# Eve product and workbench teaching

Work package: `eve-capabilities-2026-09-07`. This implements the Eve breadth and
sequencing portion of `TODOS.md` §§6, 10 and 11. It does not close the separate
whole-repository expansion, live presentation editing, or human reader
acceptance requirements.

The user's attached brief identifies missing product graphs, coverage analysis,
task tracking, discussions and decisions, and change/delivery tracking. The
source audit adds feature walkthroughs, operator investigations, studio reads
and confirmed studio actions, content delivery, and shared member tools.

## Baseline and source boundary

Implementation baseline: `8c27a8d3350fe7484c18585b570f72a561090540`. The
immediately preceding implementation has 47 guides, 470 unique slides and 493
guide-slide instances. The thirteen-guide Eve architecture route contains 110
slides. Its recorded verification is in `verification/eve-sequence-2026-09-07/`;
those reports retain their original hashes and scope. They are not fresh
evidence for this expansion.

New capability teaching pins the full baseline revision above. Earlier
architecture and implementation chapters keep their own declared revisions. The
pin establishes what the source says, not which features a particular deployment
has configured or which live provider succeeded today.

## Design brief

Visual thesis: keep the cream, ink and terracotta teaching workspace; use
readable relationships, exact status distinctions and a continuous worked case
to make Eve's product capabilities concrete.

Content plan: begin with the work a person wants to accomplish; establish the
objects and evidence; teach each operation and recovery path; end with a
traceable delivery review. Each guide has one learning outcome and each slide
one dominant relationship. Detailed explanations belong in narration and notes.

Interaction thesis: retain the existing restrained slide entrance, source
disclosure and route/sidebar transitions with reduced-motion support. Diagrams
must remain complete without animation and in PDF. Navigation should expose
product/workbench learning separately from architecture while preserving old
guide URLs and meaningful prerequisites.

Use evidence comparisons for independent facts, tables for exact mappings, flows
only for actual order, and graph relationships with labeled directions.
Illustrative records and numbers must be identified as examples. Do not draw
invented production dashboards or imply drag-and-drop actions that do not exist.
Apply the existing typography floors and inspect actual embedded sizes and PDF
page fit. Model review and automation do not establish human comprehension.

## Requirements and guide homes

| Requirement                | Teaching outcome                                                                                    | Guide                       |
| -------------------------- | --------------------------------------------------------------------------------------------------- | --------------------------- |
| Capability orientation     | Choose the right surface, identity and evidence source for a real request                           | `eve-capabilities`          |
| Product graph              | Read containment, typed relationships, provenance, inspector and bounded neighborhoods              | `eve-product-graph`         |
| Analysis                   | Interpret verification, waivers, mapping gaps, route impact and repository/live backlog joins       | `eve-coverage-analysis`     |
| Task tracking              | Capture, link, prioritize, triage, inspect and recover work through its actual lifecycle            | `eve-work-tracking`         |
| Discussions and decisions  | Anchor discussion, preserve alternatives, accept explicitly, supersede and export ADRs              | `eve-discussions-decisions` |
| Delivery                   | Follow leases, reports, artifact expectations, failures, manifest changes and release notes         | `eve-delivery-tracking`     |
| Member feature audit       | Begin/resume, walk ordered steps, mark with evidence, handle drift and close a run                  | `eve-feature-audits`        |
| Operator investigations    | Read queues, incidents, crashes, models, health, readiness and freshness with scope and time bounds | `eve-operator-analysis`     |
| Studio analysis            | Use every registered view family; distinguish stored observations from pure evaluation              | `eve-studio-analysis`       |
| Studio actions and content | Capture/promote/transition/schedule with confirmation; deliver briefs and configured ideation       | `eve-studio-actions`        |
| Shared member tools        | Locate content, resume activity, manage supported personal records and guided tours                 | `eve-member-capabilities`   |
| Integrated practice        | Find gap → evidence → discussion → tracked work → decision → delivery, with counterexamples         | `eve-workflow-practice`     |

The guide count follows these different learning outcomes. The capability
inventory must enumerate actual registry entries and map every entry to teaching
or an explicit implementation boundary. A citation alone is not reviewed
teaching. Existing material should be reused by reference where already deep; do
not duplicate whole architecture slides to inflate coverage.

## Source-review findings to preserve

- Graph Explorer runs on member web at `/assistant/graph`, behind builder
  authorization. `open_graph_explorer` returns a link; it does not operate a
  browser. Workbench writes additionally require a configured store and a
  confirmation-capable client; studio views enforce their own studio scope.
- The full compiled graph vocabulary is broader than the assistant's runtime
  graph. The latter composes curated catalog, journeys, documentation and
  routes. Repository TODO/architecture and projected live-intent queries have
  distinct artifact/projection paths.
- `get_graph_neighborhood` currently receives that knowledge-only runtime index.
  Its generic traversal supports incoming `tracks`, but its caller does not
  merge live work. Teach an explicit `list_work_items` read for live work; the
  tool description's broader wording is not evidence of a live overlay.
- Explorer inspector work references are exact-node matches; the work-list tool
  supports containment closure. The inspector's “open work” label does not
  itself filter terminal records from the board snapshot.
- The Work lens is status columns with the first graph reference printed on each
  item. There is no implemented drag mutation or separate domain swimlane in the
  rendered body despite the file header describing swimlanes.
- A mapped verifying journey is graph traceability, not a fresh passing run. A
  verified domain does not establish that all its flows are verified.
- Changes compares current manifest sections with a pasted prior manifest. It
  does not list removed-only sections or node-level changes; use Git for the
  complete artifact difference.
- Member audit progress is visited/skipped/pending for a pinned walkthrough. It
  is independent of automated coverage and work-item verification.
- “Complete course” overstates the existing architecture route. Replace that
  claim with accurate route names and specific learning outcomes.

## Verification and completion

Author and source-review first. Then run applicable builder/schema/coverage,
narration, route, changed-guide browser/accessibility and PDF checks serially
under host resource limits. Reuse exact unchanged artifact evidence with
explicit hash continuity; do not repeatedly rerun the full bank. Extend focused
checks for the new inventory, guide sequence and meaningful learning-path
behavior. Record optical/editorial findings separately from automated checks.
Update catalog, README and continuation records to the actual result, commit
with normal hooks, integrate current remote changes and verify the task tip on
`origin/main`.

Status: all twelve guides have current narration and PDF editions, and the
scoped automated checks pass. See the
[delivery receipt](../verification/eve-capabilities-2026-09-07/README.md) for
exact scope, retained visual reviews and artifact hashes. Human reader and
normal-speed listening acceptance remain separate evidence categories.

## Editorial review and sequence

Reviewer: Codex, 7 September 2026; source/code and authored explanation review.
This is model review, not a human comprehension trial.

The first six guides teach the dependent product-to-delivery chain. Graph
objects precede coverage verdicts; evidence precedes task capture; task
lifecycle precedes discussion/decision records; those records precede execution
and artifact checks. The worked enrollment example names its illustrative status
throughout. Recovery slides cover waivers, changed graph links, parked/rejected
work, superseded choices, lost leases, failed verification and incomplete
delivery evidence.

Operator and studio chapters follow that shared foundation. Studio reads precede
confirmed actions; content briefs follow the task and delivery concepts they
reuse. Member audits then establish the separate visited/skipped model before
shared member tools. The final practice chapter revisits the complete chain and
asks the learner to reject plausible but unsupported conclusions. The full path
continues into existing architecture and controls after this product practice.
The separate architecture route preserves introduction/detail pairing for a
reader taking that route independently.

All 94 slides have a defined job, dominant relationship, detailed explanation,
three distinct questions and pinned sources. The authored layouts use 37
evidence comparisons, 36 tables, 19 ordered flows, one layer map and one
ownership/handoff diagram. Tables express exact mappings; arrows express a real
sequence or named handoff, not decorative progress. Long identifiers remain in
explanations where necessary, while narrow table labels use readable operation
names. No illustrative record or diagram is represented as a screenshot of a
live production session.

The [registry inventory](eve-capability-inventory.json) is extracted from pinned
TypeScript syntax and MCP registrations. The
[teaching crosswalk](eve-capability-teaching.json) contains 54 explicit
editorial assignments covering its 128 entries. The crosswalk validator checks
exact assignment sets, actual guide membership, explanation and question
presence, and pinned registry citations. It deliberately does not turn keyword
occurrence into an automatic source-review or comprehension verdict. The
section-level assignments for the product graph's complete authority-rule and
“Two planes” units declare this package's exact source revision. The coverage
validator admits that newer slide evidence only after reading the supplemental
Git blob and matching the entire cited Markdown file to the primary inventory's
pinned file hash; changed or unattested newer implementation evidence still
cannot satisfy the older inventory. The README's package layout, build command,
Postgres/Neo4j mechanics and cross-plane query-pack units remain unassigned
until their complete details are taught.

Additional source boundaries preserved in teaching:

- Task priority is supported; ordinary tool inputs do not expose assignee or due
  date. A graph reference is traceability, not responsibility assignment.
- Accepted decisions and ADR export are distinct. Export produces files and a
  ledger result; it does not itself commit or push Git history. Accepted
  decision authority is taught through the admitted confirmation path, without
  inventing a human-actor guard inside the store transition function.
- Repository-configured shipping checks ancestry against the local `origin/main`
  ref. Without that configuration, the store can record a caller observation
  without checking repository ancestry. The check does not itself fetch Git.
- Artifact verification checks the current compiled artifact's conditions.
  Presence is not a before/after causal diff; an E2E suite file match is not an
  executed passing test. Unsupported or absent expectations remain explicit.
- Fleet and lease tools are observational. Missing cost is not zero, expired
  holds are not reclaimed by reading them, and triage recommendations do not
  themselves execute recovery.
- Member feature audits retain a pinned walkthrough and historical marks.
  Visited/skipped progress, graph coverage and work verification are three
  distinct models; closing an audit can retain pending steps.
- Studio views are classified reads with scope/configuration admission and
  bounded lists. Stored readiness records and a pure readiness evaluation answer
  different questions. A page feature is not automatically a drawer tool.
- Studio actions require confirmation and current state. Scheduling plans a
  date; it does not prove publication. A content-brief dispatch records its
  target reference and work-state changes; it does not create an external target
  job.
- Shared member tools are taught by their actual records. A Nisaba workspace is
  not evidence of a general notebook editor; absent reading progress stays
  absent. A curated tour submission is not approval or an automatic future
  notification.

## Review findings and delivery evidence

Initial narrow-screen checks exposed broken table words. Operation labels were
shortened without reducing typography or removing exact selectors from the
explanations. The complete 282-slide/view pass then found one desktop evidence
label collision in `eve-cap-operator-docs-memory`; its label was shortened to
“Docs state.” The first standalone visual review also exposed 37 captions that
repeated their takeaways; those now identify the comparison’s specific context.
This caption-only change preserves the narration text and recordings. Superseded
checks and failed attempts are retained with the delivery receipt; final passing
checks bind the corrected manuscripts and generated output.

Recovery uses the renderer's bounded batches and current manifest fingerprints:
resume only pending audio, stop the owned process group on memory pressure, and
rebuild duration-derived files after successful generation. Browser and PDF
checks run sequentially; exact unchanged prior guide and audio hashes retain
their earlier evidence instead of rerunning unrelated guides.

## Craft rubric and reader tasks

The following common criteria apply separately to each of the twelve guide rows
below. They are editorial judgments with explicit evidence limits, not averaged
scores or a declaration that the separate whole-library craft program is closed.

| Standard                    | Application and evidence boundary                                                                                                                                                                                                          |
| --------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
| 1. Visual benchmark         | Inherit the established cream/ink/terracotta teaching language and native relationship treatments. The separate library-wide external benchmark acceptance remains open.                                                                   |
| 2. Typography and alignment | Keep the shared typography floors; shorten narrow labels rather than shrinking them. Source/semantic checks inspect every new slide at 1280, 390 and 320 pixels, with embedded and print checks covering the delivered editions.           |
| 3. Quantitative comparisons | These guides introduce no measured performance chart. Windows, list caps and state counts describe source contracts; missing observations never become zero or live success.                                                               |
| 4. Continuity               | Keep exact state names and the illustrative enrollment case across the dependent graph/work/delivery chain. Distinguish member audit, graph coverage and work verification throughout.                                                     |
| 5. Image direction          | Native tables, flows and named handoffs explain these source-defined records. No generated photograph or invented product screenshot is needed for their teaching job.                                                                     |
| 6. Actual-size readability  | Automated geometry/text checks are supplemented by inspection of real browser and PDF renders. Retain the exact reviewed specimens in the delivery receipt; do not describe viewport emulation as a physical-device or browser-zoom trial. |
| 7. Reader comprehension     | Each slide has three review questions; the guide-specific tasks below define expected evidence. No willing-reader session or normal-speed listening trial has been conducted for this expansion.                                           |
| 8. Editorial subtraction    | Keep one dominant relationship per slide. Move exact selectors and qualifications into the explanation when a visual label becomes crowded; preserve the operation, condition and failure meaning.                                         |

| Guide                     | Criteria 1–8 applied in its own context; decisive reader task                                                                                                                                                                         |
| ------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| What you can do with Eve  | Establish purpose, surface and availability before mechanisms. The reader should select the correct surface and distinguish the three meanings of progress.                                                                           |
| Product graph             | Move from containment through typed links to bounded reads and inspector scope. The reader should locate a feature, follow a named relationship and explain why a separate live-work read is needed.                                  |
| Coverage analysis         | Define verdict and freshness before route impact/backlog joins. The reader should identify the exact missing evidence and avoid converting graph mapping into a fresh test result.                                                    |
| Work tracking             | Introduce the record, capture kinds, human preparation, delivery states and explicit recovery. The reader should update the intended item and graph links without inventing a due-date or assignee input.                             |
| Discussions and decisions | Keep conversational evidence separate from accepted choice and file export. The reader should preserve alternatives, accept through the proper path and supersede without rewriting history.                                          |
| Delivery tracking         | Keep observation, execution, shipping and artifact conditions separate. The reader should trace a lease/report and explain exactly what a verification result does and does not establish.                                            |
| Operator analysis         | Start with scope and observation windows before queue/detail investigations. The reader should select a bounded read, retain missing-data qualifiers and avoid treating a read as a control action.                                   |
| Studio analysis           | Group every admitted view into its meaningful workbench family; reads precede evaluation and writes. The reader should inspect a concept or risk record and distinguish stored data from pure evaluation.                             |
| Studio actions            | Begin with confirmation/current state; progress through capture, promotion, transition, scheduling and brief delivery. The reader should recover the exact target identity and distinguish recorded dispatch from external execution. |
| Feature audits            | Explain begin/resume and one observed step before skip, progress, drift and closure. The reader should preserve pinned-run history and explain why a closed run may still contain pending steps.                                      |
| Shared member tools       | Organize by actual member records and supported operations, then page context and tours. The reader should choose a real tool, retain absent-progress boundaries and distinguish submission from approval.                            |
| Workflow practice         | Recombine the dependency chain, show named responsibility handoffs and test counterexamples. The reader should reconstruct a delivery narrative whose every completion claim has the required evidence.                               |

Reader sessions can use the existing [reader-review protocol](reader-review.md)
with the exact final guide, source and artifact hashes. The present automated
checks and model review do not supply participants, observations or quotes.
