unassigned units. The guide credits the server-owned policy, render-admission, durable-media and APNG-download repairs while retaining explicit score, privacy, accessibility, caller and release gaps. Its receipt is complete for this increment. The full V1 audit, remaining creation systems, native-device and human acceptance remain open. Counts below describe historical checkpoints.

Technique delivery checkpoint · 9 September 2026: what reached the slides separates the delivered Eve fixes from the unfinished whole-library design work. Eve retains 25 guides and 217 slides. Living Scenes now adds a 25-slide technique catalog and a 19-slide compatibility guide between the composition workflow and Compose Assist. They teach all twelve technique records, all nine primitive contracts, continuity scorecards, four policy stops, seven scoring dimensions, the complete template matrix and the limits of the current callers. At this standalone checkpoint the library contained 83 guides, 1,103 unique slides, 1,126 guide-slide instances and 507 cited source paths. See the technique delivery evidence. Full V1, whole-library design and human acceptance remain open. Counts below are historical checkpoints.

Living Scenes correction delivery · 9 September 2026: the 16-slide runtime guide now includes five corrected explanations and three adjacent detail slides, with updated PDF and eight regenerated recordings. It shows validator differences, cue priority, actual scheduling states and reconnect limits. The verification receipt records the scoped source, visual, browser and narration evidence. The library has 79 guides and 1,028 unique slides; no complete handbook unit is newly assigned. Full V1, broader Living Scenes and human acceptance remain open. Earlier counts below are historical checkpoints.

Tara Mentor Presence delivery, 2026-09-09: the 43-slide guide explains the synthetic cast, score, settings, arrivals, posture, choreography, cadence, release dossiers, rendered media, private storage, preferences, fallback, disclosure and analytics. Three real component crops accompany relationships, records, decision diagrams and timelines. The PDF and 62.8 minutes of new narration are delivered; all 982 prior slides and recordings remain intact. The library now has 79 guides, 1,025 unique slides and 1,048 guide-slide instances. Twelve complete source sections are newly bound to teaching evidence; Tara has 18 reviewed, 2 structural and 10 unassigned units. The guide distinguishes the implemented tonal render from the intended embodied experience and records observed product gaps. Its receipt is complete for this increment. The full V1 audit, Living Scenes consumer audit, native-device and human acceptance remain open. Counts below describe historical checkpoints.

Teacher and ritual integration, 2026-09-09: the teacher guide — since merged into chapter five of the Tara track — and incoming ritual assembly/tone guide are now combined: 52 slides and 53.9 minutes of narration beyond the shared 930-slide baseline. The library has 78 guides, 982 unique slides, 1,005 guide-slide instances and 466 cited paths, all with complete narration. The integration receipt preserves both deliveries and records the one corrected ritual-table overflow. Tara has 17 reviewed, 2 structural and 11 unassigned handbook units. Mentor Presence, Living Scenes, the remaining V1 audits and human/native acceptance remain open in the full V1 workstream. Counts below describe historical checkpoints.

Tara teacher delivery, 2026-09-09: the new twenty-two-slide guide — since merged into chapter five of the Tara track — explains the canonical teacher record and the actual directory: identity, lineage, roles, locales, rights, voice, search, generated reviews, availability, content links and recovery. Two real component crops accompany relationships, records and decision diagrams. It includes a PDF and 27.7 minutes of narration. The receipt preserves all 930 earlier slides and recordings. The library now has 77 guides, 952 unique slides, 975 guide-slide instances and 457 cited paths. The complete Teachers handbook unit is newly reviewed; Tara has 15 reviewed, 2 structural and 13 unassigned units. Seeded profiles, generated reviews and weekday displays are explicitly qualified. Mentor Presence, Living Scenes, connected service evidence and human acceptance remain open within the full V1 audit. Counts below describe historical checkpoints.

Tara ritual, reflection and companion integration, 2026-09-09: the ritual assembly and tone guide joins the incoming reflection and companion guides on current main. The library has 77 guides, 960 unique slides, 983 guide-slide instances and 460 cited paths, all with complete narration. The integrated Tara handbook has sixteen reviewed, two structural and twelve unassigned units; global reviewed coverage is 94. The integration receipt binds all three histories, current generated artifacts, narration preservation and focused browser/center/PDF evidence. Full V1 work remains active; delivery counts below are historical checkpoints.

Tara companion delivery, 2026-09-09: the new twenty-eight-slide guide explains all six offers, their selection and attribution, Metis reflection-topic analysis, Veritas/Sophia evidence labels and Assistant context transfer. It includes a complete PDF and 34.5 minutes of narration. The receipt preserves all 902 earlier slides and recordings. The library now has 76 guides, 930 unique slides, 953 guide-slide instances and 451 cited paths. One complete handbook unit is newly reviewed; Tara is at 14 reviewed, 2 structural and 14 unassigned units. Native Metis note wiring, destination verification, Iris persistence and human acceptance retain explicit qualifications. The full V1 audit remains active. Counts below are historical.

Combined Tara delivery, 2026-09-09: reflection and continuation and the incoming duration, context and lineage guide are integrated. Together they add 48 slides and 55.1 minutes of narration to the 854-slide baseline. The current library has 75 guides, 902 unique slides, 925 guide-slide instances and 429 cited paths. The integration receipt binds both authored deliveries, their unchanged recordings and all 48 reviewed PDF pages. The full V1 audit remains active; subsequent counts below describe historical checkpoints.

Tara reflection delivery, 2026-09-09: the new twenty-four-slide guide explains the native route, all five prompt families and fifteen starters, Save/Skip/Edit/Done, completion projection and the next Continue action. Relationship, transition and gate diagrams make identity and state changes explicit. It includes a complete PDF and 30.3 minutes of narration. The receipt binds source cases, rendered review and exact preservation of all 854 earlier slides and recordings. The library now has 74 guides, 878 unique slides, 901 guide-slide instances and 426 cited paths. The guide characterizes pinned behavior; native lifecycle, durable note recovery, human acceptance and the remaining full V1 work stay open. Counts below are historical.

Tara context and offline-audio integration, 2026-09-09: the new duration, context and lineage guide joins the three audio guides and five selection-taxonomy assignments on current main. The library has 74 guides, 878 unique slides, 901 guide-slide instances and 427 cited paths, all with complete narration. All 854 baseline scripts, manifest entries and MP3 bytes remain exact. Three new source-unit assignments advance the Tara handbook to thirteen reviewed, two structural and fifteen unassigned units; global reviewed coverage is 91. The delivery receipt retains 14 source cases, 172 browser checks, focused center/PDF evidence and the unresolved application-consumer boundary. Full V1 work remains active.

Tara offline audio delivery, 2026-09-09: the new twenty-slide guide follows enrollment, checksums, quality and media fallback, recent retention, budgets and clearing. It has a complete PDF and 25.2 minutes of narration. The receipt records exact source cases, rendered-page review and preservation of all 834 earlier slides and recordings. The library now contains 73 guides, 854 unique slides, 877 guide-slide instances and 424 cited paths. This teaches the pinned implementation and its limits; native-device, reflection and other full V1 work remain open. Counts below are historical.

Tara audio and selection integration, 2026-09-09: the library now combines 72 guides, 834 unique slides, 857 guide-slide instances and 420 cited paths, all with complete narration. The 42 audio pages and the five incoming Tara taxonomy assignments are retained together. The integration receipt records the current center review, exact export-pixel comparisons and unchanged recordings. Delivery counts below are historical checkpoints. The full V1 program remains active.

Tara audio delivery, 2026-09-08: shared audio state and native playback add 42 slides, complete PDFs and 48.5 minutes of narration. The pair explains legal phase transitions, interruption, listening time, markers, native launch ordering, controls and recovery. All 792 prior slides/scripts/recordings and 70 prior guide HTML files remain exact. The library now has 72 guides, 834 unique slides, 857 guide-slide instances and 420 cited paths. The receipt retains scoped source, visual, accessibility, PDF and audio evidence. Earlier counts below are historical. The full V1 program and native-device, storage, reflection and companion acceptance remain open.

Tara selection source-unit assignment, 2026-09-08: five complete taxonomy units now bind to seventeen exact pages in the existing Tara selection guide: the canonical catalog boundary, mood, theme, modality and cadence sections. The source-scoped inventory advances to ten reviewed, two structural and eighteen unassigned units; global reviewed coverage advances from 83 to 88. The library remains at 70 guides, 792 unique slides, 815 guide-slide instances, 416 cited paths and 792 unchanged recordings. The assignment receipt records the exact unit/heading/fingerprint regression, 316 browser checks, focused center/PDF verification and current limits. Full V1 work remains active.

Tara selection and complete Eve operations integration, 2026-09-08: the current library has 70 guides, 792 unique slides, 815 guide-slide instances and 416 cited paths, all with complete narration. The two Tara guides remain unchanged and Eve operations now ends with its five-document reading map. The final integration receipt records current review and exact reuse. Counts below are historical checkpoints; full V1 work remains active.

Tara selection and Eve release integration, 2026-09-08: the current library has 70 guides, 791 unique slides, 814 guide-slide instances and 416 cited paths. Both new Tara guides and the incoming Eve release-evidence page are complete. Every guide HTML file and all 791 recordings retain their owning bytes. The integration receipt records current center/PDF checks and exact reuse. Earlier counts below are historical checkpoints. Full V1 work remains active.

Tara selection and handoff delivery, 2026-09-08: two linked guides add 44 slides: moods, themes and modalities and Tara–Lilith practice handoff. The complete feature catalogs, constraint cases, type priority and thirteen rule summaries now have readable references, record specimens, relationships and branches. New narration totals 58.5 minutes, with complete PDFs. All 746 earlier slides and recordings and 68 earlier guide HTML files remain exact. The library now has 70 guides, 790 unique slides, 813 guide-slide instances and 415 cited paths. The receipt retains source, visual, accessibility, PDF and audio evidence. Earlier counts below are historical; full V1 work remains active.

Eve operations source completion, 2026-09-08: the library contains 68 guides, 748 unique slides, 771 guide-slide instances and 414 cited paths. The new closing page maps all five documents linked from the operations handbook to their distinct safety/privacy, model-quality, platform-evidence, platform-control and continuing SMX-maintenance roles. The source-scoped report now has all ten meaningful units reviewed, no errors and complete: true. The 111.911-second local narration, complete twenty-page guide PDF and dated verification receipt carry the delivery evidence. The global exhaustive-teaching backlog remains open; this increment does not establish live control health, a passing release gate or a completed maintenance run.

Eve release-procedure teaching, 2026-09-08: the merged library contains 68 guides, 747 unique slides, 770 guide-slide instances and 414 cited paths. One new operations page preserves the standard stage, family, journey and Docs Center checks, all five high-risk suite additions, and the separate roles of the current Eve Deep Test and Polish closure and ongoing Eve SMX maintenance loop. That source now has nine reviewed and one open unit. The dated verification receipt records exact source, visual, narration, PDF and browser evidence. This increment does not establish that a particular change passed or was released.

Tara discovery delivery, 2026-09-08: the new twenty-four-slide Tara service discovery, saved items and launch guide follows recommendation reads, configured search pools, card identity, bridge ordering, favorite requests, complete/partial reconciliation, page walks and launch precedence. Native diagrams and worked records accompany 31.6 minutes of narration and a complete PDF. All 722 previous slides and recordings and 67 earlier guide HTML files remain exact. The library now has 68 guides, 746 unique slides, 769 guide-slide instances and 413 cited paths. The receipt retains 26 source cases, 174 final visual checks and focused center/PDF evidence. Earlier counts below are historical checkpoints. Full V1 work remains active.

Tara web search delivery, 2026-09-08: the new sixteen-slide Tara web search and result navigation guide explains local example content, debounce, synonyms, weighted matching, grouped results, highlights, spelling, recent history, storage failure and navigation. Two readable source captures and native diagrams accompany 20.0 minutes of narration and a complete PDF. All 706 previous slides and recordings and 66 earlier guide HTML files remain exact. The current library has 67 guides, 722 unique slides, 745 guide-slide instances and 405 cited paths. The receipt retains 117 final visual checks, nineteen source-component cases and complete focused center/PDF evidence. Earlier counts below describe historical checkpoints. Full V1 work remains active.

Tara catalog and Eve service-level integration, 2026-09-08: the current library contains 66 guides, 706 unique slides, 729 guide-slide instances and 403 cited paths. It includes the complete 22-slide Tara catalog guide and the four incoming Eve service-level slides. All 706 recordings, both manuscripts and every guide HTML file retain their owning delivery bytes. The integration receipt records the current center/PDF checks and precisely reused evidence. Earlier counts below describe their individual delivery checkpoints. Full V1 work remains active.

Tara catalog delivery, 2026-09-08: the new 22-slide Tara practice catalog and records guide follows recommendations into linked practice, content and teacher records, with taxonomy, duration, rights, accessibility, cadence and contract examples. Native diagrams, record specimens and comparisons add 29.0 minutes of narration and a complete PDF. Final review passes 160 visual checks and the complete focused center/PDF checks. All 680 previous slides and recordings and 65 earlier guide HTML files remain exact. Before integration, this delivery contained 66 guides, 702 unique slides, 725 guide-slide instances and 403 cited paths. The receipt records actual review and its limits. Full V1 work remains active: recommendation/search/launch consumers, full taxonomy selection, native callers, companions, mentors, workbenches and other V1 workstreams still need their owning explanation. Human reader/listening acceptance is unperformed.

Eve service-level teaching after Tara web-practice integration, 2026-09-08: the Eve delivery contained 65 guides, 684 unique slides, 707 guide-slide instances and 399 cited paths. Four new operations slides preserve every required latency, routing, effect, confirmation, grounding, safety, provider, continuity, client, evaluation, drift and outcome-cost measure, then bind alerts to a runbook and affected plane. That source now has eight reviewed and two open units. The dated verification receipt records exact source, visual, narration, PDF and browser evidence. This increment does not establish live thresholds or measured service levels.

Tara web practice delivery, 2026-09-08: the new 21-slide Tara practice and the web sitting player guide follows the actual course rows, player controls, timer, text/audio boundaries, start and completion records, reflection storage and later readback. Three focused source-component captures and native diagrams add 27.9 minutes of narration and a complete PDF. Five slides received composition polish; 153 final visual checks and the complete focused center/PDF checks pass. All 659 earlier slides and recordings and 64 earlier guide HTML files are preserved. The library now has 65 guides, 680 unique slides, 703 guide-slide instances and 399 cited paths. The receipt records exactly what was reviewed. Full V1 work remains active, including Tara catalog/recommendation, native callers, companions, mentors and workbenches. Human reader/listening acceptance remains unperformed.

Tara scheduling delivery, 2026-09-08: the new fifteen-slide Tara scheduling and reminders guide follows an evening return through invitation timing, timezone resolution, quiet hours, reminder construction, storage queries and reschedule identity. Worked clocks, decision diagrams and record specimens add 22.3 minutes of narration and a complete PDF. The final review corrected a narrow-screen heading overflow; 111 visual checks and the complete focused center/PDF checks pass. All 644 earlier recordings and 63 earlier guide HTML files are preserved. The library now has 64 guides, 659 unique slides, 682 guide-slide instances and 386 cited paths. The receipt records the exact scope. Full V1 work remains active, including Tara practice/player, companion, mentor and workbench depth. Human reader and normal-speed listening acceptance remain unperformed.

Tara session delivery, 2026-09-08: the new 21-slide Tara ritual sessions and recovery guide explains lifecycle history, completion thresholds, continuation, skip, recovery, projection identity and saved-return evidence through native diagrams and worked cases. It adds 32.2 minutes of narration and a complete PDF. Two older Tara explanations and their tracks are corrected; all 621 other upstream recordings and 61 other guide HTML files remain unchanged. The merged library now has 63 guides, 644 unique slides, 667 guide-slide instances and 386 cited paths. The receipt records 223 visual checks, complete focused center/PDF checks and four reviewed feature units. Full V1 work remains active, including Tara practice selection, player, scheduling, companions, mentors and workbench depth. Human reader and normal-speed listening acceptance remain unperformed.

Eve operational-dependency checkpoint before Tara integration, 2026-09-08: the library contains 62 guides, 623 unique slides, 646 guide-slide instances and 378 cited paths. Three new operations slides preserve every dependency, operational concern and required degraded behavior in the pinned nine-row dependency map. That source now has seven reviewed and three open units. The dated verification receipt records exact source, visual, narration, PDF and browser evidence. This increment does not certify live dependency health or close the exhaustive teaching backlog.

Offline continuity delivery, 2026-09-08: the new twenty-slide Offline reading, updates and queued work guide follows worker activation, reading updates, cache selection, explicit background actions, library retries/replay and session cleanup. Native diagrams keep the two queue implementations and their failure behavior distinct. It adds 30.3 minutes of narration and a complete PDF; all 600 earlier recordings and 61 earlier guide HTML files are preserved. The library now has 62 guides, 620 unique slides, 643 guide-slide instances and 378 cited paths. The offline receipt records 146 visual/accessibility checks, complete focused center/PDF checks, pinned source characterization and four reviewed source-unit assignments. The full V1 program remains active; Tara and the other room guides are next. Human reader and normal-speed listening acceptance remain unperformed.

Account guide delivery, 2026-09-08: the new seventeen-slide Account, preferences and personal controls guide follows a concrete preference change through editors, payloads, persistence and specialist readbacks. It adds memory, voice, safety, notifications, Telegram, data-rights, connection and billing depth, with 27.0 minutes of new narration and a complete PDF. All 580 earlier recordings and all 60 earlier guide HTML files are preserved. The library now contains 61 guides, 597 unique slides, 620 guide-slide instances and 365 cited paths. The account receipt records the 125-check visual/accessibility matrix, complete focused center checks, exact PDF text and all seven reviewed account-feature units. The full V1 program remains active; offline and room depth are next. Human reader and normal-speed listening acceptance remain unperformed.

V1 foundation revision, 2026-09-08: the twelve-slide Lilith web/PWA guide now has a concrete result case, four distinct launch rooms, labeled shell/account and release relationships, and both update choices. Six affected PDFs and the new recording are delivered; all earlier recordings are preserved. The parallel Eve configuration update is integrated, bringing the library to 60 guides, 580 unique slides, 603 guide-slide instances and 348 cited paths. The foundation receipt records actual review and verification. The full V1 audit remains active; this increment does not close the detailed room, account, workbench, platform or source-unit teaching work.

Original checkpoint: 2026-09-06, machine transfer requested by the user. Execution-plan clarification: 2026-09-07. Historical receipts remain scoped to their recorded artifacts; determine current work using section 0 before following an older resume instruction.

This is the continuation contract for all presentation work discussed in the conversation. The user explicitly requested committing the partially completed work to origin main so another machine can continue. This checkpoint is not a declaration that the presentation project, narration rebuild, comprehensive coverage, Eve integration, Domains audit or diagram review is finished.

Read this file, the checkpoint state, EXPANSION.md, DESIGN.md and repository AGENTS.md before continuing. Keep this checklist current with actual evidence and commit and push each completed task. Do not stop after a convenient intermediate phase.

Design adoption update, 2026-09-06: the user requires the complete content-led composition overhaul in section 10. Replace the habitual four-point openings and horizontal lists with structures chosen for the ideas being taught, while preserving the Oshun visual language. Apply this during authoring and renderer development as well as the final review, across existing guides, new guides and Eve-generated presentations. Recording these requirements does not mean the redesign has been implemented or verified.

Visual craft and comprehension update, 2026-09-06: the user also requires all eight additional quality standards in section 10: a concrete visual benchmark, typography and optical alignment, quantitative visualization, visual continuity, image art direction, readability at actual viewing sizes, representative-reader comprehension and editorial subtraction. Establish the benchmark, typography rules and reader-review method with the pilot; apply the complete standard throughout authoring, migration, Eve generation and final verification. This update records required work; it does not claim completed design implementation or reader research.

Combined integration, 2026-09-08: the Eve capability delivery below and the parallel V2 rollback-session guide are preserved together: 60 guides, 578 unique slides, 601 guide-slide instances and 347 cited paths. Use the merge receipt for current catalog verification; the two parent receipts retain their dated scope. This integration does not close the broader acceptance items below.

Eve capability delivery update, 2026-09-07: the scoped Eve product/workbench package adds twelve guides, 94 unique slides and 7,949.704 seconds of narration. The Use Eve route teaches product graphs, coverage/impact analysis, work tracking, discussions/decisions, delivery, operator/studio workflows, member audits and shared member tools. Its explicit crosswalk covers the pinned catalogs of 90 assistant tools, seven executing-agent MCP operations and 31 studio views. The full Eve learning path now contains 25 guides and 204 slides; the architecture route remains separately available. That checkpoint had 59 guides, 564 unique slides and 587 guide-slide instances. All 47 earlier guide HTML files and 470 earlier recordings are preserved exactly.

The delivery receipt records the complete affected-guide browser passes, center checks, 189 route checks, 94-page PDF text check, full narration integrity check, selected visual reviews and source/crosswalk/freshness gates. The scoped breadth gap is addressed; the whole-repository source-unit assignment program, live Eve integration, external benchmark acceptance and human reader/listening work below remain open. This increment does not relabel those separate project requirements as complete.

0. Execution contract for the continuing agent#

This section makes the requirements below executable. Updating this plan does not implement the presentation project or close any delivery checkbox. When asked to execute the project, continue through the remaining stages; when asked only to improve this plan, verify and publish the documentation change without claiming the planned implementation has been delivered.

Establish the current state before choosing work#

  1. Read applicable AGENTS.md, this contract, sections 11–12, README.md, EXPANSION.md, DESIGN.md, and the authoring records and latest verification receipts for the affected guides. Open frontend-skill before planning meaningful UI work. Preserve its composition, content and interaction briefs in authoring records; apply its app guidance to the presentation workspace and section 10's teaching requirements to slides.
  2. Inspect git status, the active branch/worktrees and running jobs; fetch origin main and review incoming changes. Preserve other sessions' edits, indexes and processes. Use an isolated worktree when they overlap the task. Follow AGENTS.md resource checks before costly work, including dependency installation, narration, inventory collection and browser suites.
  3. Distinguish the implementation commit, the declared source revisions and the artifact hashes verified by each report. They are different identities. Read meta.json, authored slide revision declarations, coverage.json and inventory metadata; account for explicitly declared supplemental snapshots as well as the baseline pin. Fetch required Git objects if absent. Never substitute moving main for a declared source.
  4. Establish a dated baseline from actual catalog, content, coverage and narration outputs using applicable section 11 freshness checks. Record guide count, unique slides, reused guide-slide instances, current/stale/missing audio, PDF coverage and teaching states with report paths and hashes. A JSON count without its applicable freshness check is provisional. Preserve failures and inspect source inputs before rebuilding; regeneration alone does not explain a discrepancy.
  5. Reconcile that baseline with this checklist. Sections 2–5 contain historical counts and completed increments; checked environment/setup tasks must still be confirmed on a new host. Do not restart the original 316-track queue, recreate an integrated draft or assume an earlier test total describes the current library. Recheck invalidated work and remaining acceptance.
  6. Select the first unmet dependency in the stage table and record a concrete work package. Fix broken current behavior within its scope before adding dependent content. Missing human review or live-provider access leaves that acceptance item open; continue authorized work that does not require it.

Stage dependencies and exit criteria#

The stages describe project progress, not guide-count targets. A passing guide increment may be committed while its project stage remains open. Final visual and listening passes follow integration, but authoring, design, accessibility and narration quality apply from the first new slide.

Stage Required inputs and work Evidence required to exit
A. Current baseline and narration (§§3–5) Reconcile artifacts and receipts; repair stale/missing narration and existing delivery defects. Current builder/audio and affected browser/PDF checks pass; remaining human listening and content findings are tracked.
B. Authoring and design foundation (§10) Review pilot candidates, complete missing treatments, record briefs, benchmark, typography and reader protocol, and implement shared contracts. Source discovery may continue under host limits. Required pilot treatments and viewing formats pass applicable automated and visual checks; required reader findings are resolved before benchmark acceptance. Candidate status stays explicit meanwhile.
C. Complete teaching and organization (§§5–7) Apply B's authoring contract to new/migrated guides; cover the full repository, V1–V10, Docs Center, generation systems, all eleven tracks and UI/UX requirements. Every meaningful source unit and required track has substantive reviewed teaching; hierarchy/prerequisites/durations and current artifacts pass; no arbitrary count cap or unreviewed duplicate exemptions.
D. Real Eve integration (§8) Begin delivery after comprehensive content coverage; discover the actual host/runtime and reuse its authority, persistence and execution contracts. Live question → persisted note → admitted edit → regenerated result works, plus all four ad hoc request families and artifact lifecycle; affected contracts, evaluations and end-to-end gates pass.
E. Exhaustive Domains audit (§9) Reconcile registry, directories, code/planned roots and navigation against the final declared source boundary. Every domain has an accurate home and teaching coverage; discovery/source-of-truth checks pass; resulting additions return through C's authoring and verification steps.
F. Complete library quality review (§10) Review final guide sequences and every reused slide context after integration. Every-slide diagram/image and every-guide eight-standard rubrics, actual-size browser/PDF reviews and normal-speed listening are current; all material findings and required reader retests are resolved.
G. Release and publication (§§11–12) Freeze candidate inputs, run applicable final gates and reconcile the evidence manifest. Exact delivered artifacts have passing gates/reviews; final source freshness is assessed, docs agree, normal hooks pass, and the task branch tip is present on origin/main.

Source review and a documented pilot candidate can progress while a reader session is pending. Do not label the benchmark accepted, inherit nonexistent reader findings or close the corresponding stage. Reopen earlier stages when later changes invalidate their exit evidence.

Define each work package before editing#

Use one coherent guide, mechanism, shared behavior or audit slice as the unit of work. Reuse authoring/, drafts/, assignments/ and existing review ledgers; avoid a competing catalog or coverage database. Give the package a stable ID, link the precise checklist item/heading, and fill in these fields with actual values. Unknowns need a discovery action, not an invented answer.

Field Required detail
Purpose and scope Audience, prerequisites, learning outcome or observable behavior; guide/slide/source-unit IDs; boundaries and remaining related work.
Baseline and dependencies Implementation commit, baseline/supplemental source revisions, input hashes, prerequisite work and findings; current versus historical evidence.
Source review Exact files/sections/symbols and owning code/tests; claims, exceptions and implementation status; resolve conflicting docs/code or label the discrepancy.
Deliverables Canonical files to edit; generated HTML/audio/PDF/catalog/map outputs affected; assignments, briefs and receipts to update.
Design and story Guide art direction, narrative arc, slide jobs, source-supported relationships, worked example, transitions, applicable failure/recovery case and benchmark treatments.
Acceptance Observable pass/fail conditions for the whole explanation or flow, edge cases and relevant negative cases; distinguish automated, model, visual, listening and human-reader evidence.
Verification Exact commands/modes, environment/fixtures, browsers/viewports, existing performance/readability limits, report locations and expected scope. Discover scripts' actual arguments before use.
Recovery How to stop owned jobs, resume partial generation, restore a known-good revision and prevent incomplete output from being advertised as ready.
Closeout Status, reviewer type/date, findings/retests, verified hashes, commit/publication evidence and next unmet dependency.

Reference shared rules instead of copying the whole project checklist into each package. Prefer completing a reviewable explanation through publication over accumulating unverified manuscripts. An intentionally scoped source-review or tooling package can finish independently; its guide-delivery item stays open.

Execute a guide or change through its whole lifecycle#

  1. Read and reconcile sources. Inspect complete relevant units and their contracts, callers, tests and failure paths. A search hit or copied excerpt is discovery, not review. Separate documented intent, implemented behavior, fixture execution, actual integration and acceptance. Explain contradictions and limitations wherever they affect the learner.
  2. Write the learning map and briefs. Establish the canonical guide home, prerequisites, reusable slides and gaps before assigning IDs. Walk one case from input through mechanism to outcome; explain why it works and where it fails. Cover every meaningful part of a unit before calling it reviewed. Split overload rather than shrinking text or hiding teaching in citations.
  3. Edit canonical inputs. Author in content/*.json, diagrams.json, catalog-source.json, meta.json and actual shared renderer/CSS/JS as appropriate. Keep unpublished work in drafts/ until delivery is ready. Preserve stable IDs and foundation coverage; trace reuse, prerequisites, excerpts and references through add/remove/reorder changes. Follow the current builder's revision/schema contracts.
  4. Review a rendered candidate. Build the actual draft/guide and run focused contracts and Playwright. Inspect standalone and embedded desktop/compact/ mobile views, real zoom, print/PDF and source/transcript access as applicable. Record findings against slide IDs. Resolve source/story/layout defects before producing large narration batches.
  5. Generate deliverables. Use canonical tools for HTML, scripts, catalogs, maps, narration and PDFs; do not repair generated bytes/metadata manually. Render missing/stale audio with the supported sequential renderer, verify encoded output, then rebuild duration-dependent artifacts. Check reviewed source/slide fingerprints when creating assignments; content, graph and inherited-scope changes can invalidate them.
  6. Verify and review the delivered result. Run the matrix below against final artifacts. Inspect actual captures/PDFs and listen at normal speed as required by the acceptance item. Include every affected reused context. A draft render or synthetic audio test does not close narrated-guide acceptance.
  7. Resolve findings and repeat invalidated checks. Keep failures, record causes/fixes, regenerate affected outputs and repeat affected checks/reviews. Do not weaken assertions, update expectations blindly, omit failing specimens or change completeness definitions to obtain green.
  8. Publish and continue. Update assignments, relevant docs and this checklist with verified scope; review the diff and stage only the package's changes. Commit with normal hooks, push to origin main, integrate remote changes and retry on rejection. Verify publication, then proceed to the next requirement when executing the broader project.

Select verification by what changed#

Section 11 remains the final release gate. This matrix selects checks for an increment; it does not permit skipping affected checks or replacing final whole-library verification with samples. Inspect scripts/configs for newer applicable suites. Run expensive targets sequentially, with resource admission and supervised logs under AGENTS.md; never launch a broad workspace sweep for a prose-only edit.

Change Required automated verification and review
Planning prose only Scoped Prettier, git diff --check, valid new/changed local links and section references, executable-path/flag review, workflow consistency and normal hooks. Rebuild/check generated documentation actually affected. UI/runtime suites apply if their inputs or behavior also change.
Guide content, diagrams, assignments or catalog Builder freshness; applicable Python/Node presentation contracts; source/assignment and inventory-report freshness; invalidated audio checks; full affected standalone guides and affected center navigation/audio/PDF paths; source, story, visual and listening review for the claimed scope.
Shared renderer, player, navigation, CSS, fonts or export Relevant unit/contracts plus full standalone-bank and center Playwright; applicable composition, table, source-artifact, short-viewport, actual-browser-zoom and print regressions; axe, keyboard/focus, offline/error/audio paths and existing performance budgets. Cover affected layouts/reused contexts and add missing regressions for changed behavior.
Inventory, source revision or assignment rules Collector/parser/coverage regressions, appropriate pinned recollection/freshness, workflow/repository inventory gates, supplemental-source validation, foundation retention and assignment invalidation/review. Record final source-boundary deltas; regenerate dependent explanations/evidence.
Narration text, pronunciation, timing or model settings Speech/narration/resource tests, actual full fingerprint/hash/duration/decode/pause checks, regenerated duration labels, affected real browser playback/race/failure paths and normal-speed listening of every changed track. A shared policy change may invalidate the entire bank.
Eve API, persistence, authority or generation Owning app/service lint/typecheck/build/unit/integration/schema/migration/evaluation/release gates; deep Playwright through actual flows and applicable native mobile automation; separate live/mocked cases. Exercise retries, stale revisions, cancellation, denied scope and partial generation/persistence failures.
Native/product execution or claims about it Existing targeted product contract/native/service checks and guide source/teaching tests. Follow local UE instructions in AGENTS.md; record exact source/host patches, fixtures and commands. An adapted-host pass cannot close unchanged-source, full-game or live-network acceptance.

Existing focused browser entry points include composition-pilot.mjs, center-short-viewports.mjs, table-browser-zoom.mjs, table-print-regression.mjs and guide-specific scripts under tools/presentations/tests/. These supplement eve-oshun-deck.mjs and presentation-center.mjs; select affected assertions, not just similarly named scripts. Record browser/version, viewport and zoom separately; CSS scaling or viewport emulation does not prove actual browser zoom.

Inspect semantic diffs before accepting generated expectations or snapshots. Retain applicable negative controls and important failure paths. Capture observed process exits; logging through a pipe must preserve the test process's exit, not just the logging command's. Interrupted, running, skipped and failed checks are not passes. An expected incomplete teaching gate can document project gaps for a valid increment; it cannot close comprehensive coverage.

Evidence, invalidation and honest completion#

Use a dated verification/<package>-<date>/ receipt, extending an existing one when it represents the same work. Store a concise README plus machine-readable results where tooling provides them. Include commands/flags and observed exits, tool/model/browser versions, fixture/live status, source revisions, tested input/output hashes, ordered slide IDs/reused contexts, capture/PDF/audio references, reviewer type, unresolved findings and retests. Keep secrets and private user content out. Preserve failed attempts and prior evidence with its original scope. Large reproducible captures may remain outside Git when their location, hashes and reproduction follow the existing artifact policy.

Use planned, in progress, blocked, verified for stated scope and published in package receipts. Keep this file's [ ]/[x] conventions: check an item only when its entire deliverable and applicable acceptance are satisfied and its commit is published. Split broad items into completed technical increments and explicit open quality items when necessary. Candidate benchmarks, mechanically current audio, source review and full guide acceptance are distinct outcomes. Non-applicability requires a specific reason; a missing required reviewer/provider/platform is a blocker, not an exemption.

Before reusing evidence, trace changed inputs through their consumers:

  • Source/revision changes invalidate affected claims, excerpts and assignments; reassess their slides, narration, diagrams and source-based tests.
  • Slide text, graph, inherited chapter/release or membership changes invalidate applicable assembled hashes, teaching review and reused contexts. Narration text/policy changes also invalidate speech and duration evidence.
  • Renderer/CSS/font/image/data changes invalidate affected browser/PDF and visual/actual-size reviews even if slide text hashes are unchanged. Material meaning/treatment/audience changes also trigger section 10's reader rules.
  • Catalog/hierarchy changes invalidate prerequisites, navigation/search/history, ancestor durations and guide-context evidence. Eve schema/runtime/persistence changes invalidate dependent contracts, live flows and evaluations.

Record changed inputs → affected artifacts/checks in the receipt and rerun that scope. Inherited passes require unchanged relevant inputs, output hashes and environment assumptions, plus a valid baseline allowed by the harness. After a merge or regeneration, verify that delivered bytes still match the evidence; rerun affected checks when they do not. Do not rebuild after the final gate and publish changed outputs under the earlier passing report.

For a blocker, record the failing command or missing input, attempted remedy, missing coverage, impact and exact next step. Complete feasible setup work and continue independent authorized requirements. Never fabricate participants, listening, native execution or real Eve receipts. Arrange required human evidence through the permitted workflow and leave it pending when no reader is available; do not contact other people without the user's authorization.

Record each published package's commit and confirm it is on fetched origin/main, with no local-only task commits. Since the publishing commit cannot contain its own hash, record tested input hashes/base commit in its receipt and report the resulting commit after publication. On handoff, include the last passing scope, open findings, owned stopped/active jobs and logs, and the next runnable action. A handoff or published increment does not close the broader project requirements in sections 1 or 11.

Design edition binding (user decision, 2026-09-11)#

The Eve redesign started on 9 September 2026 because the library's default slide system produced walls of tables, one repeated skeleton, no covers or closes, and almost no captures. The accepted replacement is the Eve edition: designEdition: "eve" on every slide, the chapter chrome, the four-actor vocabulary, the validated layout repertoire in tools/presentations/presentation_eve_edition.py, the slide and chapter definitions of done, and the voice rules, all recorded in EVE-V1-REDESIGN-TODOS.md and the brief authoring/eve-two-track-edition.md. The accepted specimen is chapter 1, Meet Eve.

Binding rule. From 11 September 2026, every slide produced through any section of this file (sections 5, 6, 8, 9, 10 and 12 included) is authored in the Eve edition. No new slide may use the legacy skeleton or the legacy layouts (table, evidence-comparison, flow, columns, layers, gated-flow, record-anatomy, case-study, relation-map, ownership-handoffs, recovery-map, product-view) as its primary visual; when a closed set genuinely needs exact values, it uses the edition's compare-panel or stat-panel and counts against the one-table-in-four budget. Where an older paragraph in this file describes how to author or compose slides (section 10's content-led composition overhaul, section 5's per-guide acceptance criteria, section 6's track recipes), the redesign file governs on any conflict and those paragraphs remain only as review criteria. Do not re-establish a "design pilot"; it exists. Do not restart the old renderer defaults for speed.

Measured on 11 September 2026: 76 content files, 1,470 slides, 521 in the edition. 23 files are fully migrated and 53 have no edition slide at all. The legacy layout mix in those 53 files is dominated by evidence-comparison (195), table (154), flow (117), columns (112) and gated-flow (93). Nine of the 53 were delivered on 9 September or later in the old system (43-tara-companions, 46-living-scenes-composition, 46-living-scenes-consumers, 46-tara-consumer-continuity, 47-living-scenes-assist, 48-living-scenes-technique-catalog, 49-living-scenes-compatibility, 69-arete-humane-engagement and the V1 sophia/iris depth pairs); their delivery paragraphs at the top of this file are historical and their guides await migration under the redesign file's Phase D tracks like every other legacy guide.

  • Moved 2026-09-18: the builder guard is A3.G1 of EVE-V1-REDESIGN-TODOS.md, in the phase that precedes chapter work.
  • Moved 2026-09-18: the chapter-budget guard is A3.G2 of the same file.
  • Name, in the redesign file's Phase D list, every track this file plans beyond Eve and V1, so one file stays the plan of record for migration: the eleven cross-cutting tracks of section 6, the V2–V10 product guides (40-v2-* through 44-v2-* today), and the whole-monorepo, generation/ComfyUI and Docs Center space tracks. For each: a track id, its guides today, and the order it is migrated in. Verify: every guide id in catalog-source.json appears under exactly one Phase D track or in the redesign file's list of guides that retire, checked by a short script whose output is pasted under the list.
  • Rule, not a task (un-boxed 2026-09-18): each of those tracks gets the redesign file's D0 setup (track declared in catalog-source.json, track brief under authoring/, glossary slide, route tests remapped, old deck files retired, route walk at desktop and mobile) before its first chapter is written.
  • Rewrite the delivery-paragraph convention at the top of this file: every delivery from now on states the edition, the layout counts (table-like share, diagram count, capture count) and the chapter receipt path, so a regression to the old system is visible in the paragraph itself. Verify: the convention paragraph at the top of this file names those three things, and the next delivery paragraph written after it carries them.
  • Covered elsewhere (un-boxed 2026-09-18): section 8 moved to Phase 0 of the redesign file, where EI.8.01 validates every proposed slide with validate_eve_edition and the A3 guards before anything is written. The other half, that section 12's "read the docs" and "see it live" controls render inside the chapter chrome rather than as a new footer element, is an acceptance line of 12.6 and 12.7 below.
  • Re-review the nine post-9-September legacy guides named above against the chapter definition of done and schedule them explicitly in the redesign file's tracks (Tara · continue, Tara · native and offline, Creation, Shared systems, The other rooms); do not treat their 9–11 September receipts as satisfying the redesign. Verify: each of the nine guide ids appears under a named Phase D track of the redesign file with a migration item of its own; today none of them is mentioned there (checked 2026-09-18).
  • Update section 13's resume order, README.md, DESIGN.md and authoring/current-design-status.md so the first instruction to a continuing agent is the binding rule above and the redesign file's milestone order, not the pre-redesign composition contract. 2026-09-18: section 13 already says this; README.md, DESIGN.md and authoring/current-design-status.md do not mention the redesign file at all, and they are what is left. Verify: each of the three opens with the binding rule and a link to EVE-V1-REDESIGN-TODOS.md.

1. Entire user intent and required sequence#

  • Preserve the original requested format: an extensive, branded, source-linked HTML presentation with voice narration about all of Eve and Oshun V1.0, V1.1 and V1.2, modeled on the earlier Kojo/TDAO reference.
  • Replace the long public reading route with a Presentation Center: collapsible left navigation with topics/subtopics, selected presentation filling the right workspace, focused standalone guides and individual PDFs.
  • Enrich the first Eve/V1 decomposition instead of merely slicing the old deck: the completed first phase had 35 guides, 349 unique slides, 349 recordings and 35 PDFs. Its 244 foundation slides were retained and 105 substantive slides added.
  • Expand to all information from V1–V10 and all information in Docs Center, including authored pages, generated references and relevant supporting sources.
  • Extend that scope to every aspect and corner of the entire monorepo: repository layout, all products/domains/packages, generation and agentic systems, software architecture, granular implementation, infrastructure, CI/CD, operations, testing and supporting tooling. The Docs Center inventory is a starting point, not the final coverage boundary.
  • Explain all generation systems, including ComfyUI workflows and agentic workflows, from approachable system overviews through detailed execution paths, node/tool contracts, implementation, failure handling and operations.
  • Complete all eleven cross-cutting presentation tracks in section 6: product purpose/journeys; identity/permissions/trust; data/content lifecycles; non-generation algorithms; money/access/resource economics; architectural decisions/evolution; platform extension; operational case studies; governance/human workflows; role-specific learning/practice; and the Presentation Center itself. Give each explicit coverage and review evidence.
  • Determine the final guide count from actual learning outcomes and complete source coverage. 35, 38, or any arbitrary small number is not a cap.
  • Make each guide understandable, suitably enriched and appropriately sized; organize and sequence guides with prerequisites, canonical homes, related links and as many hierarchy levels as the content needs.
  • Show reliable narration length for every presentation, topic, subtopic and higher/lower ancestor. Keep incomplete audio visibly incomplete.
  • First, not last (owner decision, 18 September 2026): integrate actual Eve into the center for grounded questions, persisted notes and real presentation/slide changes. The work is Phase 0 of EVE-V1-REDESIGN-TODOS.md (items EI.0.01–EI.11.06); this box closes when that phase's receipt exists.
  • Enable Eve to generate and narrate ad hoc presentations from a conversation: tradeoffs, deep research, competitor analysis and idea exploration. Support temporary drafts, searchable artifact history and saved project/collection presentations. The work is section 0.13 of Phase 0 in EVE-V1-REDESIGN-TODOS.md (EI.12.01–EI.12.11).
  • Retain the requested Oshun / OSHUN / Ọṣun = “oh-shuhn” pronunciation.
  • Complete the global pacing correction: headings, subtitles, sentences, explanatory paragraphs and takeaways must not blend unnaturally.
  • After the preceding work, meticulously ensure every domain appears under Domains, with complete source/code/planned-root coverage and appropriate guides.
  • Finally review every slide and add as many useful diagrams or additional diagram slides as reasonably clarify the ideas. The user explicitly requested this final pass; do not limit it to the few new slides or existing diagrams.
  • Fully adopt section 10's content-led composition overhaul: meaningful openings, varied compositions, expressive minimalist elements and accurate flowcharts. Complete the representative pilot, shared authoring/renderer changes, whole-library migration and regression checks. Four-column openings and horizontal arrows must no longer be the default explanation for unrelated ideas; preserve legitimate four-part concepts when the content warrants them.
  • Fully adopt section 10's visual craft and comprehension standards. Complete all eight areas, including concrete reference examples, precise implementation rules, actual-size browser/PDF evidence and real reader feedback for the pilot and material new treatments. Extend the per-guide rubric and Eve evaluations; no generic polish statement, automated score or attractive full-resolution capture substitutes for these checks.
  • Add relevant screenshots and UI mockups, including additional slides wherever needed to explain actual interfaces, proposed designs and user journeys.
  • Create extensive UI/UX design presentations for appropriate product, domain and shared-interface sections: show designs, mockups, alternatives, decisions and modern design-system composition from atoms to complete pages.
  • Review every presentation from beginning to end for exceptionally high quality: natural, human-like narration, a coherent story, clear transitions, thoughtful pacing and consistent organization throughout the entire library.
  • Finish all applicable verification and publish the resulting commits to origin main. A partial inventory or passing subset of tests is not completion.

The original reference was a local directory on the first machine: /Users/elikemagudogo/Desktop/workspace/tokenized-assets/docs/presentations/eli5/kojo-tdao-architecture-and-flows/. It may be unavailable after transfer. The implemented design, content, narration, source excerpts and tests in this repository preserve the chosen approach; do not block continuation just because that reference directory is not mounted.

2. What is saved, and what is not yet complete#

The counts and receipts in this section describe recorded increments. Additional guides or source snapshots may have arrived since those runs, including through another branch. Reconcile them with the fetched implementation and fresh checks under section 0; the word "current" inside an older checkpoint does not override newer artifact evidence.

The current catalog has 48 guides, 484 unique slides, 507 guide-slide instances, 311 cited source files and 20 legacy native SVG diagrams. All 48 PDFs are exported and all 484 recordings have current fingerprints. The rollback-session integration receipt records the fourteen-slide delivery, complete automated gates, 25 passing native cases in the explicitly adapted host and two whole-section teaching assignments. The teaching report has 40 reviewed units, 10,018 structural units and 103,286 unassigned units; the exhaustive project remains incomplete. CSS flow slides also carry diagrams, so the SVG count is not the total visual-flow count. The initial semantic recurrence treatment adds its own native SVG return connector; it is not an additional legacy diagrams.json entry.

The newest published guide entries are:

Guide Slides State at handoff
portfolio-map 12 Full standalone browser/PDF/actual-audio checks passed; two new diagrams reviewed
v2-product-promise 9 Draft desktop/compact/mobile/axe/PDF checks passed; narration pending
v2-product-surfaces 11, including one reused slide Draft desktop/compact/mobile/axe/PDF checks passed; narration pending

The later companion authority review adds a twelfth surface-guide slide and corrects its read-only/authentication wording against the pinned OpenAPI and developer reference. The historical handoff row above retains its original count.

V1 and V2 now share a Products branch. Delivery and release remain last in the reading order. The 38-guide hierarchy passed the complete smoke checks in --draft mode, including all guide selections, nested navigation, search, prerequisites, history, collapse persistence and the mobile drawer. This is not the full all-slide/all-audio portal release test.

Narration at the stopped checkpoint:

  • 64 tracks current, totaling 4,944.153 seconds of unique audio; 773 measured sentence segments are recorded in those manifest entries.
  • 316 pending: 297 existing recordings have stale speech fingerprints, and 19 V2 recordings do not exist yet.
  • The example orientation-boundaries is corrected. Its title/subtitle quiet gap measured approximately 0.965 seconds, and subtitle/notes 0.799 seconds.
  • Existing stale files are preserved for resumability and history. The builder excludes them from current playback and completed listening-time totals.
  • The owning narrator was stopped with its supervised termination path; do not copy or reuse old process IDs. No owned narration job should be left running as a requirement for this handoff to work.

The exact current/stale/missing ID lists and content hashes are saved in verification/handoff-2026-09-06/state.json. Recompute them after any edit.

Historical evidence boundaries#

  • Completed Eve/V1 phase commit: fe8137e13cf649788304379e75e3f0179564f33d.
  • Completed inventory foundation commit: 12feae7e7c.
  • Existing verification/summary.json, presentation-center.json, standalone-bank.json and current-bank-controls.json describe earlier runs. Keep their scope honest; global speech changes mean their HTML/audio hashes are not proof for the final expanded library.
  • The historical source correction changed the heading of eve-context-failure to “Fluent answers can still use the wrong evidence”; preserve that correction.
  • The newly copied three PDFs passed sequence/title/footer checks. Full current portal export checks and final visual PDF review are still required.

3. Resume on the next machine#

Destination progress is recorded in the Linux resume verification. The next V2 manuscript and narration checkpoint records 74 current tracks and the isolated 16-slide draft checks. The historical handoff counts above remain the baseline until narration finishes. The companion follow-up records the next source correction, two reviewed teaching assignments and additional narration progress. The Nyx and contracts visual follow-up records 89 current tracks, Nyx's full standalone actual-audio check and the complete isolated sixteen-slide contracts visual/editorial review. The Nyx learning map records further source-review findings without claiming additional teaching assignments. The Nisaba follow-up reaches 96 current tracks and records its source-review learning map, including the feature page's omitted study-adapter package and distinct notebook paths. The opening Eve follow-up reaches 116 current tracks and adds five reviewed teaching assignments, including three corrected explanations. It makes orientation and all six Eve guides fully narrated without declaring the global rebuild or listening review complete. Its seven full standalone suites cover 74 slides; the reused status explanation also passes the twelve-slide release-acceptance draft suite. The complete 38-guide center run covers all 407 embedded instances and 38 PDFs, with actual audio boundaries for ten complete guides and 814 hash-verified viewport captures. Only 37 images across five unique slides and six guide contexts are claimed in this increment's targeted visual review, not the final library pass. The member follow-up reaches 125 current tracks and nine reviewed teaching units. It makes V1 web/PWA, Tara and Arete fully narrated, corrects the Tara clamping exception and Arete's recorded-miss explanation, and retains additional source findings in the member learning map. Global narration and normal-speed listening remain unfinished. Three full standalone suites cover 30 slide instances, and the complete center run covers all 407 instances and 38 PDFs, with 26 desktop/mobile audio boundaries for 13 ready guides and 814 hash-verified captures. The targeted visual ledger reviews 13 images across the two corrected slides, not the complete library. All 54 Python and 22 Node tests pass; the teaching-completeness gate remains an expected failure.

The shared architecture follow-up reaches 145 current tracks and 11 reviewed teaching units. Both architecture and communication/discovery are fully narrated. Four corrected explanations and their learning map distinguish schema validation from structural guarding, event publication from consumer effects, and the curation package inventory from persisted platform behavior. Two full standalone suites cover 21 instances; the complete center run covers 407 instances and 38 PDFs, with 30 desktop/mobile audio boundaries for 15 ready guides and 814 hash-bound captures. The targeted visual ledger reviews 25 images across four slides, not the final library pass. All 54 Python and 37 Node tests pass. Global narration, normal-speed listening and teaching completeness remain unfinished.

The earlier Sophia narration follow-up is superseded for guide depth by the forty-slide Sophia delivery. The expanded route covers every meaningful unit in both pinned V1 Sophia pages: live answer and source-scoring paths, complete envelope and enum boundaries, adapter roles and unavailable states, source-set readiness and trace, real but unwired BM25 and verification engines, specification-only ingestion and publication, invalidation, evaluation and consumer evidence. Its 32 new tracks add 42.6 minutes; the complete guide runs 56.5 minutes. Psyche/Lilith, wider repository/consumer acceptance and normal-speed human listening remain unfinished. The Iris completion immediately below supersedes the earlier eleven-slide follow-up.

The Iris follow-up reaches 162 current tracks and 21 reviewed teaching units. Two new slides separate the ranker and inspection mechanisms, making Iris an eleven-slide narrated guide. Seven existing Iris explanations and the shared Eve memory explanation now distinguish write planning from recall, correct decay semantics and explain consent, tombstone and inspection integration boundaries. The learning map retains the historical review and points to the completed forty-four-slide depth delivery. Six earlier and 49 new assignments now cover every meaningful unit in the two pinned Iris pages; the source-level Iris depth gap is closed while product integration remains qualified. Both full non-draft standalone suites pass across 20 guide-slide instances. The complete center run covers 409 instances and 38 PDFs, with 34 desktop/mobile audio boundaries for 17 ready guides and 818 hash-bound captures. The targeted visual ledger reviews 69 images across ten changed slides in eleven guide contexts, including the shared Eve slide in both guides. All 54 Python and 82 Node tests pass. Global narration, normal-speed listening, full source teaching, actual consumer integration and final library visual acceptance remain open.

The subsequent source and composition work expands Psyche/Lilith to fifteen slides and adds five reviewed, hash-bound architecture units. The Psyche learning map distinguishes event construction/append/validation, turn outcomes, configured latency targets, overload instructions, fixed-priority downgrade and upgrade plans, raw versus wrapped Lilith safety methods, and consumer execution. Forty-seven new Psyche module tests and five isolated Lilith adapter tests pass. They are controlled module evidence, not live authenticated/media or complete detector acceptance.

The first two section 10 pilot treatments are implemented: independent question/evidence rows for v1-web-pwa-boundary, and optional activity/skip/return semantics for shared member-day. Explicit contracts reject unknown layouts, invalid endpoints and compulsory chains between optional peers. Nineteen new composition/audit tests pass. The complete eleven-slide web/PWA suite passes with current audio; the fifteen-slide Psyche suite passes in draft mode. The larger pilot, benchmark collection, typography standard, full library migration and actual reader evidence remain open. Canonical art direction and the neutral reader protocol preserve that scope.

At the initial composition checkpoint, audio reached 167 tracks / 14,773.101 seconds / 2,276 sentence segments, all decoded and pause-verified. 220 remain pending: 196 stale and 24 missing, including the four new Psyche tracks. The two pilot explanations and corrected Lilith policy track were recorded; Psyche still has nine pending tracks. No normal-speed listening or human comprehension pass is claimed.

The initial composition evidence binds that checkpoint's source, HTML, narration and PDF artifacts to 76 passing Python tests and 134 passing Node tests. The complete center draft/ready-audio run covers 413 instances, 826 ordered desktop/mobile captures and 38 PDFs, with 34 actual audio boundaries across 17 ready guides. The focused pilot checks cover 33 current captures, including compact/mobile scroll-to-end and a labeled font-fallback/CSS-text-enlargement case. All 24 before captures, 33 after captures and 108 guide-review images were model-inspected; no human reader or normal-speed listening evidence is implied.

The visual ledger records all eight craft areas as partial, pending or specifically inapplicable for the inspected scope. Resolve the three legacy mobile tables losing row identity, the latency-target dash's missing visible legend, the linear runtime decision overview, generic Psyche opening and misleading legacy ordinals during the remaining pilot. Only the shared member context has current Psyche PDF page-fit review; all fifteen pages have full-size inspection. Full web/PWA guide art direction, complete-guide contact sheets and willing-reader sessions remain required before benchmark promotion and library migration.

The priority-decision follow-up replaces the runtime overview with six source-ordered supplied/absent decisions, selected-signal validation and distinct valid, error and unchanged outcomes. All 192 signal-subset/current-mode cases agree with the pinned resolver. Eleven additional decision-contract tests cover branch completeness, variable condition counts, terminal roles and invalid fallthrough; the complete presentation suites now pass 88 Python and 136 Node tests, including a new complete-guide brief/order check. The fallback assignment was refreshed after source review; the totals remain 26 reviewed, 10,018 structural and 103,300 unassigned units.

Audio now reaches 176 tracks / 15,834.350 seconds / 2,439 sentence segments, all decoded and pause-verified. 211 remain pending: 191 stale and 20 missing. Psyche's fifteen tracks are current, totaling 1,868.740 seconds. Both Psyche and web/PWA pass complete non-draft standalone checks. Those checks caught a compact active-playback overflow on the new decision slide; a scoped spacing correction preserves label sizes, and the focused matrix now passes 49 captures including all four compact playing contexts. Normal-speed listening and actual reader comprehension remain unverified.

The full web/PWA guide brief now records all eleven narrative jobs, evidence anchors, relationships, emphasis, reading order and transitions. Those remaining treatment choices are plans, not implemented or accepted designs. Complete the architecture/ownership, quantitative and artifact examples, remaining full-guide migration, typography standard and real reader review before promoting a benchmark. The older table, dash-legend, opening and legacy-numbering findings remain open.

The current full center run passes all 38 guides/PDFs and 826 embedded captures, including 36 desktop/mobile playback boundaries across 18 ready guides. It is still draft/ready-audio mode, not a full narrated release pass. The new visual ledger distinguishes 56 original-detail model inspections from 87 exact-byte inherited observations. Both complete web/PWA and Psyche contact sheets were inspected; the complete web/PWA desktop sequence was also reviewed. Repeated account/design columns and implied PWA lifecycle arrows remain migration findings. The room-maturity touching-column finding is retracted below. The guide brief was corrected against effective layouts: release enforcement is already a fan-out graph and needs reviewed labels/exceptions, not replacement of an unseen manuscript fallback. Reader comprehension and full pilot acceptance remain pending.

The table-continuity follow-up retains an exact-value table for psyche-latency-targets, now with typed configured milliseconds, seven numbers, five explicitly missing targets and a visible “not configured, not zero” legend. The values map to four exports at the unchanged source pin. Narrow native tables retain their row headers while every comparison column remains reachable; numeric columns align right, and a scroll hint appears only for actual overflow. The table brief explains the choice and source scope. The reader protocol now includes neutral quantitative and mobile-identity tasks; no participant results or accepted quantitative benchmark are claimed.

The table-continuity evidence binds this increment to 102 passing Python tests, 137 passing Node tests, both complete non-draft standalone guides and the 34-capture focused table matrix. The complete center passes all 38 guides/PDFs and 826 embedded captures, with 182 desktop/mobile table-readability reports and 36 actual audio boundaries. The existing 49-capture composition regression passes with exact prior image bytes. Model review covers eight baseline, 34 final and eight full/page-fit PDF items; this is not a full-library optical or real-reader pass. All 176 current audio files, manifest and narration script retain their prior verified bytes; no new recording or normal-speed listening is claimed. Source-coverage totals remain 26 reviewed, 10,018 structural and 103,300 unassigned units. The generic opening, legacy numbering/arrows, full-guide migration and remaining pilot/craft requirements stay open.

Optical erratum: the prior room-maturity desktop spacing finding was an incorrect model observation, not a defect fixed by this change. Its new baseline capture is byte-identical to the prior saved image. Actual DOM text ranges show 195.984375px between the final row's second and third text cells, with the existing 22px right padding intact. Desktop padding was not changed. The older receipt remains historical; the separate mobile loss of row identity was reproduced and addressed.

The ownership-handoff follow-up adds a concrete Tara read detail after the retained conceptual overview. Three named code responsibilities and four labeled exchanges show the caller, shared adapter and configured API dependency without inventing deployment hops. Pinned execution checks prove concurrent read orchestration, successful composition handoffs and rejection without a partial envelope or cancellation; they do not prove downstream card models, authenticated requests or persisted effects. The ownership brief records the communication job, responsive contract and neutral reader tasks.

All 109 Python and 144 Node tests pass. The complete twelve-slide architecture standalone suite passes with actual audio, and the focused composition matrix passes 64 captures, including 15 ownership contexts/stress cases and four negative controls. The complete center passes 38 guides/PDFs, 828 embedded captures, 182 table-readability reports and 36 actual audio boundaries across 18 ready guides. It remains draft/ready-audio mode, not the final narrated gate. Model review covers the new treatment, eight changed regression captures, architecture's complete desktop sequence/contact sheet and the overview/detail PDF pair, with a separate full-size detail export. Forty-one other pilot captures retain exact prior bytes. These are scoped observations, not every-library-slide optical or real-reader evidence.

Audio now reaches 177 current tracks / 16,007.673 seconds / 2,465 sentence segments, all freshly decoded and pause-verified. The new detail is 173.323 seconds; the preceding 176 current manifest entries and files remain unchanged. 211 remain pending: 191 stale and 20 missing. Teaching totals remain 26 reviewed, 10,018 structural and 103,300 unassigned units. Ownership is another candidate, not an accepted benchmark. The architecture opening, unlabeled legacy arrows, repeated five-step compositions, annotated-artifact example, remaining full-guide migration, typography specification, normal-speed listening and willing-reader evidence remain open.

The annotated-source follow-up replaces pwa-install-update's implied install/cache/update/clear lifecycle with the exact pinned applyWaitingServiceWorker body. Three source-bound annotations distinguish the absent-worker guard, posted request and immediate local return; activation, page refresh and server acknowledgement remain separate evidence. Installation/native scope, route-sensitive reading and queue/privacy detail remain in the notes. No slide IDs, guide order or cited source counts change. The artifact brief records the communication job and neutral reader tasks.

All 117 Python and 147 Node tests pass. The complete eleven-slide web/PWA standalone suite passes with actual audio, and the focused composition matrix passes 77 captures with 11 rejected/restored negative controls. The full center passes all 38 guides/PDFs, 828 embedded captures, 182 table-readability reports, three source-artifact screen/print reports and 36 actual audio boundaries across 18 ready guides. This remains draft/ready-audio verification, not a fully narrated release pass. Model review covers eight baseline and thirteen final artifact views, 24 changed pilot regression images, the complete web/PWA desktop sequence/contact sheet and the final source PDF page at full and page-fit size. Forty other pilot captures retain exact prior bytes. Neither these observations nor automated text checks establish real-reader comprehension.

Actual narrow-screen review found awkward identifier fragments, and compact playback exposed an inherited 17px symbol style. Source-preserving wbr opportunities and the role-specific 22px symbol rule address those findings; exact characters, source ranges, provenance, font floors and clipping are checked after reflow. The replacement recording is 173.872 seconds. Fresh decode/pause checks now cover 177 tracks / 16,103.359 seconds / 2,478 sentence segments; the other 176 current entries and MP3s remain unchanged. 211 tracks remain pending: 191 stale and 20 missing. Teaching totals remain 26 reviewed, 10,018 structural and 103,300 unassigned units. This is another pilot candidate, not an accepted benchmark. Full-guide migration, additional openings, image/chart treatments, complete typography, normal-speed listening and willing-reader review remain open.

  • Fetch origin main and use a clean branch/worktree containing this handoff. Inspect local changes and active processes before starting. Do not overwrite another session's work or kill its processes.
  • Read applicable AGENTS.md files. Normal hooks, scoped commits and pushing the completed branch tip to origin/main are mandatory. Integrate remote changes and retry if a push is rejected. Do not force-push main or bypass hooks.
  • Open the mandatory frontend-skill before meaningful UI work. Also use the relevant presentation/PDF/browser skills. The old machine lacked js_repl and in-app browser tools; repository Playwright with installed Chrome was used. Use the available supported workflow on the new host and keep deep automation.
  • Do not spawn subagents unless the user or applicable instructions explicitly authorize delegation; it was not authorized for the saved work.
  • Inspect RAM, swap, disk and competing verification as required by AGENTS.md. Use targeted gates, one expensive job at a time, and resource supervision.
  • Install the repository's pinned Node dependencies using its normal setup. The dependency setup receipt records a supervised pnpm install --frozen-lockfile at the repository's pinned pnpm 10.25.0. @oshun/web resolves and executes Playwright 1.60.0 and axe-core 4.11.4; its Playwright-managed Chromium 148.0.7778.96 launches successfully. The optional system Chrome channel is absent on this host, so the saved Chrome 152.0.7977.76 runs remain historical evidence. Workspace-link warnings identify unbuilt local CLI targets and do not widen this dependency-resolution checkpoint into a built-binary claim.
  • Create a Python 3.11+ virtual environment outside tracked content, install tools/presentations/requirements.txt and requirements-narration.txt, install the English spaCy model, and provide FFmpeg, ffprobe and Poppler.
  • Use CPU narration on Linux unless a supported renderer backend is available. --device mps is only for an available Apple GPU. The current renderer exposes cpu and mps; do not assume a CUDA flag exists.
  • Confirm model/voice access or transfer the legitimate local model cache. No provider credential is needed for stock Kokoro narration. No secret was copied into the presentation tree or handoff.
  • Rebuild the ignored whole-bank fixture, then resume the pending recordings. Do not run the browser suite concurrently with the narrator you launch.

Example setup, from repository root (adjust executable names for the host):

sh
python3.11 -m venv /tmp/oshun-presentations-venv
. /tmp/oshun-presentations-venv/bin/activate
python -m pip install -r tools/presentations/requirements-narration.txt
python -m spacy download en_core_web_sm
python tools/presentations/build-eve-oshun.py --preview
python tools/presentations/render-eve-oshun-narration.py --batch-size 6 --device cpu --threads 1
python tools/presentations/build-eve-oshun.py --preview

The previous speech environment used en_core_web_sm 3.8.0. Dependency/model versions are pinned in the requirements and renderer. If those cannot install on the destination, diagnose and document the platform limitation; do not silently change voice, pronunciation, cadence or model settings while retaining old freshness claims.

The full workspace is large. On a host below 32 GiB or without swap, do not start nx run-many --all, a broad build matrix, or combined heavy gates by default. The supervisor samples memory every 30 seconds and terminates only its owned process group below the safety reserve. Narration needs 3 GiB available memory and 5 GiB free disk at admission, retains a 2 GiB running memory reserve and a 3 GiB running disk reserve, and uses fresh sequential model processes per batch. Do not treat a temporary lack of resources as evidence the feature is impossible.

4. Finish the narration correction first#

Implemented mechanism to preserve#

tools/presentations/presentation_speech.py constructs semantic blocks from title, subtitle, each notes paragraph and takeaway; explicit script overrides use blank-line paragraph boundaries. Missing final punctuation is added for speech intonation. Sentence splitting preserves versions, decimals, paths, initials, common abbreviations and quoted endings.

The renderer synthesizes whole sentences with inserted silence of 0.40 seconds between sentences, 0.85 after a title and 0.70 after other blocks. It retains short natural speech margins, a 0.18-second initial lead and a 0.55-second tail. A rare within-sentence model chunk gets a smaller 0.12-second separation.

Speech fingerprints include the actual plan, timing policy, model and voice. Manifest entries include sentence text hashes, semantic role, start/end samples and pause samples at 24 kHz. verify_pauses decodes the actual MP3 and checks the inserted silence after encoding, not just the intended values in JSON. Playback URLs include the audio SHA so a regenerated file receives a new cache entry. The visible Script panel preserves the same paragraph structure.

  • Resume all 316 pending tracks and complete the additional authored audio. The current 391-track inventory passes full encoded-audio verification: 34,841.439 unique seconds and 5,167 measured segments. The completed 388-track queue was followed by eight corrected and three new Agentic recordings; all other 380 recordings, including the original 177 current entries, retain identical metadata and file bytes. See the final renderer log and continuity result. This completes generation and mechanical audio verification, not normal-speed listening or the broader source/visual review. The following intermediate narration counts are historical. The renderer skips valid recordings, stops on failed/no-progress batches and finishes with an inventory check. Linux continuation's first manuscript checkpoint completed ten of these: 74 were current, totaling 5,673.251 seconds; 287 stale and 19 missing tracks remained. All 74 current files pass persisted hash, duration and encoded-pause verification. This is a partial checkpoint, not the full inventory gate. The companion follow-up reaches 81 current tracks / 6,145.579 seconds; 280 stale and 20 missing tracks remain, including the additional authority slide. Complete the pending handoff tracks and newly authored narration. The next follow-up reaches 89 current tracks / 6,764.814 seconds and 1,062 sentence segments, leaving 272 stale and 20 missing tracks. All nine Nyx guide tracks are current and its full standalone actual-audio checks pass; the global rebuild remains pending. Nisaba's seven remaining recordings bring the total to 96 current tracks / 7,335.062 seconds and 1,147 segments; 265 stale and 20 missing tracks remain. Completed audio does not establish source-unit teaching coverage or a normal-speed listening review. The opening Eve follow-up reaches 116 current tracks / 9,332.714 seconds and 1,460 segments, leaving 245 stale and 20 missing recordings. Twenty pending tracks and three revised explanations were rendered in four bounded CPU batches. Orientation and all six Eve guides now have complete audio; the full-library narration gate still remains pending. The member follow-up reaches 125 current tracks / 10,180.758 seconds and 1,591 segments, leaving 236 stale and 20 missing recordings. Two bounded CPU batches rendered nine pending tracks and replaced two corrected explanations. V1 web/PWA, Tara and Arete now have complete audio. The shared architecture follow-up reaches 145 current tracks / 11,898.803 seconds and 1,845 segments, leaving 216 stale and 20 missing recordings. Four bounded CPU batches render 20 previously pending tracks, including the corrected partial-failure, event-bus and curation explanations. Both shared architecture guides now have complete audio; these mechanical checks do not establish normal-speed listening quality. The Sophia follow-up reaches 153 current tracks / 12,727.069 seconds and 1,969 segments, leaving 208 stale and 20 missing recordings. Two bounded CPU batches render eight previously pending tracks, including the shared substrate context and four corrected or enriched answer explanations. Sophia now has complete audio; Iris remains incomplete. All current audio passes actual decode and encoded-pause verification, not normal-speed listening review. The Iris follow-up reaches 162 current tracks / 13,973.959 seconds and 2,161 segments, leaving 201 stale and 20 missing recordings. Two bounded CPU batches render seven previously pending tracks, two new ranker and inspection explanations, and a replacement for the corrected shared Eve memory track. Iris now has complete audio; Eve context remains complete. All 162 current tracks pass decoding and encoded-pause verification. Global narration and normal-speed listening review remain incomplete.
  • Run full render-eve-oshun-narration.py --check after completion: all files, fingerprints, byte counts, SHA-256, durations, decode and encoded pauses pass for all 391 tracks in the final renderer log linked above.
  • Verify the reported heading/subtitle sample again on the destination; inspect/listen to representative titles, questions, quoted sentences, lists, technical acronyms, longer paragraphs and actual brand pronunciation. 2026-09-18: what is left is a person listening at normal speed on the destination machine; the automated audio checks already pass. blocked:human
  • Preserve oh-shuhn, including Oshun, OSHUN, Ọṣun and decomposed Unicode. spoken() uses Misaki ˈOʃən; O represents the requested /oʊ/ vowel. Text spelling remains unchanged. Actual Misaki frontend tests already exist. Rule, not a task (2026-09-18): the pronunciation is implemented and its Misaki frontend tests exist; this line says to keep it.
  • Add regression cases if listening exposes unnatural sentence grouping, abbreviations, list fragments or lost text. Re-render affected audio whenever a speech policy/text change invalidates it; do not fake readiness metadata. Conditional follow-up, not a task (2026-09-18): it applies only if the listening review above finds a defect, and that review is a person's.
  • Rebuild HTML, scripts, catalogs and every ancestor's narration totals only after current recordings are stable. Confirm no missing audio is advertised as a complete guide/topic duration. The 391-track checkpoint passes five complete hierarchy-label audits and all 38 selected-guide and details-panel duration checks in the full center browser run.
  • Verify actual browser playback, pause/resume, seek, speed with preserved pitch, stop, auto-advance, end-of-guide, replay, guide-switch cancellation, drawer/dialog pausing, denied autoplay, delayed play promises, missing/corrupt audio, retry, HTTP range requests and portable local-file operation. The complete 391-slide standalone run and 38-guide center run pass; see the automated checkpoint.
  • Update current verification reports and README counts/durations. The canonical bank and center reports now record the complete current runs; their previous versions are preserved in the new receipt. This completes the mechanical narration checkpoint, not listening or final visual quality.

5. Complete the current content checkpoint#

The 391-slide automated checkpoint now covers the full standalone bank, all 38 embedded guides/PDFs and both V2 standalone suites. All eleven applied Agentic slides have desktop, compact, mobile and compact-playback assertions, plus embedded/PDF checks. The scoped model ledger retains 93 images, including all 17 Agentic and 21 V2 PDF pages. Source/guide acceptance and normal-speed listening below remain separate.

  • Finish verification and normal-speed listening for the three applied Agentic Studio recovery corrections before accepting that guide. Thirty-four pinned-source probes and the existing 25-test run suite confirm that checkpoint comparison covers selected fields, does not enforce the recorded resume budget, and differs from ordinary resume/fork/retry. Replay and progress helpers also have narrower boundaries than a full historical snapshot or authorization layer. The three replacements are now applied; the original dry proposal tests retain the 7167e08ce5 baseline, while separate live-content tests verify the application. The final regenerated standalone/embedded/PDF and actual-audio checks pass; complete normal-speed listening and guide review. Do not equate mechanically current audio with source acceptance. The reviewed-source assignment count is unchanged. Baseline correction: the later fingerprint membership audit establishes that none of the eight replacement IDs was current in the original 177-track baseline, including agentic-studio-failure. This supersedes the contrary sentence in the older draft/receipt; keep the original-177 preservation check intact when applying the drafts. 2026-09-18: the corrections are applied and the browser, PDF and audio checks pass; what is left is a person listening at normal speed and accepting the guide. blocked:human
  • Finish verification and normal-speed listening for the three applied invocation, grant and handoff corrections. Thirty-eight new pinned probes and 64 existing tests pass. Preflight checks supplied declarations, renewal accepts caller-provided identity/time fields, and handoff permits explicit ancestor re-entry while returning local budget calculations. The simple resolver is not the full portfolio broker; controlled broker rejection tests also pass. Four dry authoring checks bind the new slide hashes and joint six-patch catalog merge. The corrections are applied, and five further source units remain unassigned. Browser, PDF and actual-audio checks pass; finish source and normal-speed listening review. The admission, isolation, registry, modes and pipeline explanations still need their own deeper source and integration review. 2026-09-18: as above; the unassigned source units it mentions are carried by the chapter assignment items of the redesign file. blocked:human
  • Finish verification and normal-speed listening for the applied admission and execution manuscript: two replacements and three new detail slides. Thirty-one pinned runtime probes, 70 existing package tests, 25 HTTP-route tests (including three new cases) and six dry authoring checks pass. The route uses the governed runner, neither plan runner wires a token bucket, partial calls can sit under a completed run, and completion persistence can fail after an effect. Key-based telemetry redaction is not a general secret scrubber. All three correction sets are now applied: eight disjoint replacements and three insertions, giving the Agentic guide 17 slides. The old admission graph override is removed; three independent comparisons explicitly omit ordinal numbers. The application record binds exact source changes and visual refinements, and the guide brief records every slide's job and remaining review boundaries. Seven more source units remain unassigned. All 380 unaffected tracks from the completed 388-track queue, including all original 177 current tracks without an exception, are preserved. The eight corrected and three new recordings pass the full decode/pause gate. New-slide/table/flow/embedded/PDF/actual-audio checks pass; finish normal-speed listening and source/guide acceptance. These controlled examples do not establish production exactly-once execution, live durability, exhaustive Agentic coverage or reader comprehension. 2026-09-18: as above. blocked:human
  • Resolve the specific web typecheck issue recorded during the execution review. The locked html2canvas-pro 2.4.1 dependency is restored, which supplies both callback types; a fresh boolean visibility read removes the invalid narrowing without changing the opt-in or late-result guards. The screenshot follow-up records 51 web tests, 141 BFF tests, 97 V1 prompt/corpus tests and all 16 desktop/mobile browser cases with shared accessibility checks. Web, BFF and prompt-injection-evals typechecks pass, as does the direct focused compiler check for the new tests. The bounded browser heap is adjusted after a recorded Next dev restart; no baseline or dependency version was changed. Retain the failed gate evidence as historical, including its now-superseded helper hash. These scoped checks are not a full web/Eve release or Next production-build gate. The earlier receipt also records an initial harness-discovery mistake and its corrected two-suite selection; unrelated root security/persistence dependency and baseline findings are not claimed resolved by that fix.
  • Harden tools/typecheck/ratchet.mjs against pathless compiler/setup failures. The screenshot follow-up reproduced a false green: a temporary configuration outside the app emitted three global TS2688 missing-type errors, but the ratchet reported zero app-owned errors because its parser requires a file/line prefix. The raw compiler rejected that attempt; the final app-local, consumer-version-bound test configuration passes directly with tsc. Preserve the intended dependency/backlog boundary, add positive and negative diagnostic/exit-status tests, and rerun affected real gates. Do not count the discarded temporary result as verification. The compiler-gate follow-up records 40 passing regression/integration checks, all three targeted uncached web/BFF/evaluation typechecks and the direct focused test compiler pass. Fatal failures cannot update baselines; known located backlog/dependency rules remain unchanged. Cache inputs now bind the gate implementations, and OSHUN CI runs the suite. The identical suite rejects the prior implementation in 24 cases. Historical screenshot evidence retains its superseded ratchet hash; no broad release or all-consumer verification is claimed.
  • Review the three new guide source explanations against the source pin; do not describe the portfolio overview as detailed V2–V10 coverage. The source-review receipt binds all 32 unique slides (33 guide instances) to 37 exact spans across all 27 cited paths at the source revision. Every explanation is supported with recorded qualifications, including the V4 player/engine cell-count distinction, V10's concept status and ten-channel budget, and the V2 companion's scoped writes with gameplayAuthority: false. The catalog and README retain the overview-only boundary. This claim-level review creates no whole-source-unit teaching assignment and required no slide, visual, PDF or narration change.
  • Finish V2 product-promise and surface narration and run their non-draft standalone checks. Their saved draft reports deliberately omit actual audio; the new full nine- and twelve-slide reports include it.
  • Recheck the split V2 ruleset table, long API-route table and authority graph in the embedded center at desktop and mobile sizes. Seven rows overflowed the first draft; the current two-slide 4+3 decomposition passed standalone and embedded checks. The new ledger records both embedded sizes plus full-height graph/API reading captures. That checkpoint identified the narrow label defect repaired below; this is not final typography acceptance.
  • Fix the V2 companion table's narrow row label: Authentication contract wrapped as Authenticatio / n contract. The label repair uses Auth contract without changing text size, OAuth detail or narration. A 125% row-text probe found a second split word; Mobile app preserves the adjacent planned-iOS/Android qualification. The source-reviewed teaching fingerprint is refreshed. Full affected standalone, focused embedded, mobile, PDF and all-391 narration checks pass, with five rejected browser defect injections. This closes this specific repair, not full V2 typography or guide acceptance. V2 contracts integration follows below.
  • Preserve the shortened readable graph labels. The original V3 labels overflowed their nodes; detailed meaning remains in descriptions and notes. The V3 graph-label receipt locks all five compact label pairs and the fuller graph, point and note explanations. Actual SVG geometry passes in standalone and embedded desktop states, while 390px and 320px views expose the complete reading alternative. Five full-description restorations, reduced type and hidden text are rejected; the actual PDF page is visually reviewed. Git retains no verbatim pre-repair label source, so those controls are explicitly the authored full explanations rather than reconstructed historical strings. This closes one graph-label regression, not detailed V3 teaching or the broader diagram typography review.
  • Review representative V2 mobile states and all new PDF page renders, not only browser screenshots or extracted PDF text. The current V2 visual receipt binds and individually reviews all 21 current page-fit PDF renders, six named mobile states and the standard embedded ownership graph. The repaired Auth contract and Mobile app labels remain intact at 390px; the complete nine- and twelve-slide standalone guides and a 27-guide center integration run pass. This closes the current-artifact visual review, while the two acceptance tasks below stay open.
  • Remove sequence-like ordinal markers from the six independent V2 column comparisons, then repeat their standalone, embedded, mobile and PDF review. The comparison-ordinal receipt binds 24 browser captures and all six affected PDF pages. The independent comparisons retain their exact four-item content without 01–04, while all three real progressions retain those markers in standalone and embedded desktop/mobile states. An injected comparison marker and a removed progression marker are both rejected. This closes the semantic marker correction, not the separate ownership-graph or willing-reader acceptance task below.
  • Recompose and enlarge the ownership graph's standard and compact embedded detail labels. The ownership typography receipt binds the scoped 230×82-node composition, 20px/19px authored type, at least 16px effective label height in the four SVG delivery states, the ordered mobile alternative and both canonical PDF page-fit renders. Six browser captures, four injected defects, the complete affected guides and a 27-guide center integration run pass.
  • Obtain willing-reader typography and comprehension feedback on the recomposed ownership graph in standard and compact embedded views, the mobile reading alternative and PDF page fit. Record the reader, date, protocol, findings, any changes and the final retest; automated bounds and model inspection do not establish reader acceptance. 2026-09-18: needs willing readers; automated bounds and model inspection do not establish reader acceptance, as the item says. blocked:human
  • Backfill section-level assignments for the 35 completed guides and all new guides. Eighty-two reviewed, hash-bound assignments now cover two V2 surface boundaries, five opening Eve units, twenty-eight Eve detail units, one Eve product-graph authority pair, two member playback/check-in units, one partial-failure contract, one curation package boundary, four Sophia answer units, six Iris memory boundaries, five Psyche architecture units, five V2 contracts/conventions units, one complete V2 network-quality architecture unit, three V2 rollback architecture units, four offline- continuity units and four Tara session units. The other 103,244 meaningful units remain unassigned. A supplemental-revision slide can teach a primary-inventory unit only when its assignment declares that exact revision and the validator proves the complete cited source file is byte-identical in both Git snapshots. The opening learning map records three corrected explanations, two unchanged complete explanations and the remaining source-detail boundaries. The Eve invocation map adds the complete voice, tour and proactive-help boundary while preserving six partial source units. The Eve development/change map binds five compact recipes already taught by the detailed guides and records the eight meaningful units that still lack complete teaching. The Eve model-quality map adds escalation, judge-validation and outcome-economics boundaries while preserving seven units with incomplete detail. The Eve operations map adds the complete operating-state graph, nine-row operational dependency map, six-outage degradation, configuration/secrets, telemetry-schema and health/service-level and evaluation/release boundaries and preserves one partial operations unit. The Eve memory/grounding map adds continuity, memory-governance, claim-support and bounded-retrieval boundaries while preserving five incomplete units. The Eve runtime map adds the complete seven-posture failure/fallback boundary and preserves six partial runtime units. The Eve safety map adds protective precedence, typed audit and adversarial-coverage boundaries while preserving seven incomplete units. The member learning map records the next two corrections and the Tara, Arete, PWA and account detail still pending. The shared architecture map records the contract, event-bus and curation corrections, along with the architecture, transport, discovery and curation depth still required. The Sophia map now records complete meaningful-unit teaching for the two pinned V1 pages, including live scoring, adapters, lifecycle, unwired verification engines and the target architecture. Broader repository and consumer acceptance remain separate. The Iris map now records complete meaningful- unit teaching for the two pinned Iris pages, including schema, scope, retention, recall, consent, privacy/DSAR, continuity, inspection and shell integration boundaries. Durable storage, authenticated service behavior, downstream erasure and complete customer UI remain product-acceptance work. A running total, not a task (2026-09-18): assignments are made chapter by chapter, by the "re-review assignment and teaching bindings" item of each chapter in the redesign file, and the exhaustive-coverage goal is section 6 below. The figures here are the state on the date they were written.
  • Integrate the V2 contracts guide and verify its recorded delivery as described below. Final listening, reader and craft acceptance remain open.

V2 connection-quality integration#

The twelve-slide content/43-v2-network-quality.json guide adds a Netcode branch under Products/V2, with product promise and contracts as prerequisites. It preserves the frozen draft's visual fields and qualified local-code explanation; only the final manuscript-publication sentence becomes a reminder to retain the source revision and environment. Its 1,190.383 seconds of narration brings the library to 421 current tracks. All 409 prior recordings and speech scripts, 39 prior guide definitions and HTML files, and 33 prior canonical content files are unchanged. Seven new citations and their dependent inventory hashes are the only discovery projection changes; no teaching assignment is added.

The receipt binds full standalone, center and bank verification separately from the focused 168-check/ten-defect suite and frozen draft regression. Model review covers 177 images: complete standalone and actual embedded contexts, compact/scrolled playback, all delivered PDF pages at two scales, and complete-guide rhythm. Smaller embedded contexts can scroll; reachable controls are not evidence of whole-slide first-screen fit. The twenty historical adapted-host native cases retain their exact evidence and limitations; they were not rerun or broadened for presentation integration.

  • Resolve mobile score labels that split 200/ms and 40/ms, preserving exact quantities while keeping numbers with units. Extend applicable narrow/enlarged/font/zoom automation and review all invalidated contexts.
  • Resolve ordinary prose-table headers that are noticeably small at PDF page fit. Establish effective-size typography evidence rather than accepting cell bounds or a full-resolution image as sufficient.
  • Complete normal-speed listening, willing-reader tasks and retests, final craft acceptance and source-unit assignments. The integrated eight-standard ledger keeps these requirements separate from technical delivery. 2026-09-18: listening, willing-reader tasks and craft acceptance are people's; the source-unit assignments ride with the guide's migration in the redesign file. blocked:human

The shared table repair passes the current full 421-slide bank and 40-guide center, all PDF exports, typography/actual zoom, source-dialog, focused regression, tooling, audio and freshness checks. The bank's terminal exit is zero, with all twelve checks passing against the final artifact. The two concrete findings above are resolved. The receipt distinguishes exact-PNG review inheritance from the restored-edge-only PDF comparison; neither automation nor model inspection establishes willing-reader or complete craft acceptance.

  • Repair the short-viewport center navigation exposed at actual 400% zoom: retain the exact full guide title and Topics/Guide details controls in a single row; keep secondary actions and metadata in a scrollable dialog; make the entire short drawer reachable; preserve keyboard focus through CSS breakpoint transitions and guide selection. The dated receipt records all 40 guides at actual 200%/400% zoom, narrow/landscape/ordinary contexts, real download/history/audio interactions and deliberate defects. This resolves the concrete header/title/navigation issue from the table review, not whole-library typography, reader/listening or craft acceptance.

The network-quality teaching review assigns the complete “Network quality, handshake, and graceful fallback” source unit to all twelve current guide slides. It covers the whole grade table, fallback diagram, score formula and diagnostics boundary, qualifying the source's live-transition, elapsed-time, declined-offer and persistence claims against pinned C++. Reviewed documentation units now total 32; 103,294 remain unassigned. Code units, generated duplicates, other architecture sections, reader/listening evidence and the broader acceptance items remain open.

V2 rollback-session integration#

  • Deliver the fourteen-slide predicted-frame correction guide with narration, standalone HTML and PDF, source-backed worked examples and two reviewed whole-section assignments. Extend the native host from 20 to 25 executed cases, retain its fixture limitations, and verify semantic, responsive, embedded/audio, zoom, accessibility and print behavior. The receipt binds final inputs and observed exits, preserves all 470 earlier recordings and 47 standalone guides, and records the corrected editorial labels and failed verification attempts. This completes that guide's technical integration; reader/listening and comprehensive rollback acceptance remain open.

V2 contracts integration#

drafts/v2-contracts.json preserves the historical sixteen-slide proposal, unchanged from 7e3c10f43624e85b51704e20eec6c414d6348e14. It remains outside content/; current delivery uses the reviewed eighteen-slide content/42-v2-contracts.json and the prerequisite-linked v2-contracts catalog entry. All eighteen recordings, the complete center/PDF report, whole-bank suite and shared-table regressions pass. The draft covers evidence paths, reference notation and validation, target artifacts, module reconciliation, strict floating-point configuration, the worked determinism contract, cook/DDC routes, checker limits and generated gRPC descriptors. Its isolated desktop/compact/mobile/axe/navigation/source and 16-page PDF checks passed before integration. The later isolated review ledger records all sixteen draft slides and PDF pages, replacing the earlier representative-only inspection. The eighteen-slide integrated version requires its own recorded-delivery evidence; the historical draft review is not rebound to the changed guide.

The integration brief records all eighteen communication jobs and the two added comparisons: executable V1 Zod value validation versus V2 artifact checks, and a reproduced pinned gRPC freshness failure. The pinned stored descriptors contain 44 services / 666 RPCs; isolated regeneration produces 47 / 675, adding Commentary, Director and Esports while preserving every other generated C++ byte. A synthetic options-body probe also shows that later ordinary RPC declarations can be omitted while freshness passes. The guide teaches these limits without modifying product source or claiming Unreal compilation, transport or release verification.

The five assignments cover complete cross-reference, documentation-enforcement, worked-determinism, tree-enforcement and module-topology overview units. They raise reviewed teaching to 31; they do not cover the rest of the source page, the complete V2 APIs or CI declaration units. The application record binds preservation of all 391 prior slides and speech plans, 38 existing guide HTML files and 391 prior recordings. The two independent comparisons remove ordinal markers. Enlarged-text inspection also shortens two row labels and a long displayed schema name without shrinking type or changing narration.

The focused browser suite checks 82 rendered states, retains 38 captures and rejects ten DOM defects. It includes every compact active-playback state and five narrow/enlarged tables. Three exact-source tables opt into sourceWrap; optional identifier-boundary breaks preserve selectable characters and remove fragments without shrinking type. An initial enlarged-word failure and a later overwide value-column attempt were corrected before the passing run. Other guide HTML is unchanged. PDF checks preserve non-whitespace source characters and cell order after reflow.

The current eighteen-slide non-draft standalone report, complete 409-slide bank and complete 39-guide center report pass, including all 39 PDFs, 870 embedded captures and 78 guide audio boundaries. Shared-table and companion-label regressions pass with 34 and 11 captures respectively. The delivery receipt retains the reports, 157 reviewed images, final audio-integrity check and exact input/artifact hashes. Earlier label, source-wrap and interrupted attempts remain historical; a new PDF-test header-case mistake was corrected, not hidden by changing the source cells. Normal-speed listening, reader comprehension and the complete craft/coverage goal remain open.

  • Complete the final visual/editorial review of all sixteen draft slides. The hash-bound isolated review covers desktop, compact desktop, mobile initial/table-end states, every PDF page, notes and at least three useful questions per slide. The cache flow was corrected into a route comparison and rechecked. This does not complete embedded-context, recorded-delivery or full-library review; repeat any review invalidated by integration.
  • Add strict-floating-point configuration across game/server/editor targets: MSVC /fp:strict /fp:except-, other-platform -fno-fast-math -ffp-contract=off, deterministic cook and shader-code order. Qualify the overview's unity claim: game/server explicitly disable unity, but the pinned editor target has no such override; do not infer its effective setting from that absence.
  • Explain that V2_DETERMINISM and V2_STRICT_FP are reserved definitions; the pinned source says no C++ source branches on those defines. Compiler/cook configuration and contract checks are distinct from runtime determinism proof.
  • Add the worked §7.1 chain: source reference → backlog task → FloatingPointDeterminism_V2_Contract.json → check-v2-floating-point-determinism.py → v2-build.yml → actual evidence. Teach the schema/task/targets/definitions/flags/cook/automation fields.
  • Cover reference DDC, x64 build/cook wrappers, duplicated target helpers, game/server/ editor differences and the contract/checker's actual cross-file assertions. Explicit BuildGraph Cook nodes carry the DDC argument; direct wrappers do not. Static mutation regressions demonstrate that the checker does not assert unity, the targets array or automation fields, while removing a required compiler flag fails. The subsequent XML review also separates explicit Cook entries from the following BuildCookRun -cook command in each of seven graph nodes: only the former has an explicit DDC argument. Inherited environment means the missing argument alone cannot establish the effective cache.
  • Cover gRPC pre-build generation through @oshun/codegen, v2:grpc-stubs, v2:grpc-stubs:check and how stale generated descriptors are detected. Explain the metadata-only output and regex parser limits separately from transport/serialization and live RPC integration.
  • Explain the pinned 826 contracts / 1,155 checkers / 27 Source modules counts as dated inventory observations, never as current passed-release metrics.
  • Cover documentation validation: exactly 138 distinct top-level backlog numbers, resolving bare references and file/anchor links, explicit planned package names, rejected open-decision markers, and limits of static documentation checks. A regression demonstrates that a known parent can admit a missing child reference; external HTTP destinations are not checked.
  • Preserve the distinction between 30 glossary names and 27 Source modules; distinguish absent targets, actual plugins, the three extra Source modules and planned per-mode GameFeature surfaces. Do not equate a registry with activation. The separate V2Editor plugin-directory entry contains resources and a widget contract, not a plugin descriptor; it is distinct from the V2Editor Source module.
  • Finish the five source/teaching assignments, choose the eighteen-slide sequence and add a prerequisite-linked catalog entry under Products/V2.
  • Finish narration, regenerate exports and run all applicable delivery checks for the integrated guide. Final listening, reader and craft acceptance remain separate requirements.

The one-off authoring Python files on the old machine were temporary and became stale after editorial changes. Do not reconstruct and rerun them over the saved JSON. The repository content and preserved JSON draft are the continuation sources. Pinned V2 source pages are V2/features/product-promise-and-surfaces.md, V2/architecture/v2-product-promise.md and V2/docs/README.md.

6A. Track plans: the tasks of section 6 (added 2026-09-18)#

Section 6 below is a charter. Its boxes say what complete coverage means ("explain", "teach", "trace", "for every subsystem"), and not one of them is a piece of work an agent can pick up and finish: each is the size of a track, and the redesign file, which has governed authoring since 11 September 2026, plans work as tracks of chapters of slides. So the tasks of section 6 are these plans. Each turns one part of the charter into a ### Track · … section of EVE-V1-REDESIGN-TODOS.md, where the chapter and slide items that result are ordinary, granular, checkable work. The charter's boxes stay below as the acceptance list: one is checked when the chapters that satisfy it have shipped, never from a plan. The task board lists them under Later for that reason.

The recipe for a track plan (every "Plan …" task below):

  1. Read the charter bullets the task names, and their sources, at the pinned inventory revision.
  2. Write the track into Phase D of the redesign file in the form the existing tracks use: a row in the Phase D table, chapters with a learning outcome each, and under each chapter one item per slide giving the slide id, the Eve-edition layout and the source unit it teaches. Follow the chapter definition of done in that file's section 1. Teach shipped behaviour as shipped and everything else as owed, with the tracker item named.
  3. Add the track's row to authoring/cross-cutting-coverage-matrix.json (learning outcomes, audience, source units, chapters, prerequisites, worked examples, implementation status), creating the file with the first plan.
  4. Verify: end the plan with a table that maps every charter bullet the task names to at least one planned slide id; ./eve check prints "registry clean"; python3 tools/presentations/check-generation-crosswalk.py reports no failure it did not report before the plan (on 2026-09-18 it reports one, T.20.08, which Phase A of the redesign file owns). A plan adds no slide and changes no published file, so no build or browser run is owed.

Work them in this order. PC6.01 and PC6.02 are tooling and come first because the plans cite their output.

  • PC6.01 Give the inventory units below the file level. tools/presentations/inventory-repository.py already discovers every tracked path of the pinned tree; it says of itself that it is not a semantic-unit pass. Extend it (and presentation_repository.py) to record, per file kind, the units a slide can teach: exported symbols of TypeScript, Rust and Python sources, jobs of GitHub workflow files, models and enums of Prisma and SQL schemas, nodes of ComfyUI workflow JSON, and keys of the deployment compose files, each with path, identity, hash and owning package. Do it after the redesign file's A4 re-pin, because the inventory is pinned at 1ceae52c, which predates a third of what it would read. Verify: tools/presentations/tests/ gains a fixture tree with one file of each kind and asserts the units; check-center-coverage.py reports the new unit kinds as unassigned rather than ignoring them; secrets and private runtime data stay unpublished (a spec plants a key-shaped value and asserts it is not exported).
  • PC6.02 Write authoring/subsystem-teaching-index.json: every subsystem root the repository inventory lists (apps, libraries, services, native and Unreal projects, tools, infrastructure, workflows), each with the track and chapter that teaches it or the word unplanned. This is what makes the charter's "for every subsystem" measurable. Verify: a test asserts that every root under inventory/repository/roots appears exactly once and that every named chapter exists in the redesign file or in catalog-source.json.
  • PC6.03 Plan the track Find your way around the repository: layout, naming, product and domain boundaries, workspace and package configuration, the dependency graph, shared code, language and runtime boundaries, generated artifacts, developer setup, resource-aware commands and debugging (charter: "Whole-monorepo inventory", the orientation and developer-setup bullets). Start from CLAUDE.md, domains.json, tsconfig.base.json, pnpm-workspace.yaml and docs/agents/.
  • PC6.04 Plan the track The architecture in six views: system context, service, component, deployment, data flow and sequence, each linked to the implementation chapter that owns it (charter: the architecture and end-to-end trace bullets). Reconcile with the Shared systems track, which already plans contracts, tenancy, events and discovery, so that nothing is taught twice.
  • PC6.05 Plan the additions to the Delivery and release track for infrastructure and operations. The deployment that exists is one Hetzner box running docker-compose (infra/hetzner/, .github/workflows/deploy-hetzner.yml), which replaced AWS ECS; teach that as live and everything else as declared. Cover networking, identity and secrets handling, databases and storage, queues, CPU and GPU scheduling, observability, capacity and cost, backup, restore and incident runbooks (charter: the infrastructure bullet).
  • PC6.06 Plan the additions to the same track for CI/CD: source change, triggers and path filters, job dependencies and matrices, gates, artifacts, environments, release acceptance and rollback, read from the actual files under .github/workflows/ (charter: the CI/CD bullet).
  • PC6.07 Plan the track What a passing test establishes: unit, integration and contract tests, UI, mobile and native end-to-end, accessibility, performance, security, model and media evaluations and release verification, with what each gate proves and what it does not (charter: the testing bullet).
  • PC6.08 Write authoring/generation-coverage-matrix.json. Run tools/presentations/inventory-workflows.py and list every generation family it and a search of the provider adapters find (text and code, images, speech, music, video, 3D and scenes, procedural and game content, learning material). Map each to the chapter of the redesign file's Creation, Directed human video, Generation infrastructure, Models, lanes and licences, Open 3D studio or Film studio in Blender tracks that teaches it, or to unplanned, with its status (implemented, experimental, planned). Verify: a test asserts that every workflow family in inventory/ appears once.
  • PC6.09 Plan one track for the generation families PC6.08 marks unplanned (expected: text and code generation, Euterpe music, speech, procedural and game content, learning material). ComfyUI node-by-node walkthroughs use the node-graph-wiring layout the edition already has.
  • PC6.10 Plan what the charter's agentic-workflow bullets ask for beyond the Creation track's agentic-studio chapter and the Eve track How Eve works: list the subjects those chapters already carry, then plan chapters only for the residue.
  • PC6.11 Plan cross-cutting track 1, Product purpose and user journeys, from charter track 1 below.
  • PC6.12 Plan cross-cutting track 2, Identity, permissions and trust; start from docs/platform/auth-identity.md.
  • PC6.13 Plan cross-cutting track 3, Data and content lifecycles; start from docs/compliance/pii-inventory.md.
  • PC6.14 Plan cross-cutting track 4, Systems beyond generation (search, ranking, computation, optimization, simulation; Kalika and Concordia included).
  • PC6.15 Plan cross-cutting track 5, Money, access and resource economics; reconcile with the Governance and commerce track's billing chapter.
  • PC6.16 Plan cross-cutting track 6, Architecture decisions and evolution, from the ADRs and docs/migration/breaking-changes.md. The AWS-to-Hetzner move is one of its worked cases.
  • PC6.17 Plan cross-cutting track 7, How to extend the platform; reconcile sdks/ and examples/ with the paths that are supported today before promising any of them.
  • PC6.18 Plan cross-cutting track 8, Operational case studies. Use dated real incident records (docs/agents/incident-2026-08-15-jetsam-freeze.md is one) or exercises labelled as exercises; invent none.
  • PC6.19 Plan cross-cutting track 9, Governance and human workflows; reconcile with the Governance and commerce track's review and privacy chapters.
  • PC6.20 Plan cross-cutting track 10, Role-specific learning and practice: five onboarding routes built from chapters other tracks own, plus exercises; reuse docs/training/ only after checking that it is current.
  • PC6.21 Plan cross-cutting track 11, The Presentation Center itself: pipeline, coverage and freshness rules, the maintainer walkthrough, the edition's layouts and the craft standards.
  • PC6.22 Plan the V2 product track from the charter's V2 bullet. V2 has five guides today (40-v2-* to 44-v2-*) in the old system and no playable build (V2/V2_TODOS.md section 0), so every mechanic is taught as designed or owed, never as shipped.
  • PC6.23 Plan the V3 product track from the charter's V3 bullet, on the same honesty rule; V3/V3_TODOS.md says which of its measurements are hand-authored.
  • PC6.24 Plan the V4 product track from the charter's V4 bullet; its 770 asset files are textual stand-ins, which the track must say.
  • PC6.25 Plan the V5 product track from the charter's V5 bullet.
  • PC6.26 Plan the V6 product track from the charter's V6 bullet.
  • PC6.27 Plan the V7 product track from the charter's V7 bullet; V7/V7_TODOS.md section 0 lists what does not run yet.
  • PC6.28 Plan the V8 product track from the charter's V8 bullet.
  • PC6.29 Plan the V9 product track from the charter's V9 bullet.
  • PC6.30 Plan the V10 product track from the charter's V10 bullet.
  • PC6.31 Plan the Reference shelf track from "Generated and supporting references" below: APIs and contracts, schema entities and migrations, Unreal modules and Rust crates, observability, performance budgets and ownership, localization and design references, Docs Center maps and meta surfaces. Reference-heavy chapters are where the edition's one-table-in-four budget will bind; say how each chapter stays inside it.
  • PC6.32 Compare the charter's Eve bullet with the chapter lists of the redesign file's Phases B and C, and its V1 bullet with the Phase D tracks. Record every subject no planned chapter carries, then either add a chapter to the redesign file or write down why it is not taught. Verify: the comparison table is committed under authoring/ and every row ends in a chapter id or a reason.

6. Comprehensive whole-monorepo, V1–V10 and Docs Center coverage#

The eventual Presentation Center must teach every aspect of this monorepo through high-quality, impeccably organized and sequenced presentations. Cover both the big picture and granular implementation. The existing documentation inventory does not by itself establish complete code, workflow, infrastructure or operational coverage; extend discovery and verification to those sources.

Authoritative source snapshot and inventory#

The baseline presentation corpus and delivered inventory use the immutable source revision 1ceae52c7896a59ac74a4429d410f3218f0b91c6. Later slides may declare additional immutable revisions through the builder's supplemental source contract; inspect authored declarations and coverage.json, and verify every citation against its declared revision. A supplemental slide does not repin the baseline inventory or establish teaching coverage for newer material. The working branch and origin/main can move independently. Do not silently replace a source pin or present old observations as current production facts.

The delivered inventory/index.json and its 105 hash-verified shards contain:

Source family Sources/pages Heading units Meaningful units Words
Authored corpus 9,262 106,387 96,443 8,380,696
Generated reference pages 319 6,957 6,883 465,147

There are 103,326 meaningful units requiring teaching review and 10,018 structural units. The Docs Center build has 2,942 authored pages, 3,271 emitted files, 3,505 entities and 90 spaces. Earlier temporary inventories had lower counts and were incomplete; use the delivered index and shards instead.

The collector uses the actual pinned Docs Center build in an owned temporary Git worktree with GIT_LFS_SKIP_SMUDGE=1. It includes all manifest.pages, all tracked V1–V10 Markdown, domain Markdown and local Markdown links exposed by emitted references. Product documentation pages can be absent from space.pages while present in manifest.pages; do not regress to the smaller collection. Generated reader main content is parsed, preserving headings and link targets; unknown shells fail collection instead of silently disappearing.

  • Reproduce inventory integrity/freshness on the new host. The reproduction receipt binds two independently supervised, exit-0 collections at commits 07261ed5e97 and a60560f3aa8. Both match all 105 declared shard hashes and report the same 9,262 authored / 319 generated sources and 96,443 authored / 6,883 generated meaningful units. The collector and pinned inventory remain unchanged between those passes; avoid recollecting the entire tree during every small edit while that boundary remains true.
  • Decide and document the final freshness boundary. The freshness policy requires an explicit immutable source cutoff already on origin/main, an atomic repin of all inventory families and a path-level candidate comparison. New, removed, changed and generated material has explicit assignment-invalidation rules. Any later tracked change outside the Presentation Center paths forces a newer cutoff and another review; the current pinned inventory remains an expansion baseline, not a current-main claim. The executable gate reports current drift and requires literal cutoff/claim SHAs for final acceptance. The verification receipt retains the working audit and expected final-gate rejection.
  • Review every meaningful unit and map it to specific explanatory slides in specific guides. Include table rows, code/config contracts, schemas, state and error inventories, diagrams, dependencies, ownership and dated evidence.
  • Backfill old Eve/V1 guides against the same unit-level standard. The earlier 191-source citation map does not prove that every section has been taught.
  • Keep unassigned, assigned, reviewed, stale and structural states separate. Store assignments in assignments/*.json with source hash, exact slide hash, guide membership, citations including the relevant excerpt, an explanation of what is taught and a substantive review rationale.
  • Use the assembled slide hash, including graph, inherited release and chapter. A later slide or source edit must invalidate review when appropriate.
  • Treat repeated source units as review candidates, not automatic exemptions. Shared explanations may serve several source contexts only when they actually teach each one; retain the mapping and rationale for all occurrences.
  • Keep source excerpts available and safe, but do not count a citation or an unprocessed document dump as an enriched presentation explanation.
  • Maintain source → unit → guide → slide traceability and a readable coverage report. Show the remaining gaps clearly during development.
  • Require check-center-coverage.py --require-complete to pass with no invalid, stale, unassigned or merely assigned meaningful units before claiming complete coverage. The current thirty-two-reviewed / 103,294-unassigned result is an expected incomplete gate, not complete teaching of the corpus. These totals are historical; read the freshly checked teaching report for the current counts.

Tools to preserve and use:

  • tools/presentations/presentation_inventory.py: inventory unit identity, generated-page parsing, shards, exact hashes and assignment validation.
  • tools/presentations/inventory-center.py: isolated pinned collection and check.
  • tools/presentations/check-center-coverage.py: report and final complete gate.
  • tools/presentations/tests/test_inventory.py: 16 coverage regressions.
  • coverage-map.json: cited-source/slide/guide relationships.
  • foundation-coverage.json: immutable original 244 IDs and foundation HTML hash f21f0b91ba2444d2902601840cda9cd5f9a9dde853a5e7d4d21dabe28ee96f58.

Decomposition and guide-count decision#

  • Build the complete learning map from the inventory before setting a final count. Record a justified plan by topic/subtopic and revise it as teaching review discovers detail. A word-count estimate alone is not a decomposition.
  • Pair product features with their owning architecture, source contracts, operational dependencies, worked paths, failure/recovery and evidence.
  • Keep a short overview, then focused mechanism and worked-case guides; separate reference-heavy inventories where that improves comprehension.
  • Give each guide a clear learning outcome, prerequisite context, complete explanatory narration, useful questions and a canonical place in the hierarchy.
  • Preserve all reviewed foundation material. Corrections or superseded evidence require deliberate replacement/mapping, not silent deletion.
  • Avoid arbitrary equal-sized cuts, thin generic slides, duplicate giant guides, placeholder topics and claiming complete systems from one overview.
  • Keep progress reviews honest: presentation artifact verification is distinct from acceptance of the underlying product, provider, native app or deployment.

Whole-monorepo inventory, architecture and implementation depth#

The pinned Git discovery index accounts for 140,123 tracked entries across 43 root groups, including all tracked dot-directories, one symlink and one gitlink. It binds 3,324 package and 3,505 project manifests to their exact source objects and compares declared pnpm patterns, registry roots, documentation membership and presentation citations. There are 130,861 paths outside the documentation inventory, 44 non-root package manifests outside the declared patterns and ten repeated package names. These are discovery/reconciliation inputs, not taught units or automatic dependency defects. Effective graph resolution, non-Markdown semantic units, actual ownership/implementation review, external payloads and teaching assignments remain open. No checklist item below is closed by file enumeration alone.

The workflow declaration inventory extends non-Markdown accounting to all 106 pinned workflow files and two local action definitions: 308 jobs, 1,767 steps and 3,718 meaningful units. Their exact ranges account for all 1,065,561 bytes; unsupported sources cannot silently disappear, and reviewed assignments must bind the whole workflow context as well as the unit and assembled slide. All 3,718 units remain unassigned. This does not parse shell statements, validate effective CI execution, establish ownership, teach the workflows or close the other repository families below.

  • Inventory every repository root and nested subsystem, including tracked dot-directories, apps, libraries, packages, services, native/Unreal projects, plugins, schemas, assets, configuration, scripts, tooling, documentation, tests, evaluation fixtures, infrastructure and CI/CD definitions. Reconcile filesystem/Git discovery with workspace manifests and dependency graphs so subjects missing from Docs Center navigation are still included.
  • Extend the source/teaching inventory beyond Markdown to meaningful code, configuration, workflow, schema and asset units. Record revision, path, symbol/section or node identity, hash, responsibility, owner, dependencies, implementation state and guide/slide assignments. Explain generated or vendored families through their origin, contracts, local integration and regeneration/update process; account for repeated units without silently dropping unique behavior. Keep secrets and private runtime data unpublished.
  • Create repository-orientation guides covering the monorepo layout, naming, product/domain boundaries, workspace/package configuration, dependency graph, shared code, language/runtime boundaries, generated artifacts and where to find implementations, tests, design decisions and runbooks.
  • Teach the overall software architecture through system-context, service, component, deployment, data-flow and sequence views. Explain client/server boundaries, domain ownership, shared substrates, trust/tenancy, persistence, events/queues, external providers and cross-product integration, then link each view to focused implementation guides and representative end-to-end paths.
  • For every subsystem, cover purpose and user/operator value, public APIs and contracts, entry points, key modules/functions/classes, algorithms, state machines, data models, concurrency, caching, error handling and recovery wherever applicable. Explain actual code with small annotated excerpts and worked inputs/outputs; a directory listing or symbol dump is not teaching.
  • Trace representative behavior from UI/CLI/event entry through routing, validation, authorization, orchestration, execution, storage and returned results. Include alternate/error paths, lifecycle transitions, relevant schemas/configuration and the tests that demonstrate each important contract.
  • Cover developer setup, local environments, dependency installation, code generation, build graphs, caching, package boundaries, resource-aware commands and debugging. Explain configuration precedence, feature flags, migrations, compatibility and dependency/toolchain upgrades where present.
  • Create detailed infrastructure and operations guides: actual deployment topology, infrastructure as code, containers/orchestration, networking, identity/secrets management, databases/storage, queues, CPU/GPU scheduling, observability, performance/capacity/cost, backups, restore, disaster recovery and incident runbooks. Distinguish declared infrastructure from live evidence.
  • Create detailed CI/CD and release guides tracing source change → triggers and path filters → job dependencies/matrices → build/test/evaluation gates → artifacts → environments/deployment → release acceptance and rollback. Explain workflow permissions, credentials boundaries, caches, migrations, supply-chain checks and failure diagnosis from the actual workflow files.
  • Cover unit/integration/contract, UI/mobile/native end-to-end, accessibility, performance, security, model/media evaluations and release verification wherever present. Explain fixtures, determinism, evidence, known gaps and exactly what passing each gate establishes about the underlying system.
  • Give every meaningful unit a substantive reviewed teaching assignment; maintain explicit missing, stale and unresolved queues. Extend inventory tests and the complete-coverage gate to reject unaccounted repository roots, subsystem families and meaningful non-Markdown units. Passing the existing documentation-only gate is insufficient for whole-monorepo completion.
  • Review source changes before final publication and invalidate affected assignments, diagrams, screenshots, narration and walkthroughs. Preserve a dated, reproducible coverage boundary and a maintenance process so new packages, workflows and infrastructure do not disappear from the center.

All generation systems, ComfyUI and agentic workflows#

  • Discover every generation entry point, engine, orchestrator, workflow family, provider adapter and consuming product/domain across the monorepo. Include text/code, images, audio/speech/music, video, 3D/animation/scenes, procedural/game content, learning material and other modalities wherever present. Explain shared systems once in depth and cross-link their distinct integrations; do not limit coverage to the currently named studio guides.
  • Build a generation-system and workflow coverage matrix linking each family to its high-level guide, granular implementation guides, source definitions, UI/workflow visuals, contracts, runnable examples and verification evidence. Include implemented, experimental and planned systems with clear status.
  • Teach the complete generation lifecycle: user intent/brief and input assets → workflow/model selection → validation/admission → scheduling/execution → streaming/progress → outputs/post-processing → evaluation/review → storage, provenance and publication. Show which component owns each transition and how cancellation, retries, partial failure and recovery affect the result.
  • Provide dedicated ComfyUI overviews and detailed workflow walkthroughs. Explain graph structure, node types and connections, typed inputs/outputs, parameter binding, custom-node registration and dependencies, workflow templates/API representations, model/checkpoint/LoRA selection and relevant conditioning, sampling, decoding and post-processing stages where used.
  • Walk representative ComfyUI workflows node by node using readable graph screenshots/diagrams, focused detail slides and concrete inputs/intermediate outputs. Explain why each node and parameter exists, how data changes along each edge and how variations change results; cover every material workflow family and its unique behavior, not just one attractive example.
  • Explain ComfyUI submission/queueing, execution callbacks or polling, output collection, asset/model paths, workflow/custom-node versioning, reproducibility, seeds where applicable, resource/VRAM constraints and CPU/GPU/provider placement from the actual integration. Cover admission, disallowed nodes, permissions, timeouts, cancellation and failure diagnosis.
  • Provide dedicated agentic workflow and system guides at both overview and granular levels: goals, planning, agent roles, orchestration/graphs, branching/loops, prompts/models, context/memory/retrieval, tool schemas and execution, delegation where implemented, human review, stopping conditions, durable state, resumption and result validation.
  • Trace concrete agentic runs step by step, including messages and state changes, tool inputs/outputs, authority/confirmation, budgets, retries, idempotency, observability and recovery. Explain how agents invoke generation workflows and how workflow execution differs from model/agent decisions. Map each step to the actual modules, contracts and relevant tests.
  • Explain model/provider registries, capability routing, fallbacks, local versus remote execution, streaming/batching, training/fine-tuning and model promotion where present. Cover modality-specific preprocessing, quality controls, evaluation, provenance/rights, retention and output handoff.
  • Include complete worked examples and dedicated implementation/operations slides with UI screenshots, workflow graphs, sequence/state diagrams, annotated code/configuration and representative artifacts. State inputs, versions, parameters, expected results and failure cases; distinguish live execution evidence from fixtures, mock outputs and proposed workflows.
  • Verify runnable examples through the applicable service, contract, workflow, model/media evaluation and release gates; add coverage for missing important paths. Track unavailable provider/hardware evidence explicitly while still teaching the documented architecture and implementation honestly.

Required cross-cutting presentation tracks#

The user explicitly requested all eleven tracks below. Several subjects already appear in individual product/domain guides; give them coherent, discoverable learning routes with complete teaching assignments and cross-links. Reuse reviewed explanations where appropriate, while covering each distinct integration and journey. Eleven tracks is not an eleven-guide cap. Each track needs as many overview, mechanism, implementation and worked-example guides as its scope requires.

  • Maintain a coverage matrix for all eleven tracks: learning outcomes, audience, source units, canonical guides, ordered slides, prerequisites, worked examples, implementation status and current verification/editorial evidence. Apply the same screenshot/mockup, diagram, narration and quality standards as the rest of the library; a catalog label alone is not coverage.

1. Product purpose and user journeys#

  • Explain who each product serves, the problems it solves, its core promise, business rules, constraints and documented success criteria. Show how products and domains fit together; distinguish intended outcomes from measured results.
  • Walk complete journeys across relevant products/domains, from discovery and onboarding through the user's main task, result and return visit. Connect visible actions to shared services and data, including permissions, handoffs, alternate paths, accessibility needs and recovery when the journey fails.
  • Trace each journey to actual routes/screens, contracts, owners and tests. Use one coherent worked scenario across the overview and implementation guides, with screenshots/mockups and honest implemented/planned boundaries.

2. Identity, permissions and trust#

  • Teach authentication, identity federation/SSO, sessions, refresh/revocation, recovery and step-up verification; then roles/scopes, permissions, entitlements, service identities, consent and tenant isolation wherever applicable. Explain canonical contracts and any parallel or partially integrated runtime paths.
  • Trace allowed and denied requests from the client through gateway/BFF, domain authorization, storage and external integrations. Show where identity is established, authority is checked, data isolation is enforced and audit evidence is recorded, including expired/revoked grants and cross-tenant cases.
  • Connect trust-boundary/threat diagrams to actual policies, code, contracts and security tests. Include impersonation/delegation boundaries and human confirmation where present; start from docs/platform/auth-identity.md and its governing decisions, then reconcile each product's implementation.

3. Data and content lifecycles#

  • Map collection/ingestion → validation/transformation → storage/indexing → synchronization/distribution → publication/use → retention/archive → export/deletion. Include authored content, generated media, datasets, embeddings, derived assets and metadata wherever present.
  • Explain sources of truth, schema ownership, identifiers, lineage, event contracts, consistency, conflict resolution, idempotency and schema evolution. Show how updates and deletions propagate through caches, search indexes, replicas, object storage and downstream consumers, including failure cases.
  • Walk concrete data/asset examples through consent, access, residency, retention, export, deletion and backup/restore boundaries. Link actual inventories such as docs/compliance/pii-inventory.md to implementation and tests; distinguish documented obligations from verified enforcement.

4. Systems beyond generation#

  • Inventory search, retrieval, ranking, recommendation, mathematical/symbolic computation, statistical analysis, optimization, simulation and decision systems across all domains. Include relevant Kalika and Concordia mechanisms; do not treat the generation/model pipeline guides as complete algorithm coverage.
  • Teach each material algorithm's problem, inputs/outputs, assumptions, constraints, mathematical intuition, execution steps, data structures, complexity and tradeoffs at suitable levels of detail. Use worked examples, diagrams and annotated implementation excerpts with prerequisite explanations.
  • Explain correctness conditions, numerical precision, determinism, convergence or approximation limits where relevant, and compare alternatives using actual benchmarks/evaluations. Cover edge cases, failure behavior and test evidence; distinguish theoretical guarantees from measured implementation behavior.

5. Money, access and resource economics#

  • Connect plans/subscriptions, prices, invoices, usage metering, quotas and entitlements to product access. Explain payments, ledger/accounting entries, settlement, refunds, chargebacks and reconciliation wherever implemented or planned, preserving rail/provider/release differences.
  • Trace purchase/renewal/cancellation and failed/duplicate/delayed payment scenarios through events, state transitions, authorization and customer support. Show when access is granted or removed and how discrepancies are investigated; link contracts, idempotency rules and automated verification.
  • Explain CPU/GPU/storage/network/model-provider costs, per-operation resource accounting, budgets and capacity tradeoffs. Use reproducible, dated examples with explicit assumptions; separate configured limits, estimated economics and measured charges. Start from the support/billing and payment-substrate sources, including V1/architecture/support-billing-and-crypto.md.

6. Architecture decisions and evolution#

  • Build a coherent architecture-evolution route from governing ADRs, migrations, consolidation work and deprecation records. Explain the original problem, constraints, alternatives, selected approach, tradeoffs and consequences; distinguish accepted, superseded and proposed decisions.
  • Show before/after architecture and code/contract examples, including legacy coexistence, partially completed migrations and unresolved seams. Trace compatibility across APIs, SDKs, events, data schemas, configuration and supported clients rather than presenting only the desired final design.
  • Explain upgrade paths, compatibility/deprecation policies, migration order, rollback and the tests proving transition safety. Link dated source evidence, including docs/migration/breaking-changes.md; do not invent historical motivations or present old compatibility claims as current without review.

7. How to extend the platform#

  • Provide worked contributor guides for adding a domain, package, API, SDK integration, plugin, agent tool, ComfyUI node and workflow wherever the repository supports those extension points. Explain shared steps once and teach the distinct registration/runtime contracts of each extension type.
  • Trace implementation through naming/ownership, manifests and dependency boundaries, schemas/types, registration/discovery, configuration, permissions, UI integration, packaging/deployment, versioning and documentation. Show how an extension becomes reachable and how to disable or remove it safely.
  • Provide minimal complete examples, expected behavior, failure diagnosis and appropriate unit/integration/contract/end-to-end verification. Reconcile sdks/, examples/ and existing scaffolding with current supported paths; label proposed extension mechanisms instead of implying they already work.

8. Operational case studies#

  • Create worked incident guides for failed deployments, provider outages, stuck/broken generation jobs, capacity exhaustion and data recovery wherever relevant. Follow detection → triage → diagnosis → containment → mitigation/ rollback → recovery validation → follow-up prevention.
  • Connect symptoms to logs, traces, metrics, queues, affected dependencies and user impact. Explain incident ownership/escalation, blast radius and the documented service objectives, error budgets and recovery targets; link docs/operations/slos.md and relevant runbooks to implementation evidence.
  • Use dated real incident records or clearly labeled controlled exercises, including failure injection, restore/rollback checks and expected observations. Verify examples in appropriate isolated test environments; distinguish declared recovery targets from measured results and production experience.

9. Governance and human workflows#

  • Teach moderation, review, approvals, appeals, support and escalation as complete workflows: actors, authority, intake, evidence, decisions, handoffs, notifications, resolution and recorded outcomes. Explain ownership boundaries between user, operator, reviewer, administrator and automated systems.
  • Show state transitions, separation of duties, conflicts of interest, audit trails and policy/version changes wherever documented. Include worked approval, rejection, appeal and correction cases tied to actual interfaces, contracts, persistence and applicable service/UI automation.
  • Cover asset/model/code licensing, provenance, rights/consent, attribution, procurement and distribution restrictions where applicable. Connect compliance documentation and governance policies to enforcement and review evidence; clearly distinguish policy claims, legal review and automated technical checks.

10. Role-specific learning and practice#

  • Create sequenced onboarding routes for developers, designers, operators, reviewers and creators, with shared prerequisites, a glossary, clear learning outcomes, increasing depth and links into canonical technical/design guides. Include complete practical tasks appropriate to each role.
  • Add exercises, worked solutions, expected results, troubleshooting and comprehension checks. State required setup, versions, test data and how to reset/retry; reuse and verify applicable material from docs/training/, workshops, tutorial notebooks and examples instead of duplicating stale steps.
  • Validate runnable exercises with the applicable automation and review the learning sequence for understanding and task completion. Record actual learner feedback when obtained, resolve confusing steps, and keep content/test passes distinct from evidence that a person can successfully perform the task.

11. The Presentation Center itself#

  • Explain the center's architecture and publishing pipeline: pinned sources → discovery/inventory → teaching assignments → authored content/diagrams/assets → narration/scripts → catalog/navigation → standalone/embedded HTML and PDFs → verification and publication. Map each stage to its actual files and tools.
  • Teach coverage/freshness rules, source/slide/asset hashes, generated versus authored files, speech fingerprints, pronunciation/pacing, measured durations, hierarchy/shared-slide semantics, portable assets and accessibility. Explain how these contracts prevent stale or incomplete artifacts from appearing ready.
  • Provide a maintainer walkthrough for adding/updating/splitting a guide, changing a screenshot/diagram or narration, regenerating affected artifacts, reviewing previews, running relevant gates and publishing. Include deliberate missing/stale-source, broken-link, failed-audio and layout failures with recovery steps; keep the center's own guides current as its implementation evolves.
  • Teach section 10's slide briefs, composition selection, semantic visual primitives, diagram contracts and whole-guide rhythm review. Show how to extend the renderer and verify a new treatment without reintroducing a fixed point count, generic column fallback or compulsory template cycle.
  • Teach the visual benchmark and all eight craft/comprehension standards from section 10 in the maintainer route. Demonstrate font/scale and chart contracts, stable visual identities, image crops, actual-size review, reader-study evidence, subtraction and review invalidation with real authored examples. Keep the reference collection and authoring guidance synchronized with the delivered renderer.

Product-specific remaining work#

These are scope prompts, not substitutes for enumerating each product's actual inventory units. Consult each product's feature, architecture, documentation, backlog, dependencies, decisions, audits, generated references and code evidence.

  • V1: expand beyond the first 35 Eve/V1 guides to all inventoried V1 and Docs Center information. Preserve member/builder separation; Tara, Arete, Nyx, Nisaba, Lilith, Veritas, Metis; web/PWA/native/watch; release and payments; creation/studio/living scenes; privacy/review/grounding/memory; channels and Telegram; contracts, operations, delivery and acceptance.
  • Eve: the completed foundational guides are not a blanket full-inventory review. Cover the full handbook, runtime/actions/tools, four planes, context, six tool layers, authority/confirmation, trust propagation, task families, model/quality/evaluation, cost/budgets, observability, failure recovery, fleet, evidence and current charter; reconcile dated implementation qualifications.
  • V2: finish the third orientation/contract guide, then combat/defense/game feel, rulesets/tag/assists, mode catalog/training/replay/signature events, roster/stages/story/relationships/progression/creator suite, online/ranked/ tournaments/esports/integrity, commerce/live service/community/support/AI, racing/vehicles/crossover/open world, accessibility/localization/platform, balance/wellness/privacy/launch. Cover deterministic GAS/frame data/input/ animation/rollback, HUD/AV/XR, services, authored data/build/cook, security, performance/telemetry/tests/release and all detailed source units. The remaining orientation map identifies whole-unit gaps and separates replication declarations from actual engine consumers. The rollback source map records the complete architecture-page review and concrete tick/recovery, prediction, fallback and diagnostic discrepancies to exercise against actual C++. The native probe attempt reproduces an unchanged-pin UnrealHeaderTool failure on local UE 5.5.4. Its explicitly adapted diagnostic host passes header generation, then encounters seven unchanged V2Input compiler errors. The five-case harness translation unit compiled independently, without executing its assertions. The subsequent Linux diagnostic run adds ten exact dependency-expression adaptations, disables unity merging and completes all 321 project-local build actions. All five cases are now discovered, executed and passing in that explicitly adapted host, with process exit zero. Twenty-two preparation/integrity tests pass. Preserve the original failures and all fixture qualifications: this is not an unchanged-pin build, Blueprint acceptance, full-game target or network integration pass. No product-source repair is included. Neither map nor this attempt creates a teaching assignment. The expanded native suite now executes nineteen cases successfully, including bounded malformed codecs, parity/datagram records, handshake/fallback thresholds, history, structured restore and presentation interpolation/event queues. Only the test module is recompiled/relinked; all copied product-source bytes and fixture adaptations remain unchanged. The receipt retains the failed discovery-directory permission attempt, corrected discovery, build, runtime and exact dependency-module continuity. This is not real socket, encryption, audiovisual-consumer or complete snapshot-substate coverage. Finish the remaining source-unit review and integration automation before claiming complete rollback coverage; passing cases prove only their named assertions. The connection-quality manuscript now develops a coherent twelve-slide explanation of measurements, handshake, pressure, fallback, validation, score and diagnostic boundaries. Its draft receipt records a twentieth passing native case, complete isolated browser/PDF checks and mobile numeric-heading repair. A returned decision can fail its separate validator; valid config is not default-policy alignment. The source-annotation browser audit now accepts explicit draft specimens while retaining canonical-source checks. That draft receipt is historical. The later recorded integration adds twelve delivered slides and recordings plus complete embedded/PDF review; it does not close the rollback source map. Reader/listening acceptance, reviewed assignments and the two explicit typography defects remain open. The subsequent rollback-session guide delivers fourteen narrated slides and two reviewed architecture units: the session loop and prediction/trigger sections. Five additional native cases exercise confirmation consumption, key boundaries, correction order, duplicate confirmations and frozen retry/reset. All 25 cases pass in the same adapted host; no product source is repaired. Shared typography and short-viewport receipts resolve the earlier two display defects. Full rollback source coverage, real consumers, reader/listening acceptance and unchanged-pin/full-game/network acceptance remain open.
  • V3: embodied world services, avatars/rooms/presence/spatial audio and client tiers; Tara Studio, Saraswati Stage and Lilith Commons; authoring, identity/social/rights boundaries, shared systems, world/service/runtime and platform delivery, failure and acceptance. Do not equate content targets with actual authored assets or a client tier with weaker trust requirements.
  • V4: tactical-action genre cells and their different local mechanics and networking; common roster/Shadow War/creation/online/production; authored content, plugin activation, platform and runtime evidence. Preserve the distinction between real code/manifests and activated authored binary content.
  • V5: six genre cells, Cross-Continuum Bureau, Mind Palace, investigation, narrative and relationships/consequences/progression/persistence; common runtime/compiler/authoring, source-module versus feature-plugin state, security/operations/build/acceptance and V8 reuse.
  • V6: Egbe, Ori, Orun, stewardship, Chronicle, autonomy/refusal/negotiation, Aye Bridge, persistent continuity, native/stream/web surfaces, identity and governance, providers and end-to-end evidence. Keep live providers explicitly gated where the source says so.
  • V7: Mawu realms and creator experience; executable UGC, server authority, Wasmtime sandboxing, PubGrub dependencies/locks, realm persistence/recovery, capabilities, identity firewall, community governance/appeals, creator economy, discovery, hosting, production/integration gaps and acceptance.
  • V8: Ariadne player promise/case board, symbolic construction, visible clues, solver proof of both uniqueness and the intended solution, narrative/ media realization after proof, suspect/world content, V5 compiler reuse, release decisions, delivery and operations. Preserve the distinction between the full eight ReleaseDecision gates and focused seven-gate contracts.
  • V9: Metis Wonder/Atlas/Lesson/Thread, grounded claims and source pins, computed kernels, prerequisites/learner model, generation and interaction, assessment/memory, quality/provenance gates G1–G7, delivery and provider seams, privacy, operations and real learner evidence. Artifact pass is not a learner study.
  • V10: Rail attention policy and channels, channel-owned content/outcomes, glance/micro-act/live/dive, grants/dayparts/budgets, absence-free behavior, media arbitration with one audio and one video lane, discretion/brand rings, identity inherited from V1, shell/surface families, providers/configuration, recovery/ops and dated design versus implemented-state claims.

Source facts requiring careful interpretation#

These are reminders from the reviewed sources, not newly verified production facts:

  • Oshun Synthetics LTD is the company; Oshun the shared platform; Lilith the V1 member app/assistant; Eve the builder/operator assistant. Shared Lilith policy infrastructure and V3 Lilith Metaverse have distinct meanings.
  • V1.0 is web/PWA, four named rooms and crypto; V1.1 adds native/watch and fiat; V1.2 adds Veritas and Metis. Operator preparation is not member release exposure.
  • Earlier Eve measurements include four planes, six tool layers, 17 writes and a pinned threat inventory of 144 cases / 25 routes / 87 tools. Do not reuse those historical counts as current without checking the dated source.
  • Eve trust labels use a least-trust join; model-facing framing is not admitted authority. The dense index can exist while BM25F still serves retrieval.
  • Sophia's extractive Jaccard/citation thresholds and Iris's eight canonical scopes versus eleven adapter names are distinct concepts. Citation counts are not truth. Iris's 30-day decay constant is an e-folding time, not a half-life; write protection and recall deduplication can select different winners. Iris tombstones retain content in the local store and do not make it the FSRS owner or prove erasure across consumers.
  • V2 public projections and the planned companion do not settle match truth; documented tournament operations and the real broadcast toolkit differ.
  • Existing copied source excerpts may contain broader or older claims. The presentation explanation must retain governing release scope and dated evidence.

All Docs Center spaces must be accounted for#

For every row below, review its actual source/unit lists, assign appropriate guides and complete teaching verification. The count is the authored Docs Center space count, not the entire corpus size for that subject or a guide-count target.

Reviewed Space Kind Authored pages
[ ] v1-oshun — Oshun Platform product 133
[ ] v2-fighting — Fighting Game product 699
[ ] v3-metaverse — Lilith Metaverse product 128
[ ] v4-tactical — Tactical Action product 76
[ ] v5-openworld — Open-World Narrative product 170
[ ] v6-egbe — Egbe Companions product 49
[ ] v7-mawu — Mawu product 46
[ ] v8-ariadne — Ariadne product 23
[ ] v9-metis — Metis product 28
[ ] v10-rail — The Rail product 28
[ ] wt-web — V1 Web PWA walkthrough 863
[ ] wt-admin — Admin Cockpit walkthrough 48
[ ] wt-tenant — Tenant Admin Console walkthrough 17
[ ] wt-mobile — Mobile App walkthrough 13
[ ] wt-telegram — Telegram Mini-App walkthrough 3
[ ] disc-decisions — Decisions (ADRs) discipline 83
[ ] disc-reference — Reference discipline 28
[ ] disc-runbooks — Runbooks discipline 50
[ ] disc-operations — Operations discipline 4
[ ] disc-security — Security discipline 4
[ ] disc-compliance — Compliance discipline 27
[ ] disc-integrations — Integrations discipline 14
[ ] disc-guides — Guides discipline 3
[ ] disc-training — Training discipline 20
[ ] disc-conventions — Conventions discipline 5
[ ] disc-performance — Performance discipline 1
[ ] disc-infrastructure — Infrastructure discipline 3
[ ] disc-migration — Migration discipline 3
[ ] disc-audits — Audits discipline 113
[ ] disc-research — Research discipline 6
[ ] disc-testing — Testing & QA discipline 2
[ ] disc-launch — Launch discipline 14
[ ] disc-proposals — Proposals discipline 66
[ ] disc-api-internal — Internal API notes discipline 2
[ ] disc-content — Content discipline 1
[ ] domain-aglaea — Aglaea domain 3
[ ] domain-airmid — Airmid domain 3
[ ] domain-aja — Aja domain 3
[ ] domain-aje — Aje domain 3
[ ] domain-annapurna — Annapurna domain 3
[ ] domain-aphrodite — Aphrodite domain 3
[ ] domain-arete — Arete domain 3
[ ] domain-asase — Asase domain 3
[ ] domain-athena — Athena domain 3
[ ] domain-bellona — Bellona domain 3
[ ] domain-brigid — Brigid domain 3
[ ] domain-calliope — Calliope domain 3
[ ] domain-concordia — Concordia domain 3
[ ] domain-contracts — Contracts domain 3
[ ] domain-cybele — Cybele domain 3
[ ] domain-demeter — Demeter domain 3
[ ] domain-euterpe — Euterpe domain 3
[ ] domain-freya — Freya domain 3
[ ] domain-gaia — Gaia domain 3
[ ] domain-galatea — Galatea domain 3
[ ] domain-hathor — Hathor domain 3
[ ] domain-hestia — Hestia domain 3
[ ] domain-iris — Iris domain 3
[ ] domain-isis — Isis domain 3
[ ] domain-kalika — Kalika domain 3
[ ] domain-kuanyin — Kuanyin domain 3
[ ] domain-lakshmi — Lakshmi domain 3
[ ] domain-lilith — Lilith domain 3
[ ] domain-maat — Maat domain 3
[ ] domain-maya — Maya domain 3
[ ] domain-meditation — Meditation domain 3
[ ] domain-metis — Metis domain 3
[ ] domain-mnemosyne — Mnemosyne domain 3
[ ] domain-neith — Neith domain 3
[ ] domain-nike — Nike domain 3
[ ] domain-nisaba — Nisaba domain 3
[ ] domain-nous — Nous domain 3
[ ] domain-nyx — Nyx domain 3
[ ] domain-openapi — Openapi domain 3
[ ] domain-oshun — Oshun domain 3
[ ] domain-oya — Oya domain 3
[ ] domain-phoebe — Phoebe domain 3
[ ] domain-proto — Proto domain 3
[ ] domain-psyche — Psyche domain 3
[ ] domain-saraswati — Saraswati domain 3
[ ] domain-seshat — Seshat domain 3
[ ] domain-shakti — Shakti domain 3
[ ] domain-shared — Shared domain 3
[ ] domain-sophia — Sophia domain 3
[ ] domain-tara — Tara domain 3
[ ] domain-themis — Themis domain 3
[ ] domain-uzume — Uzume domain 3
[ ] domain-veritas — Veritas domain 3
[ ] domain-yemaya — Yemaya domain 3
[ ] platform — Shared Platform platform 7

Generated and supporting references#

  • Cover APIs/OpenAPI, shared contracts, configuration, dependencies and events.
  • Cover schema entities/enums/relationships, migrations and lifecycle states.
  • Cover Unreal modules/plugins/build/automation, Rust crates and workflow assets.
  • Cover observability alerts, recording rules, dashboards, scrapes and telemetry.
  • Cover performance budgets/thresholds, testing maps and project ownership.
  • Cover localization, glossary/changelog, design and accessibility references.
  • Cover product-space maps, journeys/results, role lenses, repository maps, generated diagrams/galleries and Docs Center coverage/meta surfaces.
  • Follow relevant supporting Markdown linked by generated pages and Domains, including detailed Isis and Lilith trees. Large volume is not an exclusion rule.
  • Preserve unknown/missing-source failures in discovery and review. A source omitted from navigation can still belong to the coverage universe.

7. Complete organization, comprehension and navigation#

  • Establish coherent learning routes for repository orientation, products, domains, UI/UX/design systems, generation/agentic systems, architecture and implementation, infrastructure/operations, and CI/CD/testing/release. Sequence each route as orientation → concepts → architecture → worked flow → granular implementation → operations/verification, adapting to the subject and audience. Cross-link shared concepts and avoid forcing newcomers through dense internals before they understand the system's purpose. 2026-09-18: the routes are data (readingRoutes in catalog-source.json; tools/presentations/tests/eve-learning-routes.mjs checks Eve's). One route per subject listed here is added when the track it runs through exists, so this follows the plans of section 6A. Verify: each route is in catalog.json, the center renders it, and a route test like Eve's asserts that every step resolves to a published guide in order.
  • Give all eleven cross-cutting tracks and role-specific onboarding routes clear entry points, canonical guide homes and links to relevant product/domain material. Validate preparation → explanation → worked example → practice/ verification sequences, including routes that cross several domains; preserve stable deep links and avoid duplicating entire guides for each audience. 2026-09-18: the eleven cross-cutting tracks do not exist yet; this waits for PC6.11 to PC6.21 and the chapters they plan. blocked:upstream
  • Extend the existing collapsible sidebar to the complete portfolio, shared platform subjects and Domains; keep the presentation as the primary workspace. 2026-09-18: the sidebar renders whatever the catalog declares, so it grows as D0 track setups land. Verify: with the full planned catalog loaded as a fixture, every track and domain is reachable from the sidebar by keyboard, and the deck remains the primary pane at 1440 and 390 pixels wide.
  • Keep Products/V1–V10, cross-product domains and shared mechanisms distinct but cross-linked. Product-specific and domain-wide guides can have different scopes without duplicating an entire guide under several homes. Rule, not a task (un-boxed 2026-09-18): it constrains how the plans of section 6A place a guide; nothing is built by it.
  • Preserve canonical IDs and stable deep links. Plan redirects or explicit mappings for intentionally renamed/split guides and slides. Rule, not a task (un-boxed 2026-09-18). The redesign file already keeps every slide id through a merge; a rename or split writes its mapping in the same commit.
  • Validate that prerequisites precede dependents, every slide has a home, no guide repeats a slide unintentionally and there are no orphaned sources/links. Verify: one builder check with test_build.py cases: a prerequisite that follows its dependent, a slide with no home, a slide repeated in one guide without being declared shared, and a source link with no target each fail the build and name the offender.
  • Add explicit related-guide relationships and navigation where needed; recommended preparation alone is not the entire cross-reference experience. Verify: catalog-source.json gains a related list per guide, validated both ways by the builder (a relation to an unknown guide fails), rendered in the center, with one chromium check that the links navigate.
  • Fix mixed hierarchy nodes when introduced: center.js currently renders child groups before the node's own guides. Parent overview guides should come before their detailed children when that is the intended reading sequence. Verify: a center.js unit test renders a node that has its own guides and child groups, and asserts that the parent's guides come first.
  • Derive every ancestor's listening total from current audio. Count shared context once per guide where a reader hears it, and distinguish that from unique-recording totals. Mark pending totals instead of presenting partial time as the complete duration. Explain rounding and 1× playback semantics. 2026-09-18: tools/presentations/tests/center-duration-labels.mjs implements and tests the label policy (rounded-up minutes at 1×, "Audio pending" for incomplete narration). Verify against the rest of this item (ancestor totals, shared context counted once per guide, the explanation of rounding) and close, or name what is missing.
  • Scale search beyond the current small catalog: meaningful topic aliases, source concepts, product/domain names, clear results and keyboard recovery. Keep navigation responsive and measure it with the expanded data size. Verify: with a catalog fixture of the planned size (about 150 guides and 4,000 slides), a search for a topic alias, a source concept and a product name each returns the expected guide first, keystroke-to-result stays under 100 ms in chromium, and Escape returns focus to the field.
  • Verify deep nesting on narrow screens: readable labels, selection visibility, useful breadcrumbs, mobile drawer focus/inertness/Escape/backdrop, no trapped scrolling, persistent collapse and correct history behavior. 2026-09-18: tools/presentations/tests/center-short-viewports.mjs exists and authoring/center-short-viewports.md records what it covers. Verify against this list (drawer focus and inertness, Escape, backdrop, persistent collapse, history) and close, or name what is missing.
  • Keep guide selection and dialogs from leaving prior narration playing. Preserve stop/replay/end boundaries and clear loading/failure states. Rule, not a task (un-boxed 2026-09-18): the behaviour exists and the audio-flow checks guard it (tools/presentations/tests/audio-flows.mjs); this line says to keep it.
  • Retain standalone HTML, embedded source excerpts, scripts/questions, local audio and individual PDFs. The whole-bank .qa-preview is a test fixture, not a reintroduced giant public presentation. Rule, not a task (un-boxed 2026-09-18): it says what must not be removed.
  • Test self-contained copying and static HTTP serving. Source-file links need the repo; embedded excerpts and basic playback/navigation should remain useful when the center folder is copied on its own. Verify: a test copies the published files (the EI.1.02 allowlist) to a temporary directory, serves them with a plain static server and, separately, opens index.html from disk; navigation, the embedded excerpts and local audio work in both, and the repository links are absent rather than broken in the copy.

8. Integrate real Eve — moved to the redesign tracker#

Moved on 18 September 2026 (owner decision). Integrating Eve into the Presentation Center is now the first priority, not something that waits for content coverage. The fifty items that stood here were written before the code had been read and several of their premises did not hold (the web assistant dock cannot be bundled; the BFF has no browser session; decks are single files with no slide-changed event; nothing can verify a slide edit today). They were rewritten against the code as Phase 0 of EVE-V1-REDESIGN-TODOS.md, which is the plan of record. Nothing below is tracked here any more.

What this section asked for Where it is now (Phase 0)
Discover and reuse the actual architecture; choose the host 0.0, EI.0.01–EI.0.03, EI.1.01–06
Pass guide, slide, notes and sources as page context EI.2.01–05, EI.3.01–07, EI.4.01–04
Grounded answers with citations; least trust for source text EI.5.01–05, EI.6.01–04
Notes bound to guide, slide and revision EI.7.01–05
Propose, preview and enact real slide changes; receipts; undo EI.8.01–06, EI.9.01–08
Add, remove, reorder slides; diagram and guide changes EI.10.01–03
Invalidate audio; regenerate HTML, MP3 and PDF; durations EI.9.05–07
Ad hoc presentations from a conversation (all four groups) EI.12.01–11
Required real integration evidence EI.11.01–06

The secret-handling rule that closed this section still binds Phase 0: source keys only into the process that needs them, check presence and length only, and record a live-testing gap rather than fabricating a run.

9. Final exhaustive Domains audit#

The user requested both complete coverage and visibility under Domains. Current product guides about a domain do not by themselves satisfy that final navigation and source-of-truth requirement. The pinned registry has 54 domains.

  • Compare the registry with all domain directories, code roots, planned roots and generated Docs Center discovery. Do not discover only directories with a features file; that can hide a missing canonical file.
  • For every domain, cover its canonical docs/domains/<name>/deep-dive/{features,architecture,specifications}.md, relevant supporting docs, contracts, APIs, events, ownership, product uses, implementation/planned state, failure behavior, operations and evidence.
  • Give every domain a clear canonical home under Domains, with sensible introductory/detail guides and links to its product/shared-system uses.
  • Retain planned-only domains with their actual status. Discover and add any new domains beyond the 54 in the pin before final acceptance.
  • Inspect and run tools/docs_center/domains.py and tools/docs/verify-domains-source-of-truth.mjs (pnpm domains:check) with their applicable tests. Preserve code-root allowlists and product/domain distinctions.
  • Resolve the 13 known preliminary failures (recheck current main first): eight generated HTML TODO mirrors were misclassified as authored TODO sources because the filename matcher lacks a Markdown extension filter; five root ledgers lacked registry ownership rows. Add regressions for the classification fix rather than broadening an exclusion to hide legitimate sources.
  • Ensure ownership for these five root ledgers if still missing: EVE_DEEP_TEST_AND_POLISH_TODOS_2026-08-04.md, EVE_EVERYWHERE_TODOS_2026-08-19.md, EVE_SMALL_MODEL_EXCELLENCE_TODOS_2026-08-16.md, EVE_SOTA_GAP_CLOSURE_TODOS_2026-09-01.md, V1_PRODUCT_GRAPH_RESTRUCTURE_TODOS_2026-08-04.md.
  • Keep this continuation ledger's ownership row in docs/domains/README.md. It maps presentation coverage to the actual product/domain owners and does not declare a new domain. No additional missing-ledger failure should be introduced.
  • Rebuild affected generated documentation, validate freshness and inspect Domains and Presentation Center navigation on desktop/mobile after changes.
  • Pass source-of-truth, source-unit teaching coverage and browser gates before claiming the Domains phase is finished.

Per-domain final review checklist (all remain pending for the expanded library):

  • aglaea — sources, complete teaching, Domains navigation, related guides and verification.
  • airmid — sources, complete teaching, Domains navigation, related guides and verification.
  • aja — sources, complete teaching, Domains navigation, related guides and verification.
  • aje — sources, complete teaching, Domains navigation, related guides and verification.
  • annapurna — sources, complete teaching, Domains navigation, related guides and verification.
  • aphrodite — sources, complete teaching, Domains navigation, related guides and verification.
  • arete — sources, complete teaching, Domains navigation, related guides and verification.
  • asase — sources, complete teaching, Domains navigation, related guides and verification.
  • athena — sources, complete teaching, Domains navigation, related guides and verification.
  • bellona — sources, complete teaching, Domains navigation, related guides and verification.
  • brigid — sources, complete teaching, Domains navigation, related guides and verification.
  • calliope — sources, complete teaching, Domains navigation, related guides and verification.
  • concordia — sources, complete teaching, Domains navigation, related guides and verification.
  • contracts — sources, complete teaching, Domains navigation, related guides and verification.
  • cybele — sources, complete teaching, Domains navigation, related guides and verification.
  • demeter — sources, complete teaching, Domains navigation, related guides and verification.
  • euterpe — sources, complete teaching, Domains navigation, related guides and verification.
  • freya — sources, complete teaching, Domains navigation, related guides and verification.
  • gaia — sources, complete teaching, Domains navigation, related guides and verification.
  • galatea — sources, complete teaching, Domains navigation, related guides and verification.
  • hathor — sources, complete teaching, Domains navigation, related guides and verification.
  • hestia — sources, complete teaching, Domains navigation, related guides and verification.
  • iris — sources, complete teaching, Domains navigation, related guides and verification.
  • isis — sources, complete teaching, Domains navigation, related guides and verification.
  • kalika — sources, complete teaching, Domains navigation, related guides and verification.
  • kuanyin — sources, complete teaching, Domains navigation, related guides and verification.
  • lakshmi — sources, complete teaching, Domains navigation, related guides and verification.
  • lilith — sources, complete teaching, Domains navigation, related guides and verification.
  • maat — sources, complete teaching, Domains navigation, related guides and verification.
  • maya — sources, complete teaching, Domains navigation, related guides and verification.
  • meditation — sources, complete teaching, Domains navigation, related guides and verification.
  • metis — sources, complete teaching, Domains navigation, related guides and verification.
  • mnemosyne — sources, complete teaching, Domains navigation, related guides and verification.
  • neith — sources, complete teaching, Domains navigation, related guides and verification.
  • nike — sources, complete teaching, Domains navigation, related guides and verification.
  • nisaba — sources, complete teaching, Domains navigation, related guides and verification.
  • nous — sources, complete teaching, Domains navigation, related guides and verification.
  • nyx — sources, complete teaching, Domains navigation, related guides and verification.
  • openapi — sources, complete teaching, Domains navigation, related guides and verification.
  • oshun — sources, complete teaching, Domains navigation, related guides and verification.
  • oya — sources, complete teaching, Domains navigation, related guides and verification.
  • phoebe — sources, complete teaching, Domains navigation, related guides and verification.
  • proto — sources, complete teaching, Domains navigation, related guides and verification.
  • psyche — sources, complete teaching, Domains navigation, related guides and verification.
  • saraswati — sources, complete teaching, Domains navigation, related guides and verification.
  • seshat — sources, complete teaching, Domains navigation, related guides and verification.
  • shakti — sources, complete teaching, Domains navigation, related guides and verification.
  • shared — sources, complete teaching, Domains navigation, related guides and verification.
  • sophia — sources, complete teaching, Domains navigation, related guides and verification.
  • tara — sources, complete teaching, Domains navigation, related guides and verification.
  • themis — sources, complete teaching, Domains navigation, related guides and verification.
  • uzume — sources, complete teaching, Domains navigation, related guides and verification.
  • veritas — sources, complete teaching, Domains navigation, related guides and verification.
  • yemaya — sources, complete teaching, Domains navigation, related guides and verification.

10. Final every-slide visual, UI/UX design and presentation-quality pass#

8 September Eve completion: the seven-specimen repair was only an initial step. The subsequent completion package finishes the listed design repairs across all 25 Eve guides / 204 slides: concrete product and Studio cases, authentic pinned component captures, source-accurate recovery/ADR/verifier relationships, continuous example identity, scoped typography, subtraction and complete sequence review. Four inherited architecture diagrams receive readable wrapped labels. Product-image panning preserves the current slide. The last category-only “Question” header now names its cost dimensions; remaining question headers contain actual questions.

  • Finish the twelve product-guide gaps in the earlier Eve review.
  • Review all thirteen architecture-guide sequences and fix discovered defects.
  • Record all 204 retain/change decisions and exact slide evidence in the Eve quality ledger.
  • Bind shipped Eve browser, source, narration and PDF evidence in the completion receipt.
  • Obtain willing-reader observations against the final Eve candidate and record acceptance or further revisions using the neutral reader protocol.

These scoped items do not close the whole-library overhaul below. The initial repair receipt remains a dated record of its partial scope. Model optical review and passing automation do not establish reader comprehension or an accepted visual benchmark.

The user requested an exhaustive diagram review and then extended the requirement to screenshots, UI mockups, extensive UI/UX design presentations and a meticulous quality review of every presentation. These are required continuation tasks, not completed work at this checkpoint. They apply to every final presentation and slide, including the old foundation, new product/domain/design guides and slides added by Eve or later content review. Add useful visual and design content while expanding the library, then perform the complete final review below.

Required content-led composition overhaul#

Superseded in part, 11 September 2026. The authoring and renderer contract this subsection asked for now exists as the Eve edition (section 0, "Design edition binding", and EVE-V1-REDESIGN-TODOS.md). The items below stay as review criteria; where they describe how to compose a slide, the redesign file governs. "Old defaults" below means the pre-edition skeleton, which no new slide may use.

Added 2026-09-06 in response to the user's two slide examples and explicit request to adopt the complete improvement approach. All unchecked items below are required implementation and review work. The scope includes every existing guide and reused slide instance, all future authored guides, standalone and embedded output, PDFs, and Eve's generated or revised presentations.

Baseline, priorities and coverage#

The source audit at commit 5883cfe9c9 found 33 of 38 guides opening with exactly four numbered columns, and 252 of 381 unique slides using columns or horizontal flows (137 columns and 115 flows). The remaining layouts were 83 tables, 24 layers, 21 graphs and one cover. These are dated baseline counts, not caps or target distributions. The audit uses catalog opening IDs and effective layouts after diagrams.json overrides.

  • Reproduce and record the audit against the current source revision: per-guide openings, point counts, ordered layout families, repeated silhouettes and runs of similar compositions. Count unique slides and reused guide-slide instances separately; retain the dated baseline above and record changes without claiming that numeric variety proves quality.
  • Audit the current causes in tools/presentations/build-eve-oshun.py, deck.css, content/*.json and diagrams.json: implicit column fallback, identical flow/sequence rendering, automatic numbering, uniform graph rectangles and content authored to fit those structures. Review both the authoring decisions and renderer limitations.
  • Track each guide and slide's communication problem, proposed treatment, implementation status, source review and current visual/test evidence in the existing review ledgers. Review unaffected-looking slides as well; retain a suitable composition with a recorded reason where appropriate.
  • Adopt this contract before expanding further with the old defaults. Establish the renderer and pilot while continuing independent source and narration work; use the new approach for subsequent authoring, then complete the migration and final full-library pass. The pilot is an implementation stage, never the completion boundary.

Author the meaning before selecting a composition#

  • Give every guide a clear audience, prerequisite knowledge, purpose and intended learning outcome. Choose a narrative arc appropriate to that purpose, with transitions that make the next slide necessary and a close that resolves the opening.
  • Write a short per-guide art-direction brief alongside the learning objective before selecting slide treatments. Describe the appropriate visual character, such as human and experiential, analytical, architectural, exploratory or instructional, and explain how it serves this subject and audience. These are examples, not compulsory categories. Translate the direction into concrete choices for imagery, composition, visual density and pacing. Preserve the shared Oshun palette, typography and accessibility standards while letting a meditation journey and a permissions architecture each express their subject appropriately. Keep the brief with canonical authoring records, outside audience-facing copy and narration, and update it when the guide's purpose or audience changes.
  • Create a brief for every slide before choosing its layout: one narrative job, one primary takeaway, supporting evidence or concrete example, the relationship to explain, intended emphasis and reading order. Record whether relationships are sequence, choice, comparison, recurrence, containment, dependency, ownership, convergence, change or another source-supported structure. Keep these planning fields out of visible copy and narration.
  • Let meaning determine element count and composition. Preserve necessary distinctions, including genuine four-part concepts; do not pad, split or merge ideas to meet a layout's slot count. Remove automatic numbering from unordered concepts and equal visual weight from unequal ideas.
  • Redesign each guide opening around the question, concrete example, tension, claim or system view that best introduces its subject. Use four introductory categories only when they are the strongest introduction. Do not substitute another universal opening formula.
  • Use takeaway-led titles and concise visible support. Put explanatory depth in notes or dedicated detail slides, remove redundant copy and split overload before shrinking text. Preserve source-backed substance and accessible explanations while improving scanability.
  • Judge each composition by whether a viewer can identify the important relationship before reading every paragraph. Require a clear focal point and deliberate hierarchy; do not confuse uniform columns, empty space or extra decoration with clarity.

Required repertoire and selection guidance#

Implement and document reusable treatments for the communication jobs below, using actual center content to demonstrate their use. These are available choices, not a mandatory rotation or a requirement to use every family in each guide. Extend the repertoire when an idea needs a different expression.

Communication job Suitable treatment and semantic requirement
One central claim Large statement with one concrete supporting example or evidence anchor
A meaningful distinction Paired comparison with aligned attributes and explicit differences
A decision Branching path with labeled conditions and clear outcomes
A recurring journey Open loop with an explicit, explained return relationship
Responsibilities across actors or services Swimlanes with ownership and labeled handoffs
Containment, scope or hierarchy Nested regions, layers or hierarchy with clear boundaries
Several inputs producing one result Converging paths with the result emphasized and input relationships explained
Change over time Timeline with meaningful milestones and accurate interval semantics
A mechanism in practice Annotated example, artifact or relevant screenshot with focused callouts
A tradeoff Matrix or plotted comparison with justified dimensions and supported values
  • Retain straightforward lists, tables and linear sequences where they genuinely explain the content. Use columns for parallel peers, arrows for supported direction and timelines for time; avoid implied ordering, causality, proportionality or runtime behavior that the sources do not establish. Label schematic spacing when it does not encode elapsed time.
  • Vary composition, emphasis, density and pacing through the guide as the argument develops. Review adjacent slides and complete thumbnail sheets; fix repetitive openings and monotonous silhouettes with editorial judgment. Do not randomize layouts, impose variety quotas or create a new repetitive template cycle.

Minimalist elements and flowchart craft#

  • Preserve cream, ink, terracotta, Georgia/Arial, generous spacing and calm presentation chrome. Establish shared typography, spacing, line-weight, node, connector and emphasis rules in DESIGN.md and the renderer's style contracts. Use scale, position, alignment and contrast to establish importance while keeping the existing visual identity coherent.
  • Give the minimal element vocabulary explicit meanings: thin rules separate ideas; brackets group them; dots mark events; arrowheads show direction; connector labels identify requests, permissions, evidence or saved state; enclosing outlines identify ownership or scope. Use terracotta sparingly to emphasize the key relationship or conclusion, with text/shape cues so color alone never carries meaning.
  • Use rectangles when they represent an actual entity or operation; support suitable event, decision, boundary and text treatments instead of drawing every concept as an identical box. Add no ornamental icons, background effects or card mosaics that fail to explain the idea.
  • Design flows around explicit start/end states, decisions, conditions, optional paths, handoffs, merges, failures and recovery where relevant. Make the principal path and any return relationship readable; distinguish a worked scenario from a required sequence or complete execution model.
  • Label meaningful connections and branch outcomes. Route connectors deliberately with consistent attachment points, visible arrow direction, space around junctions and adequate separation from text. Draw edges behind nodes; avoid unexplained crossings, ambiguous merges, orphaned nodes, obscured arrowheads and lines through labels.
  • Provide node sizing and label wrapping that preserve legibility. Split complex diagrams into overview/detail slides when needed, preserving their relationship and references; never fit everything by shrinking labels or deleting important conditions.
  • Preserve graph meaning across responsive layouts and accessible text alternatives: branches, conditions, direction, ownership and returns must survive mobile reflow and screen-reader descriptions. A flat list of node names alone is insufficient for a relational diagram.
  • Use progressive reveals only when they help follow a complex explanation. Keep the complete meaning available without animation, in reduced-motion mode, transcripts and static PDF output; preserve keyboard navigation, focus and narration synchronization if reveals are added.

Required redesigns of the two supplied examples#

  • Redesign member-day (Arrive / Practice / Apply / Explore / Return), including its use in psyche-lilith and v1-web-pwa. Show one illustrative session with optional activities and a spacious return relationship labeled Library, history and permitted memory connecting back to arrival. The existing notes explicitly say this is not a mandatory sequence: Tara, Arete, Nyx and Nisaba must not appear to be compulsory consecutive steps. Preserve release availability, consent, interruption and honest continuity boundaries; do not imply hidden memory capture or fabricated history. Validate the final path against the pinned sources and explain its meaning in the narration.
  • Redesign v1-web-pwa-boundary (Route / Release / Capability / Outcome) around the claim A completed journey needs its own evidence. Retain all four distinct questions, pair them with the corresponding evidence in a compact comparison and emphasize the completed outcome. Verify the evidence examples against the sources. Do not imply that route presence, release admission or service capability proves completion, or turn the distinctions into a funnel or sequential gate model without source support.
  • Capture before/after renders and a short rationale for both examples, documenting the relationship made clearer, source fidelity, reading order and accessibility. Check every reused guide context and current standalone, embedded, compact, mobile and PDF outputs.

Shared implementation, pilot and complete adoption#

  • Extend the canonical authoring schema and renderer with explicit composition intent, emphasis, relationships, labeled edges, decisions, groups/ownership and accessible descriptions as required by the repertoire. Validate supported values, required fields and references; unsupported layouts must produce an actionable authoring error instead of silently falling back to generic columns. Migrate legacy content deliberately and keep generated HTML reproducible.
  • Separate semantic relationships from rendering geometry. Support appropriate node sizes, routed connectors, grouping and responsive/print adaptations without making authors express every idea as a list of title/description pairs or a fixed rectangle grid. Keep diagrams source-reviewable and editable through native SVG/CSS and shared data.
  • Build a representative pilot of roughly twelve existing slides covering both examples, several guide openings, an architecture/ownership explanation, a decision, a failure/recovery case, a comparison and an annotated concrete example. Let coverage determine the exact pilot size; use real content and include a full guide sequence to assess rhythm. (Reconciled 2026-09-18 with the design edition binding in section 0: the pilot exists — chapter 1, Meet Eve — and section 0 and section 13 both say not to establish another. Do not build a second pilot from this box. It stays open only as the review criterion the 11 September note describes: the accepted specimen must cover the slide kinds listed here, and a kind it does not cover is added to the edition under the redesign file, not to a new pilot.)
  • Use the pilot to establish the visual benchmark, typography rules and representative-reader review described under the craft/comprehension standards below. Exercise all eight areas on suitable real examples; use the same pilot and evidence records rather than a disconnected second redesign. Resolve material findings before propagating its treatments. (Reconciled 2026-09-18: "the pilot" is chapter 1, Meet Eve, per the design edition binding; the benchmark, typography rules and reader review are measured on edition chapters, as section 13 step 2 says.)
  • Implement and review the pilot end to end, resolve source, design and accessibility defects, and document the accepted treatments and selection guidance in DESIGN.md and maintainer documentation. Then migrate every existing guide and enforce the same contracts for new authoring and Eve generation/revision. A renderer-only update does not complete the editorial migration; a pair of redesigned examples does not complete the renderer or library adoption. (Reconciled 2026-09-18: the first sentence is delivered by the accepted specimen and is not to be repeated; what remains open here is the second — migrating every existing guide — and the plan of record for that is Phase D of EVE-V1-REDESIGN-TODOS.md.)
  • Preserve stable IDs, deep links, foundation coverage, source mappings, prerequisites and shared-slide ownership. Give additional slides stable IDs and update guide order and coverage. Invalidate affected review records; regenerate HTML, PDFs, scripts, changed narration and duration totals under the existing freshness rules. Reuse audio only when its content/fingerprint and spoken references remain valid.

Verification and completion criteria for the overhaul#

  • Add meaningful schema/renderer tests for each new composition and relationship type, invalid layouts/references, optional branches, recurrence, labeled handoffs, grouping and accessible alternatives. Include long labels, uneven content and legitimate four-part comparisons; verify semantics and failure behavior, not only markup snapshots.
  • Extend deep Playwright coverage for the pilot and all affected flows in standalone and embedded presentations at desktop, compact and mobile sizes. Check overflow, clipping, label/connector collisions, reading order, accessible names/descriptions, keyboard/focus, reduced motion, navigation, narration and any reveal/detail interactions. Exercise image loading and unavailable assets where annotated examples use them.
  • Verify PDF/static rendering for every treatment: complete content, readable labels, preserved relationships, correct page sequence and absence of motion-dependent omissions. Run the existing applicable builder, contract/unit, lint/format, accessibility, performance and browser/PDF gates under the repository's resource-safety rules.
  • Review contact sheets of every complete guide for meaningful openings, repeated silhouettes, emphasis and rhythm, then inspect every final slide and PDF page at readable size. Record conceptual accuracy and visual clarity separately from automated geometry/accessibility results. Passing counts, screenshots or tests alone do not establish expressive quality.
  • Require a per-guide pass for source-faithful relationships, meaningful opening, appropriate composition and point count, clear focal point, readable connectors, coherent pacing, accessible mobile/print expression and current narration. Record reviewed retain-as-is decisions and resolve every material defect; no aggregate score may hide a failing guide.
  • Complete the audit and current evidence for every final guide and slide instance, including revised shared contexts and Eve's generation and editing evaluations. Keep the overhaul open until authoring, renderer, migration, documentation and verification are all complete; adding this checklist, passing the pilot or increasing layout counts is not completion.

Required visual craft and comprehension standards#

Added 2026-09-06 after the user's request to comprehensively adopt the eight additional quality recommendations. These standards extend the composition overhaul above. They apply to the existing library, new guides, reused slide contexts and Eve-generated or revised presentations, with format-specific evidence for standalone HTML, the embedded center, mobile and PDFs. The reader review has the explicit sampling scope below; do not claim every future ad hoc deck has been tested with people.

1. Concrete visual quality benchmark#

  • Build a small, versioned reference collection from the completed pilot using real Oshun content. Include polished examples of typography, diagrams, comparisons, quantitative charts, imagery and a complete guide sequence demonstrating pacing. Select enough examples to cover the relevant communication jobs and supported formats without turning the collection into a compulsory template catalog.
  • For each reference, retain its editable authoring data, stable IDs, renderer/style/font versions, source and asset provenance, final renders and review evidence. Document the intended audience, takeaway, focal point, reading order, spacing and the relationship made clearer. Make the collection reproducible and accessible to maintainers and Eve through the supported artifact contract.
  • Include clearly marked rejected or weaker alternatives paired with the selected treatment. Explain specific defects and improvements: hierarchy, unsupported arrows, text wrapping, label size, crop, clutter or weak continuity. Keep these instructional examples outside audience-facing guides unless the guide is teaching design itself.
  • Calibrate the quality rubric against the collection. Compare candidate slides at the same viewing size and equivalent content difficulty; record concrete reviewer reasons and resolve material disagreements in the criteria. An automated similarity score or resemblance to a reference is insufficient evidence of appropriate design.
  • Document when to adapt a reference or introduce a new composition. Accept references only after factual, visual, accessibility and applicable reader-review findings are resolved. Version changes and invalidate affected examples when renderer behavior, tokens, fonts, sources or teaching claims change; preserve historical evidence as historical.

2. Typography and optical alignment#

  • Define a complete type and spacing specification in DESIGN.md and shared renderer tokens: title, subtitle, section label, body, diagram label, caption, citation, numerical value and code/exact identifier roles. Specify sizes, weights, line heights, text measures, emphasis, grid, margins and spacing relationships for each supported viewing context. Preserve the established Georgia/Arial and cream/ink/terracotta identity.
  • Perform an optical pass on every final slide: align perceived edges and centers, balance text with adjacent diagrams/images, and inspect hanging punctuation, uneven gaps, isolated short lines, awkward title breaks, widows/orphans and inconsistent baselines. Use repeatable component rules and document justified optical offsets instead of accumulating unexplained slide-specific positioning patches.
  • Establish consistent capitalization, punctuation, quotation marks, dashes, symbols, units, dates, precision and number alignment. Protect meaningful number/unit pairs from awkward breaks and keep exact API, code, path and identifier text unambiguous. Shorten visible prose or use a detail treatment without altering required technical meaning.
  • Declare font availability and fallbacks for supported hosts and offline use. Verify font loading before capture/export, missing-font behavior, layout shifts and reflow with the actual fallback stack. Check permitted font embedding/substitution, glyph rendering and selectable/searchable text in PDFs; do not assume an authoring-machine font is installed on another reader's machine.
  • Include realistic typography stress cases: long domain names and labels, multi-line titles, mixed prose and identifiers, large/small numbers, accented Oshun spellings and mathematical symbols where used. Test text enlargement and narrow layouts without clipping or hiding content; include supported localization variants when that scope is introduced.
  • Add appropriate font-readiness, fallback, overflow and text-preservation checks. Pair them with full-size and actual-viewing-size visual review; automated bounding boxes cannot establish optical balance. Fix overload by rewriting, recomposing or splitting before reducing effective type sizes below the documented readable minimums.

3. Charts and quantitative comparisons#

  • Define chart selection guidance for comparisons, trends, distributions, relationships and parts of a whole. Use a table or qualitative comparison when that communicates more accurately; do not invent scores, rankings, percentages or plotted values to fill a visual treatment.
  • Give every quantitative visual traceable data and computation provenance: source/revision/date, population or cohort, period/timezone where relevant, numerator/denominator, units, transformations, aggregation, rounding and sample size where applicable. Keep concise labels on the slide and full reproducible detail in source context. Distinguish measured, estimated, forecast and illustrative values explicitly.
  • Establish honest scale rules: bar lengths use a meaningful zero baseline; any necessary scale break, truncated line-chart domain or logarithmic scale is explicit and justified. Comparable panels share scales or clearly explain differences. Match mark length, position and area to the intended quantity; avoid decorative 3D/perspective and misleading area or volume encodings. Do not equate correlation with causation.
  • Preserve missing, unknown, unavailable, not-applicable and zero as distinct states. Show relevant uncertainty, ranges and limitations with an explanation of what they mean; never invent error bars or imply more precision than the evidence supports. Explain rounding discrepancies and changing denominators when they affect interpretation.
  • Use readable direct labels where practical, aligned units and concise annotations that connect the evidence to the takeaway. Keep legends near their marks, maintain categorical/series identities across slides and use restrained semantic color plus non-color cues. Provide an accessible description and a usable data table or equivalent text representation.
  • Extend data/schema and renderer tests for calculations, label/value agreement, scales, negative/zero/missing values, long categories and uncertainty cases. Inspect HTML/mobile/PDF outputs against the actual source data; validate underlying totals and comparisons, not merely that a chart rendered or matched a snapshot. Recheck affected narration and cited conclusions when data or transformations change.
  • Maintain a source-linked mapping for recurring entities, service/actor names, symbols, ownership lanes, series and semantic color roles. Reuse those identities consistently within a guide and related learning routes; do not let the same visual cue silently acquire a different meaning. Document intentional distinctions without assigning every domain a new decorative color scheme.
  • Connect overview, detail and worked-example slides visibly. Preserve recognizable context and stable relative positions where the topology permits; identify the region or path being examined with restrained highlighting or a contextual inset when useful. Explain justified rearrangements rather than preserving positions that misrepresent the actual relationship.
  • Carry a coherent worked case through successive slides when it helps teaching. Keep actors, example values, terminology and states consistent; explicitly introduce a different scenario, branch or source revision. Ensure narration and reading order track the same entities and changes.
  • Frame before/after and alternative comparisons comparably: viewport, scale, crop, data, state and other relevant conditions should match, with substantive differences disclosed. Do not make an option look better by giving it a larger image, cleaner example or incompatible measurement.
  • Add identity/reference validation where the shared representation can check it, and inspect consecutive-slide strips plus overview/detail pairs manually. Review reused slides in each local context and test any highlight/reveal transitions in reduced-motion and static output. A shared entity or example change must invalidate its dependent reviews.

5. Image art direction and representative outputs#

  • Give every image a specific teaching job and focal subject. Plan its placement, aspect ratio, crop, scale and relationship to the text before sourcing or capture. Establish coherent framing, caption and annotation treatments for actual UI, proposed UI, photographs, scenes and generated artifacts while allowing the subject to determine the appropriate form.
  • Where a guide explains generated imagery, a game world, a scene or another creative artifact, show representative real output when available and useful. Identify relevant model/workflow/version or capture conditions and whether the example is representative, selected, exploratory or proposed. Preserve unavailable-output gaps honestly; a polished illustration must not masquerade as a captured product result or measured typical quality.
  • Capture enough resolution for the intended output and inspect the final resized asset. Avoid blur, compression artifacts, accidental stretching, distracting interface chrome and crops that remove critical context. Balance local asset size, decoding cost and visual quality against the existing performance budgets. Retain an original and documented crop or derivative so adjustments remain reproducible.
  • Keep callouts sparse, accurate and anchored to what they explain. Match annotation type, stroke weight and contrast to the slide language; keep text readable and do not cover the subject. Recompose captions and callouts for narrow displays and PDF as necessary, with accessible descriptions explaining the same focal detail.
  • Preserve the existing asset provenance, rights/status and implementation labeling requirements. Record material cropping or editing, retain source links/hashes and do not retouch evidence in a way that changes its claim. Use matching conditions for image comparisons and verify that captions, annotations and narration agree with the actual final asset.
  • Review both complete-slide composition and the image at its displayed size. Test loading, broken/absent assets, responsive derivatives, aspect ratios and print output; provide useful fallback text without claiming visual review passed when the required asset was missing.

6. Readability at actual viewing sizes#

  • Declare the supported viewing matrix and effective slide dimensions: embedded center with navigation expanded/collapsed, relevant panels open, standalone presentation, compact desktop, mobile orientations and PDF page-fit/normal reading conditions. Record viewport, browser zoom, device pixel ratio and export scale with evidence. Add projection or screen-share conditions when those are intended delivery contexts.
  • Establish and verify practical minimum effective sizes for text, captions, diagram labels, connectors and arrowheads in each context. Measure the final displayed result after canvas transforms or fitting; large source-canvas font values do not prove the embedded result readable. Use the pilot and reader feedback to justify thresholds instead of choosing numbers solely to make an existing layout pass.
  • Inspect normal browser zoom and actual embedded captures, not just enlarged full-resolution exports. Check that terracotta rules, fine strokes, screenshots, annotations and dense comparisons survive downscaling with sufficient contrast and separation. Also exercise supported text/browser enlargement for accessibility; normal-size review does not replace it.
  • Adapt density and composition where available space is insufficient: reflow labels, split a diagram or offer a clearly indicated detail view with accessible controls. Preserve an understandable main takeaway without requiring zoom, hover or repeated panel closing. Avoid shrinking the entire slide into unreadability or hiding content to satisfy bounds.
  • Extend Playwright to exercise supported workspace/panel states, font readiness, effective label sizes, clipping, scroll reachability and detail navigation. Inspect PDFs for page-fit legibility, font substitution and stroke survival, as well as full-size geometry. Keep meaningful content accessible in static exports without depending on browser-only controls.
  • Save labeled evidence at the actual reviewed scale and record any unsupported context precisely. Require all declared viewing contexts to pass; a desktop screenshot or a PDF inspected only at extreme zoom cannot close a mobile or embedded-readability defect.

7. Representative-reader comprehension review#

  • Define a practical qualitative review for the pilot, major new visual treatments and material changes to previously validated explanations. Select available, willing readers representative of the intended knowledge level and viewing context. Cover novice/expert or accessibility needs where relevant; document recruitment limitations and sample scope without treating a small convenience sample as statistical proof.
  • Write neutral tasks and expected source-backed interpretations before sessions. Ask readers to identify the main point, explain the depicted relationships, follow a worked case, recognize optional/failure paths and distinguish evidence from assumptions. Allow a natural first look and normal reading; do not coach the answer or require arbitrary speed scores.
  • Include slide-only understanding and the intended narrated/interactive experience where applicable. Observe where readers hesitate, misread an arrow, infer an unsupported sequence, overlook a qualifier or need help locating the detail. Aesthetic preference can inform refinement, but it cannot compensate for a misunderstood claim or inaccessible treatment.
  • Record the exact guide/slide/artifact versions, viewing setup, task, expected meaning, observations, assistance provided and result. Capture only the participant information needed for the review; keep real reader observations separate from author interpretation, model simulations, automated checks and product-usability studies of the UI being depicted.
  • Turn findings into specific defects and revisions, then retest affected explanations and relevant transfer examples. Resolve material comprehension failures before promoting a treatment as a benchmark or scaling it through the library. Record what was retested and what the evidence supports; do not fabricate participants, quotations or successful outcomes when readers are unavailable. Keep missing required evidence explicitly pending while independent implementation work continues.
  • Define inheritance and invalidation of findings: unchanged uses of a reviewed treatment can use that design evidence with per-guide review, while new semantics, material density changes or different audience needs may require renewed reader testing. Calibrate Eve evaluations using these findings, but never label every generated deck as human-tested or imply a representative sample establishes understanding of every guide.

8. Editorial subtraction and final polish#

  • Give every visible title, subtitle, number, rule, arrow, caption, takeaway, footer, badge and decorative element a clear job: communicate, orient, provide evidence or support an actual action. Inspect each for duplication and competition with the focal idea. Remove unnecessary numbering, redundant borders, repeated statements and decorative chrome that adds no useful meaning.
  • Review title/subtitle/takeaway combinations together. Let each contribute distinct value and vary their presence when the idea does not need all three; update authoring validation and rendering where rigid requirements currently force filler. Preserve required accessible slide names, orientation, useful navigation and source access.
  • Preserve substantive teaching coverage when simplifying. Move supporting detail into accessible notes, source context or a dedicated detail slide with an explicit mapping; do not delete essential qualifiers, conditions, units, limitations or citations for a cleaner appearance. Reconcile foundation/source coverage and keep the visible claim accurate on its own.
  • Compare before/after versions at actual viewing size and in the guide sequence. Retain an element when removal harms comprehension, navigation or evidence. Let useful whitespace remain, but do not replace information with arbitrary blank expanses or make every slide equally sparse.
  • Inspect narration and transcripts after subtraction. Remove obsolete spoken references to labels, numbers or arrows, preserve the teaching explanation and synchronize any changed reveals. Regenerate affected audio/PDFs and repeat invalidated source, visual and reader reviews.
  • Record the final subtraction pass in each guide's quality ledger with material changes and retained-element rationale where needed. Verify the resulting static and interactive behavior through applicable schema, builder, navigation/accessibility, source, narration and PDF checks.

Adoption, evidence and final acceptance for all eight standards#

  • Establish the benchmark, typography specification and reader-review protocol with the existing pilot first; use its chart, image, continuity, actual-size and subtraction results to refine the complete standard. Promote only treatments with resolved material findings, then apply all eight areas during the whole-library migration and new authoring.
  • Add per-guide rubric entries for all eight standards and reference the supporting artifact hashes, benchmark version, data/image provenance, viewing matrix, review date, reviewer, defects and retests. Record a justified not-applicable entry when a guide has no chart or image, and the exact scope of inherited reader evidence; do not add unnecessary visuals merely to satisfy a checklist.
  • Implement the necessary shared tokens, authoring/schema contracts, renderer/export behavior and Eve guidance/evaluations. Add focused tests for changed behavior and run all applicable existing gates under the repository's resource limits. Preserve deterministic artifact generation, offline portability, accessibility and current source/audio evidence.
  • Keep measurable verification, aesthetic/editorial judgment and reader comprehension as separate evidence categories. Require a pass in every applicable area, with no unresolved material defects or missing required reader evidence hidden behind average scores, reference similarity or screenshot counts. Record exclusions and limitations honestly.
  • Finish current evidence for every final guide and reused context, updating DESIGN.md, the reference collection, maintainer guidance, Eve contracts, verification records and this checklist together. Recheck affected results after font/token/data/asset/layout/content changes. The eight-standard work remains incomplete until its implementation, adoption and scoped verification are complete; this documentation update alone closes none of those future execution tasks.

Every-slide diagrams and visual explanation#

  • Create a review ledger keyed by every current slide ID and content hash. Review all slides, not only those without an existing SVG or those recently edited.
  • Ask what relationship, process, choice, state, hierarchy, dependency, owner, comparison or worked example a diagram could make easier to understand.
  • Add diagrams wherever useful: architecture/component maps, sequences, data/authority flows, state transitions, decision trees, hierarchies, lifecycles, ownership boundaries, cause/effect and worked journeys. Use readable tables for comparisons when they communicate better than arbitrary arrows.
  • Add separate diagram slides when the explanation deserves room. Split busy diagrams and overcrowded slides; do not merely shrink labels or crowd the page.
  • Enrich narration to explain how to read the diagram, what the arrows/states mean and which path the worked case follows. Distinguish composition diagrams from claims that every runtime request traverses every box.
  • Keep visual status faithful: planned components and provider/integration gaps must remain visible; diagram polish is not a shipment certificate.
  • Use the established cream/ink/terracotta, Georgia/Arial design language and native SVG/CSS for technical diagrams. Introduce imagery only when it adds meaning; do not turn the technical center into decorative card mosaics.
  • Check all node/edge references, ownership direction, non-overlap, label fit, reading order, contrast, legend/status, screen-reader title/description and mobile explanation. Preserve useful context when desktop diagrams reflow.
  • Review actual desktop, compact desktop, embedded center, mobile and PDF renders. Automated bounding boxes are necessary but do not prove conceptual clarity. Open full-size images if a contact sheet appears cropped or unclear.
  • Extend graph/visual regression checks for new diagram structures and interactions. Keep reduced motion, keyboard navigation and focus behavior.
  • Regenerate affected narration/PDFs and update source coverage, assignments, learning order, counts and ancestor durations for all new/reordered slides.
  • Repeat any invalidated source and visual reviews and the final complete verification suite. This pass is finished only when every final slide has been reviewed against its current content and there are no unresolved visual defects.

Relevant screenshots, UI mockups and additional visual slides#

  • Inventory the interfaces and user journeys taught by each guide. Map each appropriate UI section to actual screenshots, proposed mockups or both; record a rationale where neither would improve understanding. Include existing guides, web/PWA, mobile/watch, admin, studios and game/spatial interfaces where relevant.
  • Add dedicated screenshot and mockup slides wherever a visual needs room. Show enough of the surrounding interface to orient the viewer, then use legible crops, numbered callouts and detail slides to explain the important controls.
  • Capture implemented screens from a reproducible build and representative test data. Record the route/screen, product/release, revision, viewport, UI state, capture date and asset path; reconcile differences from the pinned source. Remove private data and secrets before publishing captures.
  • Clearly label each visual as implemented UI, proposed mockup or exploratory alternative. Link it to its source or design decision, preserve release limitations, and never present a polished concept as a shipped feature.
  • Show each journey's actual structure: entry/context → user action → feedback/result → next step where ordered, with explicit choices, optional paths and return relationships where applicable. Follow the content-led composition contract above. Include relevant loading, empty, error, recovery, permission, offline and success states, plus responsive or input-mode variants.
  • Give every screenshot/mockup a meaningful caption, accessible description and narration explaining what to notice, how it works and why it matters. Ensure callouts and spoken references identify the same elements in order.
  • Keep assets local and portable, with sufficient resolution and sensible compression. Verify legibility, contrast, cropping and captions in standalone, embedded, desktop, mobile and PDF output; preserve readable text alternatives.
  • Extend the visual ledger with asset provenance, hashes, guide/slide mappings and review results. Extend Playwright checks for image loading, missing assets, layout, accessible alternatives and any zoom/detail interactions; refresh stale captures and affected narration/PDFs when the represented interface changes.

Extensive UI/UX design presentations and decision walkthroughs#

  • Create a discoverable UI/UX and Design Systems learning route with an overview, focused foundations/composition guides and detailed product/domain design walkthroughs. Cross-link them from the relevant feature presentations; determine depth and guide count from actual design coverage, without a fixed cap.
  • Build a coverage matrix of appropriate UI surfaces, journeys, components and material design choices → dedicated guides → explanatory slides → source evidence/mockups → review status. A generic design overview alone does not satisfy extensive coverage of the particular interfaces and their decisions.
  • Research current, state-of-the-art UI/UX practice during implementation, using authoritative design-system, accessibility and interaction references. Cite sources and dates, explain applicability and tradeoffs, and distinguish evidence-backed practice from stylistic preference or an untested hypothesis.
  • Teach atomic design: atoms → molecules → organisms → templates → pages with worked Oshun examples. Trace a control into a composed interaction, a larger interface section, a layout and a populated screen; map these to actual components where available. Explain responsibilities and reuse boundaries, including when another composition model better fits the interface.
  • Cover foundations and design tokens: semantic color, typography, spacing, grids, sizing, iconography, elevation, motion and themes. Show how tokens support consistency, accessibility and variation across brands, products and platforms.
  • Cover user goals, information architecture, navigation, hierarchy, affordances, progressive disclosure, cognitive load, feedback, consistency, error prevention/recovery and useful microcopy. Demonstrate each principle on the particular screen or journey rather than listing terminology alone.
  • Cover component anatomy, variants, states, composition and interaction contracts; responsive/adaptive layouts; touch, keyboard, controller or spatial input where relevant; localization, long content and realistic data density. Explain which behaviors stay consistent and which adapt to the platform.
  • Include accessibility and inclusive design throughout: semantic structure, reading/focus order, contrast, target sizes, reduced motion, screen-reader behavior, alternatives to visual-only feedback and understandable recovery. Explain how each illustrated design supports and verifies these needs.
  • Present material choices as a clear sequence: user problem and constraints → alternatives with comparable mockups → evaluation criteria and evidence → selected or proposed direction → tradeoffs, rejected options and open questions. Link existing decisions; distinguish accepted choices from proposals needing validation, and avoid inventing research findings or decisions.
  • Use annotated wireframes, polished mockups, before/after comparisons and interaction/state diagrams as appropriate. Include full worked journeys and dedicated detail slides, with narration connecting visual choices to user needs.
  • Cover design-to-implementation handoff, component reuse, token ownership, documented exceptions and design drift. Relate designs to actual code and contracts where they exist and identify planned implementation explicitly.
  • Document usability questions, evaluation methods, observed findings and remaining hypotheses. Keep automated accessibility/interaction checks distinct from human usability evidence; review the resulting guides for practical depth, source fidelity and coherent placement in the broader learning sequence.

Every-presentation editorial, narration and organization review#

  • Maintain a review ledger for every guide, including its ordered slide IDs and hashes, narration/visual asset hashes, reviewer, review date, defects, fixes and final result. Review reused slides in each guide's local context; adding, editing or reordering content must invalidate the affected review.
  • Read and watch every presentation from beginning to end as a first-time viewer would. Establish audience, prerequisites and learning outcome; open with purpose/context, build concepts in a logical sequence, use worked examples and close with useful takeaways and a clear connection to the next relevant guide.
  • Check every transition within and between sections: explain why the next idea follows, introduce terms before using them, maintain consistent examples and remove abrupt jumps, unexplained references, contradictions and repetition.
  • Rewrite narration for natural, human-like spoken delivery: direct language, varied sentence structure, concrete explanations and purposeful transitions. Remove robotic list-reading, repetitive stock openings, empty enthusiasm, jargon piles and narration that merely reads visible text without teaching it.
  • Listen to every final recording in guide order at normal speed, alongside its slides. Check warmth, intelligibility, emphasis, intonation, pronunciation, sentence rhythm, pauses and continuity; fix clipped words, rushed passages, monotonous delivery and unnatural silence. Mechanical pause checks and a few representative audio samples do not satisfy this complete listening review.
  • Ensure narration accurately guides attention through screenshots, mockups, diagrams and comparisons. Match the visible reading order, captions, labels, claims and takeaway; explain enough context for someone using the transcript or listening without being able to inspect every visual detail.
  • Balance depth with pacing and cognitive load: one clear teaching purpose per slide, enough time to understand complex visuals, useful section breaks and recaps. Split overloaded slides or guides while preserving their substance.
  • Audit organization everywhere: center hierarchy, guide titles, section headings, slide order/numbering, prerequisites, breadcrumbs, related links, terminology, scripts, source panels and PDFs. Eliminate orphaned guides, dead ends, inconsistent naming and accidental duplication across reading routes.
  • Review every final slide and PDF page for composition, typography, density, alignment, spacing, legibility, brand consistency and visual balance. Inspect full-size renders of uncertain details and resolve all material defects.
  • Review each guide against its current art-direction brief. Confirm that the visual character suits its audience and subject, that imagery, composition and pacing carry the direction coherently through the guide, and that individual slides retain accurate meaning and the shared Oshun identity. Record mismatches and fixes in the existing quality ledger.
  • Apply the content-led overhaul's per-guide acceptance criteria above. Review openings and complete thumbnail sequences for repeated silhouettes and weak emphasis, then inspect full-size diagrams for the actual relationships. Keep every unresolved composition or flowchart defect in the guide's quality ledger until fixed and reverified.
  • Complete all eight visual craft/comprehension rubric entries above, including comparison to the benchmark, typography/optical refinement, faithful charts, visual continuity, image art direction, actual-size readability, scoped reader evidence and editorial subtraction. Resolve every applicable failing area and repeat invalidated reviews against the exact final artifacts.
  • Record a per-guide quality rubric covering factual accuracy, teaching depth, story/flow, organization, visual/UI explanation, natural narration, accessibility and delivery consistency. Require an explicit evidence-backed pass in every category; content counts, successful builds or a high average score cannot compensate for an unresolved failing category.
  • Regenerate changed recordings, scripts, HTML, PDFs, catalog/coverage maps and measured durations; rerun affected checks and repeat invalidated listening, visual and editorial reviews. Mark the library reviewed only when every final presentation passes against the exact artifacts being delivered.

11. Final verification, documentation and publication#

Required current checks#

  • Whole-monorepo coverage reconciliation passes for every repository root, subsystem and meaningful documentation/code/configuration/workflow/asset unit; no unreviewed architecture, implementation, infrastructure or CI/CD family remains hidden behind a passing documentation-only inventory.
  • All generation-system/workflow families have reviewed high-level and granular guides, including ComfyUI node walkthroughs and agentic execution traces; worked examples and relevant automated gates have current evidence, with implementation status and any missing live evidence stated accurately.
  • All eleven cross-cutting tracks have complete teaching matrices and reviewed guide/slide assignments, coherent navigation/prerequisites and current source, visual, narration and editorial evidence. Check each track explicitly; a broad whole-monorepo coverage total cannot hide an omitted learning route.
  • Extension recipes, onboarding exercises, algorithm examples, lifecycle walkthroughs and operational case studies have reproducible expected results and appropriate current automated verification. Record any missing live or human-evaluation evidence honestly and resolve all material teaching defects.
  • Presentation Center maintainer guides accurately describe the delivered pipeline and have been checked against the actual authoring, regeneration, freshness/coverage, preview, failure-recovery and publishing workflows.
  • Unit tests: source escaping/excerpts, graph contracts, catalog and prerequisite coverage, foundation retention, duration aggregation, stale media, resource/batch recovery, pronunciation, speech grouping and encoded pauses.
  • Inventory integrity/freshness and assignment validation; full teaching gate.
  • Builder freshness for delivered portal/decks/catalog/maps/scripts and the ignored whole-bank fixture. Never test one artifact then ship a rebuilt variant without establishing that the relevant behavior remains verified.
  • Ruff, scoped ESLint and Prettier; appropriate typechecking/build/tests for any actual app/service changes from Eve integration or other scope.
  • Full narration inventory/fingerprint/hash/duration/decode/pause check.
  • Full standalone bank Playwright: every slide on desktop/compact/mobile, accessibility, sources/transcripts, navigation/dialogs/history, print bounds and PDF sequence, actual playback/seek/speed/end transitions for every track, HTTP failures/races and portable-file behavior.
  • Full Presentation Center Playwright: all catalog/source/prerequisite links, performance budgets, navigation/search/history, hierarchy/timing, collapse, mobile drawer and focus, every guide slide on desktop/mobile, actual guide start/end/replay/stop boundaries, all PDF exports, no runtime errors and stable tested artifact hashes.
  • Real Eve service/integration/contract/evaluation/release gates and end-to-end question → note → proposed change → admitted execution → persisted result → regeneration → verified presentation. Separate live evidence from mocks.
  • Final Domains source-of-truth and generated documentation checks.
  • Every-slide visual ledger and PDF inspection complete; all invalidated source/teaching reviews resolved.
  • Section 10's content-led composition overhaul fully adopted: current audit, slide briefs, meaningful guide openings, complete reusable repertoire and semantic primitives, both example redesigns, validated authoring/renderer contracts and whole-library migration. All guides pass the composition/diagram rubric; new authored and Eve-generated or revised decks follow the same contract. Current regression, accessibility, responsive, performance, narration and PDF evidence covers the delivered artifacts. Neither the pilot nor a layout-diversity statistic closes this gate.
  • Screenshot/mockup coverage and provenance complete for appropriate UI sections; asset loading, accessibility, responsive/print legibility and any visual interactions verified through the applicable automation and visual review.
  • All eight section 10 craft/comprehension standards implemented and adopted: versioned benchmark, shared typography/font/spacing contracts, validated quantitative visuals, consistent recurring identities, reviewed image treatments, all declared viewing sizes, real scoped reader findings and subtraction. Every guide has current rubric evidence or an explicit justified non-applicable entry; all material findings are resolved.
  • New font/data/image/layout behavior passes applicable unit/contract, builder, accessibility, actual-size Playwright, performance, narration and PDF checks. Eve generation/editing evaluations use the same benchmark and craft rules without overstating human evidence. Reference examples, reader findings and regression artifacts are versioned and fresh; automated passing results do not replace optical or comprehension review.
  • UI/UX design coverage matrix complete, with substantive atomic-design, foundations, interaction, accessibility and particular design-choice guides; source-backed explanations, alternatives and implementation status reviewed.
  • Every-guide editorial and quality ledger complete against current slide, visual and narration hashes; full normal-speed listening and story/organization reviews passed, with no unresolved failing quality categories.
  • Update README, guide map, source/coverage maps, verification records and this checklist to reflect the actual final catalog. Keep historical reports labeled as historical, and record exact hashes and verification scope.
  • Stage only relevant reviewed changes, run normal hooks, commit clearly, integrate current origin/main, push and verify no local-only task commit. Do not stop with a rejected push or diverged task branch.

Reproduction commands#

Use the prepared Python environment and resource checks before each expensive command. Run the heavy commands separately; do not paste them into one concurrent build/test/narration batch.

sh
python tools/presentations/build-eve-oshun.py --check --preview
python tools/presentations/render-eve-oshun-narration.py --check
python -m unittest discover -s tools/presentations/tests -p 'test_*.py'
node --test tools/presentations/tests/*.test.mjs
ruff check tools/presentations
node_modules/.bin/eslint --config tools/presentations/eslint.config.mjs docs/presentations/presentation-center/*.js tools/presentations/tests/*.mjs
node_modules/.bin/prettier --check docs/presentations/presentation-center tools/presentations
python tools/presentations/check-center-coverage.py --check
python tools/presentations/check-center-coverage.py --require-complete
python tools/presentations/inventory-repository.py --check
python tools/presentations/inventory-workflows.py --check
python tools/presentations/check-workflow-coverage.py --check
python tools/presentations/check-workflow-coverage.py --check --require-complete
python tools/presentations/run-supervised.py node tools/presentations/tests/eve-oshun-deck.mjs
python tools/presentations/run-supervised.py node tools/presentations/tests/presentation-center.mjs
node tools/docs/verify-domains-source-of-truth.mjs

For an intentional inventory/source revision review:

sh
python tools/presentations/run-supervised.py python tools/presentations/inventory-center.py --check

Focused browser commands are useful while editing:

sh
python tools/presentations/run-supervised.py node tools/presentations/tests/eve-oshun-deck.mjs --guide portfolio-map
python tools/presentations/run-supervised.py node tools/presentations/tests/eve-oshun-deck.mjs --guide v2-product-promise --draft
python tools/presentations/run-supervised.py node tools/presentations/tests/presentation-center.mjs --smoke --draft

Use OSHUN_DECK_QA_DIR or OSHUN_CENTER_QA_DIR to retain a distinct report folder. --draft, --smoke, --controls-only, --audio-only, --layout-only and --panels-only prove only their explicit subsets. --guides in the center requires a passing full baseline and unchanged HTML for every unselected guide. The global speech change invalidates that shortcut against the historical 35-guide baseline; run the full current suite before final completion.

During the narration rebuild, presentation-center.mjs --draft --ready-audio keeps all-guide layout/PDF checks and adds actual embedded start/end/replay/stop checks for every fully narrated guide. The report lists the selected and skipped audio guides. It rejects smoke/focused combinations and an empty eligible set; it remains a draft report and cannot serve as the full narrated release baseline. Ordinary --draft still skips audio, and non-draft mode does not silently omit incomplete guides.

The internal .qa-preview and large temporary screenshots are reproducible and not required transfer files. Source excerpts come from Git's pinned objects, not temporary old-machine copies. Saved machine-readable reports retain their original artifact hashes; later rebuilds require checking scope again.

Complete center runs now retain an initial-viewport screenshot of every embedded guide-slide instance on desktop and mobile, with ordered IDs and verified image hashes in embeddedSlideCaptures. These are review inputs, not an automatic visual pass. Standalone full-height mobile/table-end captures and PDF renders remain necessary for content outside an embedded viewport.

Decision (user, 2026-09-11): the Presentation Center and the Docs Center stay generated static sites. They are not rewritten as React applications and they do not gain a CMS. The repository is the content source, git is the authoring workflow, and the deterministic builders with their freshness gates are the properties this section must preserve. Everything dynamic is added as a thin runtime host plus small embeddable islands on top of the static output.

Current state (measured 2026-09-11): the four surfaces exist but do not link to each other. The Presentation Center cites sources by repository path and by docs-center/... relative path, and inlines excerpts into the Explain dialog at a pinned sourceRevision; those are copies, not links to a served page. The Docs Center has no link to any deck and no navigational link into a running app; its only localhost mentions are inside code samples. Neither apps/oshun/web nor apps/oshun/admin links to either center; the web /docs routes are product documentation. Eve's docs search (apps/oshun/bff/src/assistant/docs-search.ts, corpus built by tools/build-assistant-docs-search.mjs) returns repo-relative reader links such as docs-center/tara.html#anchor and no base-URL setting exists, so the admin chat cannot open them; the presentation catalog is not in that corpus. Neither static center embeds Eve (center.js, deck.js and the Docs Center reader.js contain navigation and narration only). There is no single development launcher: the root pnpm dev runs every Nx dev target in parallel, which the machine rules in CLAUDE.md forbid, and scripts/dev/ holds only the V1 domain-service runner. Section 8 owns Eve's grounded question, note and enacted-change contracts for the center; this section owns the host, the address book, the links, the liveness/freshness signal, the demo state and the Eve dock island that carries section 8's contracts into the static pages. Do not duplicate section 8's items here; reference them.

Host decision changed on 18 September 2026. Measured that day: the BFF authenticates by Bearer token only, parses no cookie and serves no static files, so a page served from it has no operator session to reuse; the admin console has the cookie, the assistant proxy routes and the gate. The admin app therefore serves the centers, and Eve's dock is the chat already mounted in AdminShell. Where an item below says "the host", read the admin origin. Sections 12.1, 12.3 and 12.8 moved to Phase 0 of EVE-V1-REDESIGN-TODOS.md; the rest of this section stays here and follows that phase.

12.1 Architecture decision record#

Moved: EI.0.02 of Phase 0 writes authoring/four-surfaces-architecture.md, with the host decision above and the conditional later additions this subsection listed.

12.2 One address book#

  • Define the canonical addresses of the five surfaces. Three are origins (web 3010, admin 3020, BFF 4010); the two centers are paths on the admin origin, because the host decision of 18 September 2026 serves them from the admin app (/presentations/files/… from EI.1.03, and the Docs Center equivalent from 12.3), so no new port is reserved (this item first asked for two). Check the per-worktree offset convention (lsof before binding) and document the addresses in CLAUDE.md's ports line and in the two center READMEs. Verify: the unit tests of this subsection's last item, and the three documents name the same five addresses.
  • Introduce environment names for the three origins (web, admin, BFF) and the two center directories (OSHUN_PRESENTATION_CENTER_DIR from EI.1.01 and its Docs Center counterpart) in one place. Names go in the drawer env file convention (~/Desktop/workspace/drawer/env-master.env on the Mac, ~/.config/oshun/env-master.env on the dev server); values never enter the repository. The BFF has no central environment schema to add them to (checked 2026-09-18: OSHUN_WEB_BASE_URL, the one origin it knows, is read ad hoc in apps/oshun/bff/src/auth/verification-email-sender.ts and routes/calendar-connect.ts), so the address-book library of the next item is the single reader: it returns a typed not_configured naming the variable instead of a silent default, and those two call sites move to it. Verify: the library spec covers a missing origin, and neither call site reads process.env.OSHUN_WEB_BASE_URL directly any more.
  • Add one shared library, libs/oshun/surfaces (@oshun/surfaces, path-mapped in tsconfig.base.json, with project.json, package.json, tsconfig.json, real tests and a clean build), that exposes the address book as typed data to Node and to browser bundles. The browser never gets values baked at build time: a page served by the admin app fetches address-book.json, a virtual file the EI.1.03 route answers from the environment, and a page opened from disk gets nothing and shows no cross-surface links. Verify: npx vitest run in the library passes, npx tsc --noEmit is clean, and the route spec of EI.1.03 gains a case for address-book.json.
  • Unit tests: every surface has an origin; an origin missing in the environment is reported as not_configured with the variable name; a per-worktree offset produces distinct origins for two worktrees.

12.3 The host process#

Moved and changed: EI.1.01–EI.1.06 of Phase 0 serve the Presentation Center from the admin app (read-only, allowlisted, traversal-safe, fail-loud when unconfigured, with Range support and scoped frame headers). Serving the Docs Center the same way is the one item left here:

  • Serve the generated docs-center/ output from the admin app with the same module and handler shape as EI.1.01–EI.1.05 (a second root, a second allowlist taken from the Docs Center generator's outputs, the same header scoping), after Phase 0's 0.2 has landed. Verify with the same five route specs over a Docs Center fixture.

12.4 A dev up that respects the machine#

  • Add one launcher, scripts/dev/up.mjs (beside the existing run-v1-domain-services.mjs), that starts the BFF plus one Next app chosen by flag (--web or --admin), never both, per CLAUDE.md. Both centers are served by the admin app, so they are up only with --admin, and the launcher says so when started with --web. It sources the throwaway env script, records every PID it starts in /tmp/claude-<session>.pids, and refuses to adopt a listener it did not start on the three ports. Verify: the dry-run and refusal specs of this subsection's last item.
  • Add the matching down that kills only the tracked PIDs and then verifies with lsof -nP -iTCP -sTCP:LISTEN over all five ports that nothing of its own remains, printing the survivors if any.
  • Print the address book on start (three origins and the two center paths, which app is up, which is deliberately down on this machine, and whether the centers are reachable) so the hub state is visible without opening a browser. Verify: the dry-run spec asserts the printed table for --web and for --admin.
  • Do not make the root pnpm dev target call this; leave the Nx target as is and document that the launcher is the supported way to bring the four surfaces up on the Mac. On the Linux dev server both Next apps may run; the launcher accepts --web --admin only when uname -a reports dev-server and free -m shows more than 3 GB available.
  • Tests: a dry-run mode that prints the planned processes and ports without starting anything; a spec that the two-app combination is refused on darwin.

12.5 Stable ids and machine-readable indexes from the generators#

  • Presentation Center: emit link-index.json from the builder with, per guide and per slide, the stable id, the deck URL fragment, the cited source paths with their pinned sourceRevision, the Docs Center hrefs already present in citations, and any product routes the slide's captures or narration name. Reuse catalog.json (presentations, hierarchy, tracks, readingRoutes) and coverage-map.json (slidePresentations) as inputs; do not create a parallel catalog.
  • Docs Center: emit the equivalent index from tools/docs_center/ with, per page and per anchor, the stable id, the source markdown or code path, the product tint, the journeys that verify it and the app routes those journeys walk (the reader already computes verified-by-journeys rollups on app entity nodes; reuse that computation).
  • Both indexes are deterministic outputs covered by the existing freshness --check gates and by unit tests that assert every slide and every page appears exactly once.
  • Define one id grammar for cross-surface references (for example deck:<guide>/<slide>, doc:<space>/<page>#<anchor>, src:<path>@<rev>, app:<surface>/<route>) in the shared library from 12.2 and validate both indexes against it.
  • Implement resolve(id) → URL and reverse(URL) → id in the shared library over the address book and the two indexes. src: ids resolve to the GitHub permalink at the pinned revision and to the local worktree path; app: ids resolve to the running origin and report when that surface is down on this machine.
  • Ship the resolver as one self-contained browser script (esbuild, no framework) that the builders reference from every generated page and that does nothing until address-book.json answers, which it does only when the admin app serves the page. It upgrades existing citation markup in place: Presentation Center source excerpts gain "open in Docs Center", "source at pinned revision" and "see it live" links; Docs Center pages gain "see the deck" where a slide cites them and "open in app" where a journey walks a route. The upgraded controls render inside the Eve edition's chapter chrome, never as a new footer element. Verify: the Playwright spec of the next item.
  • Unit tests for the resolver: every id kind round-trips; an unknown id fails loud; a down surface yields a typed "unavailable" result instead of a broken link. A Playwright spec on chromium (workers=1, the repo's inspect harness) shows the upgraded links present on a served deck page and absent on the same page read from disk.
  • Docs Center: render "see the deck" backlinks from the Presentation Center index at generation time as plain relative markup (so they survive offline reading) and let the resolver island only upgrade them to absolute origins.
  • Docs Center: render "open in app" on pages whose journeys walk routes, naming the surface (web/admin) and the route, with the release-scope state from libs/oshun/navigation/src/release-scope.ts shown when the route is outside V1.0.
  • Presentation Center: render "read the docs" on every slide whose citations include a Docs Center href, next to the existing source excerpt control, and keep the excerpt so the deck stays useful offline. Render it inside the Eve edition's chapter chrome, validated by validate_eve_edition, never as a legacy footer element.
  • Apps: add "docs for this screen" to the studio and admin shells, driven by a route-to-doc map exported from the Docs Center index into a small JSON the apps import; no hand-maintained list. The member web shell shows it only for builder-plane routes (Eve is builder-side; Lilith members do not see engineering docs).
  • Unit tests: the route-to-doc map covers every studio route that has a journey; a route with no doc renders nothing rather than a dead link. Regenerate both centers and confirm their --check gates pass.

12.8 Eve dock island for the static centers#

Moved and changed: there is no bundled island. EI.2.01–EI.2.05 give the static pages a message contract, EI.3.01–EI.3.07 put the center inside the admin console beside the existing Eve chat, EI.4.01–EI.4.04 carry the slide in the page context, and EI.5.04–EI.5.05 add the slides to Eve's corpus with openable links. The Docs Center half is the one item left here:

  • Give the Docs Center reader the same message contract (page id, section anchor, source path) and an admin page beside Eve, reusing @oshun/shell-assistant/presentation-messages with a doc message type, after Phase 0's 0.4 has landed. Verify with the component and chromium checks of EI.3.02–EI.3.07 over a Docs Center page.

12.9 Context handoff across surfaces#

  • Define a from=<id> query contract (using the 12.5 id grammar) that every cross-surface link carries, and have each surface's assistant entry read it into the page context so Eve knows where the user came from: apps/oshun/web/src/navigation/assistant-entry.ts and assistant-continuity-journey.ts on web, and on admin the presentations page of EI.3 (there is no dock inside a deck: Eve beside a deck is the admin chat, and the deck tells it the slide through the message contract of EI.2). Verify: the unit test of this subsection's last item.
  • Continue the assistant session across surfaces through the BFF session store rather than per-origin browser state: a thread started beside a deck in admin is the same thread when the same authenticated operator opens the studio assistant on web. (Beside a deck and elsewhere in admin it is already one chat, by the Phase 0 design.) Verify: the journey spec of the next item.
  • Tests: a journey spec that starts in a deck, follows "show me this in the app", and asserts the admin dock shows the deck as the origin of the thread; a unit test that a malformed from= value is ignored, never echoed into the page.

12.10 Liveness and freshness strip#

  • Add an admin API route (operator-gated like the rest of /api) that reports, for each address in the address book, whether it answers on this machine, and the git revision of the working tree and of HEAD. Verify: the route spec of this subsection's last item.
  • Ship a liveness island that renders a compact strip in both centers: which apps are up, which are deliberately down on this machine, and for the current page or slide, "describes revision X; working tree is at Y" using the pinned sourceRevision from the indexes. A stale slide is flagged with the count of cited paths that changed since the pinned revision (computed server-side from git diff --name-only).
  • Reuse the Docs Center's corpus-anchored staleness rule (anchor = newest walk, never the build clock) so the strip never breaks the byte-diff gate.
  • Tests: the route reports a down surface as down (not an error); a slide whose cited file changed is flagged with the correct count; the strip renders nothing under file://.

12.11 Demo state for "see it live"#

  • Inventory the fixture harness the deck captures already use (tools/presentations/tests/eve-product-capture.mjs and the pinned PNGs under assets/eve-product/) and record which app routes and seeded tenant/user states each capture depends on.
  • Expose that seeding as a named demo mode of the launcher (dev up --demo <fixture>), so a "see it live" link from a slide lands on the same route in the same state as the screenshot. Seeding runs against the local Postgres only and refuses any non-local database URL.
  • Add the fixture name to the app: id of each slide that has a capture, so the resolver can request the matching demo state or, when it is not loaded, tell the user which fixture to start.
  • Tests: each capture's fixture seeds without error on a fresh local database; a "see it live" link from a captured slide reaches a page whose key elements match the capture's recorded selectors on chromium.

12.12 Verification, documentation and closure for this section#

  • Both generators regenerate and pass --check; both centers' Playwright reader suites stay green on chromium; the docs-search specs, the address book and resolver unit tests, the host integration tests and the launcher specs are green. Record the counts in a short evidence block, not a ledger.
  • Run the adversarial stub grep from .claude/rules/task-checkbox-verification.md over the new library, the host code, the islands and the launcher; zero actionable hits before any item above is marked. Rule, not a task (un-boxed 2026-09-18): the scan is part of verifying every item above, per the rule file; it is not a separate deliverable.
  • Update CLAUDE.md (ports line, launcher name, one-Next-server wording), both center READMEs, docs/agents/ with the architecture record's pointer, and section 3 of this file so the resume recipe starts the host instead of a file server.
  • Memory check on the Mac before every heavy step in this section (memory_pressure | tail -2, below 35% free → tear down); the host, the Next app and a chromium run are the whole budget. Dev servers die with the session; verify the five ports are empty before ending. Rule, not a task (un-boxed 2026-09-18). With the admin app as host there are three ports to find empty, not five.

13. Where to resume, and when it is actually finished#

On continuation, establish section 0's current baseline and follow this order. Skip implemented steps only when their required evidence remains valid; complete open review and acceptance even when an artifact already exists.

  1. Validate the environment and latest checkpoint against current inputs. Run builder/narration freshness checks and resume only actually missing or stale recordings. The original 316-track queue has a recorded completion in section 4; do not repeat it based on the handoff count. Finish invalidated narration and current-library verification, retaining open listening review.
  2. Reconcile saved drafts with integrated guides and finish the next remaining coherent source-backed work package. V2 contracts already has an integration receipt in section 5; complete its outstanding reviews rather than duplicate its delivery. Extend V1–V10/Docs Center mapping to the entire monorepo, including source code, generation/ComfyUI/agentic workflows, architecture, implementation, repository layout, infrastructure, CI/CD and operations. Continue enriched content until no meaningful source information remains unassigned or unreviewed. Include the extensive UI/UX design guides, particular design-choice walkthroughs and relevant screenshot/mockup slides as part of that content expansion. Author every one of those slides in the Eve edition under section 0's design edition binding and EVE-V1-REDESIGN-TODOS.md; the accepted pilot is chapter 1, Meet Eve, so do not establish another. Migrate existing guides through the redesign file's Phase D tracks and set up any new track with its D0 recipe first. Section 10's craft/comprehension benchmark, typography specification and representative-reader review still apply, measured on edition chapters.
  3. Complete all eleven cross-cutting tracks in section 6, including their worked examples, extension recipes, role-specific exercises and maintainer guides. Finish complete hierarchy, guide count, cross-links and measured durations; verify learning routes and track-level coverage as well as source-unit coverage.
  4. (Moved to the front on 18 September 2026.) Real Eve questions, notes and enacted changes, and conversation-driven presentations, are Phase 0 of EVE-V1-REDESIGN-TODOS.md and are worked before steps 1–3 here.
  5. Complete the exhaustive Domains navigation/source/code/planned-root audit.
  6. Review every slide and add useful diagrams, screenshots, UI mockups and additional visual slides; finish section 10's composition overhaul across every guide and slide instance, including opening and thumbnail-sequence reviews, source-faithful flowcharts, mobile/accessibility and PDF checks. Complete visual and UI/UX coverage and verification.
  7. Review every presentation end to end for excellent storytelling, natural human-like narration, pacing and consistent organization. Listen to every recording and complete all eight craft/comprehension rubric entries, including optical refinement, chart/image fidelity, continuity, actual-size readability, required reader retests and editorial subtraction. Fix all quality defects, regenerate affected artifacts and repeat invalidated reviews and checks.
  8. Tie the four surfaces together per section 12: record the architecture decision, land the address book, the BFF host for both centers, the machine-respecting launcher, the generator indexes, the link resolver and generated backlinks, the Eve dock island (carrying section 8's contracts), cross-surface context handoff, the liveness/freshness strip and demo state. The centers stay generated static sites throughout; do not migrate them to a React application or a CMS.
  9. Pass every applicable final gate, update evidence, commit and push to main.

The original transfer request stopped work on the old machine after publishing that checkpoint. It does not instruct a continuing agent to stop after each increment or restart the old renderer. Resume authorized work from verified current state under section 0. The broader presentation objective remains unfinished until every requirement above is implemented and verified; do not mark it complete merely because this transfer checkpoint or one of the intermediate phases is complete.