Eve library

Lilith content workbenches — state of the art for V1.0 to V1.4: scope (2026-09-29)

PRD

Revision 18

On this page

Status: draft for Elikem's approval; the §9 decisions were settled by Elikem on 2026-09-29. Author: Claude (claude-mac-worker), 2026-09-29, from Elikem's brief of 2026-09-28. Board family: content-workbenches, filed 2026-09-29 (315 tasks and 8 epics, 1,150 dependency edges). Every task is later until this scope is approved, except the five urgent fixes to the live product (TA01, TA02, AR02 and NI02–NI04), which Elikem asked to be filed as ready. The family is open only because the board accepts ready work solely in open families. Seven research documents in Eve carry the evidence, one per area. Each has a numbered source list and a feature catalogue (§11).

An independent read-only review (Codex, 2026-09-29) was folded in before filing. It checked the claims about code against the repository and raised 29 findings, all addressed:

  • Each room now has a fail-closed release gate that its launch task depends on, and 35 cross-family dependencies on the Metis and Veritas ledger phases are wired.
  • Foundation tasks are now tested with fixtures instead of the room work built on them, which removed hidden dependency cycles.
  • New tasks cover sensitive-data governance, security for untrusted content and code execution, audience and age enforcement, harm intake before launch, an operational-evidence programme kept separate from software completion, server-enforced practice eligibility, territory-aware serving of existing passages, and an ownership map for the V1.4 overlaps.
  • Acceptance tests that contradicted the recommended decisions were corrected: no clinical exception to the claims rules, review never replaced by a label, due weight without numeric quotas, no betting markets, the sacred-text policy matrix, publisher-specific export profiles.
  • The descriptions of Nisaba's notebooks, Veritas's agent plans, the Aila evidence and ESA licences were corrected.

Revision 3 (2026-09-29) answers the unattended review of revision 2 and adds a web-search verification pass:

  • Consented outcome and harm measurement moved from M3 to M1, so it is part of every launch gate.
  • The UK KJV rights fix is listed as its own urgent item, making five in all.
  • The exact planned task bodies, dependency graph and a feature-to-task acceptance matrix are now in a linked evidence document. The seven research documents are linked at pinned revisions.
  • The Nyx, Arete, Metis and creative research, first done without web search, was redone with live web search (§3). It added 20 features and revised 85, plus a task for the 6 February 2027 annular eclipse over Accra, and it corrected the claims noted in §3.

0. The brief (Elikem, 2026-09-28, verbatim)#

so this time focusing on the workbenches for creating content for all of oshun V1.0 to V1.4 please do thorought meticulous research and analysis to ensure the workbenches are fully equipped and have all features necessary to develop sota industry leading content in all applicable categories across all of tara, nyx, lilith, metis, nisaba, veritas, arete. for instance when it comes to meditation thinking deeply about how courses are strucutred, themes, develoipmental phases or whatever to call them, like for instance sam harris intro course on waking up like builds up concentration, and like basic awareness, then like does looking for the self and all sorts of other stuff, so kinda should be able to deconstruct approach of various traditions, build courses based on them but like malleable also in tweaking as appropriate and like veritas the news segments should be veyr well researhc and balanced, insiration some from groundnews and like analysis of events and sources and all sorts of requirements and like avatar design and all the creative brainstorming and design and so much i havent even mentioned but can you infer the type of deep analysis and research i need you to do?

How this document reads the brief: the question is not "which screens are missing" but "what must an authoring workbench be able to do so that what it produces is as good as the best in the world in that category". For every room that meant studying how the best producers design, structure, research, write, produce, check and improve that kind of content, then turning each practice into a workbench capability with a test that proves it.

1. Summary#

The finding in one line: Lilith's rooms can show content, but only Tara can author it. Even Tara authors in a structure too flat for the content you described. There is no Lilith persona to write for.

Where things stand today:

  • Tara has a real production workbench. Its courses, though, are flat day lists with no phases, techniques, traditions or progression, and the live room shows a hardcoded demo catalogue with invented ratings and credentials (§2.1).
  • Nyx, Arete and Nisaba have no editorial authoring workbench. Members can write their own Nisaba notebooks, and one studio page renders 3D lunar-phase briefings, but nothing authors publishable lessons, lore, tours, programmes, editions, commentary or study plans.
  • Arete has no programme model.
  • Metis's courseware plan is deep but lacks the learning-science and explainer craft that separates the best lessons.
  • Veritas's newsroom plan is rigorous but was written for a human newsroom. It contains nothing of what makes Ground News, Tangle or Semafor useful: outlet intelligence, coverage distribution, blindspots, perspectives, or event "known/unknown" ledgers.
  • There is no character-design tooling, and no definition of Lilith herself.

What state of the art means here is the same twelve-stage anatomy in every category (§5): explicit domain models, deconstructed exemplars, arcs with reasons, sentence-level evidence, blocking domain gates, named expert review, measured production, honest labels, and outcome and harm measurement. Your two examples show what this means in practice:

  • Meditation: the research deconstructs Waking Up, Healthy Minds, MBSR 2025, The Mind Illuminated, Goenka, Tergar, Zen, Tibetan, Christian, Sufi and yoga nidra paths into a common practice ontology. Courses are composed from tradition-derived arc templates that stay malleable through typed edits: compress, swap the object, reorder, insert a module, re-voice or adapt culturally. Safety and permission rules block; craft rules warn.
  • News: Veritas classifies every disputed point as fact (verify), contested evidence (weight it) or values (steelman each side). It rates outlets on dated, per-country axes (a US left–right line does not fit Ghana) and counts coverage by independent origin, not by syndicated copies. Every sentence traces to a frozen evidence pack.

The plan: a new board family of 315 tasks plus 8 epics across a shared studio core and seven pillars (Tara, Nyx, Arete, Nisaba, Metis, Veritas, and Lilith with the creative studio). It runs in five milestones, from M0 foundations and five urgent fixes to M4 horizon items. Every task carries the research's own "done when" test, and each room's launch passes through a fail-closed release gate that lists every control it needs. Shared engines are built once and reused by the rooms, by the V1.3 personal agent and by the V1.4 channels. Metis and Veritas tasks extend their existing ledger phases rather than duplicating them.

What needs you:

  • Five urgent fixes to the live product (filed as ready tasks on 2026-09-29): the Tara demo catalogue; intense breathwork (contraindicated for sickle-cell trait, which WHO Africa puts at 20–30% in Ghana); the Seven Habits trademark; non-commercial corpus licences; and the KJV served to UK members, a separate territory-rights issue (Crown letters patent apply in the UK only).
  • Around forty decisions in §9, each with a recommendation; Elikem accepted all of them on 2026-09-29. The ones that shape the most work are paid expert reviewers, licence posture, faith-integrated variants, how traditional knowledge is handled, and how human the AI teachers and guides may appear.

2. What exists today (code and board, 2026-09-28)#

Measured from the repository at main and the Eve board on 2026-09-28. Detail is in the research documents; the inventory below is what decides the scope.

Room Authoring workbench Content model Real content Board coverage
Tara Real: /studio/tara-workbench (13 pages, 70 BFF endpoints, 15 tables, 143 library tests): spark → premise → outline → script → grounded → review → approved → recorded → published; script blocks (intro, settling, body scan, breath cue, silence, bell, closing, free guidance); gates for invitational language, duration, grounding, captions; audio QA is an honest "unavailable" seam A course is a flat day-indexed lesson list: no phases, techniques, tradition or lineage fields, structured prerequisites or unlock rules. The only course guideline in Eve prescribes one arc for everything (Foundation → Core → Integration → Application) 1 course (7 days, 1 lesson recorded), 1 meditation, 1 teacher. The live Tara room shows a hardcoded demo catalogue (see §2.1) tara-content-workbench 113/115 closed
Nyx None for lessons, lore, tours or event copy (/studio/generation/nyx-3d renders lunar-phase briefings only) Lesson framework and a mythology model in code; 5 courses / 27 lessons hand-written in TypeScript; 18 hardcoded myths; 4 hardcoded planetarium tours Strong computed data: real ephemeris, events, nightly highlights None beyond V1.4 channel tasks
Arete None 48 tables of user data (goals, habits, journal, affirmations…). No program or course model and no behaviour-change technique taxonomy; affirmations are rule-based; the "coach" is pure functions Journal prompts only None beyond V1.4 channel tasks
Nisaba None for editorial content. Members can create their own notebooks (/nisaba/notebook/new → /v1/nisaba/notebooks), which the Yemaya study workspace reads ~850k lines of library code (editions, witnesses, apparatus, study plans, TEI, CTS, IIIF, 15 canons, 9 corpus connectors) with almost no app consumers 9 public-domain passages None beyond V1.4 channel tasks
Metis (V1.2) Split across a learner app and an admin app; host decision open Deep: curriculum briefs, concept/prerequisite graph, lesson blocks, item bank, IRT/BKT/FSRS-5, tutor configuration, lecture pipeline with Manim Hand-authored demo courses v1-domain-workbenches Phase M, 1,358 tasks, largely shipped
Veritas (V1.2) None (the web app is the consumer room; the CMS is memory-only with no auth) Three disjoint stores; retraction cascade real on the V1 side — v1-domain-workbenches Phase V, 1,471 tasks: 6 inventory items shipped, the rest later; a human-newsroom model that already plans agent roles (V3.17–V3.21)
Lilith (host) and characters None for character design No Lilith host-persona definition exists. A governance-grade persona registry (13 launch personas, 4 drafted Tara mentors with voice profiles and avatar packs) and a legacy personas table with prompt fields, but no create/update authoring, no moodboards, reference sheets or voice design — V1.4 scopes channel characters (C, D, V, P pillars)

Shared layers that exist: the workbench kit and shell (362 studio routes), a 12-state editorial lifecycle, versioning, taxonomy and translation memory (libs/oshun/studio-authoring), the creative-autonomy plane (idea synthesis, bandit portfolio, LLM/persona judges, review queue, steering) with only four content categories (veritas.story-explainer, nyx.sky-event-brief, nisaba.study-passage, tara.seasonal-ritual; none for Arete or Metis), Hathor ideation and Yemaya writers'-room libraries that no room studio uses, and Isis generation consoles that are mostly analyzers.

What the existing plans do not cover (checked by title search over 13,807 tasks in the five relevant families):

  • Veritas Phase V has no tasks for outlet lean or reliability ratings, coverage distribution, blindspots, ownership transparency, headline and framing comparison, perspective building, OSINT media verification, synthetic-media assessment, ClaimReview, event "known/unknown" ledgers, forecasts, regional "view from" perspectives or broadcast due impartiality.
  • Metis Phase M has no tasks for worked examples and fading, interleaving, multimedia-principle checks, metacognition, motivation design (one task), explainer craft or subject-specific pedagogy.
  • No plan covers Nyx, Arete or Nisaba editorial authoring, tradition deconstruction, course arcs, or character design for room guides. Lilith appears in the V1.4 plan only as the network host for the channels (lilith-v1-4:C02); no plan defines her persona across the app.

2.1 Found during this research: the live Tara catalogue is demo data#

apps/oshun/web/src/components/domains/tara/SessionLibrary.tsx renders DEFAULT_SESSIONS (the state is initialised from it and never replaced) and TaraCourses.tsx renders a hardcoded COURSES list. The entries carry invented ratings and play counts (for example "Wim Hof Power Breathing", rated 4.8 by 312 people, 18,700 plays), invented instructor biographies ("Navy SEAL-trained breathwork instructor and Wim Hof certified guide") and a breathwork course with Wim Hof, tummo and holotropic lessons. The only safety text is a note that the effects "vary from person to person". The one real teacher record in apps/tara/content/teachers/sarah-chen/metadata.json is no better: it credits the teacher with certifications from real institutions (Spirit Rock, the UMass Center for Mindfulness, Yoga Alliance), a book and thousands of students, and marks her verified, with no evidence behind any of it. Hyperventilation with breath retention is contraindicated for sickle cell disease or trait in the ILAE/IFCN clinical EEG standard, and WHO Africa puts sickle-cell trait at 20–30% in Ghana, the first audience. No board task covers replacing this catalogue. Tasks TA01 and TA02 below do, in M0.

3. How this was researched#

  • Inventory first. Code search over the rooms' content models, studios, generation paths and persona layer. A title-level coverage check over the 13,807 tasks in v1-domain-workbenches, tara-content-workbench, yemaya-study-workspace, lilith-v1-3 and lilith-v1-4, so that nothing here duplicates planned work.
  • Seven research passes, one per area, each producing a report with a numbered source list and a feature catalogue: Tara, Veritas, Nisaba, Nyx, Arete, Metis, and Lilith with the creative studio. Every feature carries what it does, why (with evidence), a priority (P0: without it the content is not state of the art or is unsafe; P1: strong differentiator; P2: later) and a testable "done when". 284 features were collected (Tara 50, Veritas 40, Nisaba 34, Nyx 32, Arete 35, Metis 53, creative studio 40). A second, web-search pass (below) added 20 more and revised 85, giving 304. The build folded features shared across rooms into the core pillar.
  • Primary sources were preferred: programme curricula and teaching guides, standards documents, peer-reviewed studies and meta-analyses, regulators' own texts, and product and help pages. Claims that could not be verified are marked "(unverified)" in the research documents.
  • Spot checks by the author of the claims that carry the most weight. The ILAE/IFCN clinical EEG standard does list "sickle cell disease or trait" among contraindications to hyperventilation. The FDA general-wellness guidance cited is dated 8 January 2026. Crown letters patent govern the KJV in the UK only. The Veritas report's figures for the EU AI Act Article 50(4) exemption, the EBU/BBC study and CNET's corrections match their sources. The DfE generative-AI product safety standards (updated 19 January 2026) do ask products to avoid names, avatars or characters that give an impression of personhood, except in time-limited, pedagogically justified roleplay.
  • A second pass with live web search. The session's web-search allowance ran out during the Nyx pass, so the Nyx, Arete, Metis and creative passes first worked from direct page fetches and the Crossref, OpenAlex, Semantic Scholar and PubMed APIs. At Elikem's request those four passes were redone with live web search: Codex, read-only, on the ChatGPT subscription, 2026-09-29, 676 searches in all.
    • The redo checked every claim marked unverified and the fifteen most decision-relevant claims per area, and filled the gaps left by sites that refused automated reads.
    • It corrected several claims. Examples: the Akan and Yoruba sky literature exists; the 6 February 2027 eclipse is annular over Accra; implementation intentions show an overall d = .36; the AI-coach comparison was not a head-to-head trial; Khanmigo now has field evidence; the DfE page is guidance; image models' reference limits; and C2PA is at 2.4.
    • It also added 20 features and revised 85.
    • Each supplement is appended to its research document (revision 2) and governs where it conflicts with the earlier text. Every revision appears in the affected task's body.
  • Limits. Some producer sites still refused automated reads, and a few items remain unverifiable; the supplements say which. This document avoids relying on claims the research marked unverified; where one matters, it says so.

4. Principles (binding on every task in this family)#

  1. Structure before prose. Every content item is composed from typed domain models: a practice, a sky claim, a behaviour-change technique, a text locator, a concept, an outlet or a perspective. Free text alone is never the source of truth. The models are what make content comparable, checkable, adaptable and reusable by the personal agent (V1.3) and the channels (V1.4).
  2. Every sentence is traceable. Claims bind to frozen evidence, exact locators or computations. Quotations of scripture, sources or people are inserted by reference and verified by exact match; an agent never types them.
  3. Traditions and communities hold authority over their own knowledge. Permissions are graded, restricted knowledge is never generated, named reviewers from the tradition sign off where the tradition requires it, and attribution is shown.
  4. Due weight, not false balance. Facts are verified, contested empirical questions are weighted by evidence, and value disagreements are given in their strongest good-faith forms. This applies to every room that makes claims, not only Veritas.
  5. Malleable within invariants. Templates derived from exemplars can be compressed, reordered, re-voiced or adapted. Craft invariants warn and require a written justification to override; safety and permission invariants block.
  6. Safety gates block; they do not advise. Contraindications, crisis signposting, claims rules, eye safety, restricted knowledge and licence limits stop publication until resolved.
  7. AI drafts; accountable humans decide; disclosure is recorded. Every published item records who or what drafted it, who reviewed it and at what depth, and the label the member sees (EU AI Act Article 50, applying since 2 August 2026, with a marking grace period to 2 December 2026 for systems already on the market).
  8. Quality comes from structure, gates and review, not from model size. Harnesses bind the cheapest adequate model through OpenRouter (deepseek/deepseek-v4-flash-0731, price-sorted) under the repository's cost rule.
  9. Ghana and the UK first. Location (the sky over Accra is not the sky over London), language, law (UK ASA/CAP and Ofcom, Ghana NMC and criminal false-news provisions), faith and culture are first-class inputs, not localisation afterthoughts.
  10. Honest states only. An unavailable gate says so; nothing ships with invented ratings, counts, credentials, scores or reviews. The live Tara demo catalogue (§2.1) is the counter-example this family removes first.
  11. Reuse before building. Extend the workbench kit, the Tara workbench library, the persona registry, the Nisaba libraries, Nyx's computation, the Metis and Veritas ledger phases and the V1.4 craft pillars. A task that overlaps an existing one names it and extends it rather than duplicating it.
  12. Measure outcomes and harms, not just engagement. Streaks and time-in-app are never success measures on their own; consented outcome and harm measures and expert review are.
  13. Thresholds are decided, not chosen by the worker. Where a sample size, pass mark or dataset is not set in a task, the first step is a measurement plan decided by Elikem or his delegate.
  14. Sensitive data and untrusted inputs have controls before first use. Health, faith, political and biometric data need a lawful basis and a retention rule before collection. Agents that read untrusted material or run generated code need a sandbox and adversarial tests.
  15. Review and publication are separate, attributed acts. Workers build, gate and submit. Independent reviewers review. The authorised publisher publishes.

5. What a state-of-the-art content workbench is#

Across all six rooms and the creative studio, the research found the same twelve stages. Leading producers do every one of them, and weak AI pipelines skip most of them. Each stage below names the shared engine (§6.1) and where the room-specific layer sits.

# Stage What the best producers do Shared engine Room layer (examples)
1 Know the domain Work from explicit models of the field CO02 knowledge models Tara practice ontology; Arete technique ontology; Veritas outlet registry; Nyx sky claims
2 Study exemplars Take the best work apart before making new work CO03 deconstruction canvas ≥12 tradition canvases; coaching programmes; explainer craft; balanced formats
3 Plan the catalogue Know what the audience needs and what is missing CO25 coverage map, CO24 calendars Seasonal programmes; sky-event planning; syllabus coverage
4 Ideate Diverge widely, then converge on evidence Tara divergence today; CR pillar for all rooms Programme briefs; show concepts
5 Design the structure Arcs with reasons: phases, fading, spacing CO04 arc composer, CO18 formats Tradition-derived course arcs; programme templates; study plans
6 Research and draft Every claim sourced; quotations exact CO05 evidence packs, CO06 quotation lock, CO20 agent pipelines Sky claims bound to computation; misconception-first scripts
7 Check Domain gates that block what is unsafe or wrong CO10 claims, CO08 authority, CO07 rights Contraindications; backfire checks; due weight; observability
8 Produce Measured audio and visual standards; human audition CO13 audio chain; CR pillar Voice direction; recitation; sonification
9 Review Named experts, blind scoring, sign-off tiers CO11 expert review MBI:TAC rubric; cross-spectrum panels; tradition reviewers
10 Publish honestly Disclosure of how it was made; licences enforced CO09 disclosure ledger, CO07 "How we know" panels; attribution renderers
11 Measure and learn Outcomes and harms, not only engagement CO21 measurement, CO12 evaluation Outcome instruments; accuracy audits; pilots
12 Correct and adapt Corrections propagate; content adapts within limits CO19 corrections, CO27 personalisation, CO28 channel feeds Loop-aware on-air corrections; V1.3 variants

6. Scope#

315 tasks and 8 epics, in eight pillars. Task keys below are the board keys in family content-workbenches. Each task's body carries what it does, why (with the research's evidence), a testable "done when", the existing code or tasks it builds on, and its dependencies. "Feature" names the source entry in the pillar's research document.

Pillar Tasks M0 M1 M2 M3 M4 Research
§6.1 Shared studio core (CO) 29 17 6 0 6 0 —
§6.2 Tara: contemplative content (TA) 46 2 21 0 20 3 Eve 13d4487f-57de-4523-81d6-e919e7b2afba
§6.3 Nyx: sky content (NX) 35 2 14 0 16 3 Eve 788820e4-8a29-450b-b31c-86104a08c7a4
§6.4 Arete: coaching content (AR) 35 2 21 0 9 3 Eve c479a954-f5e0-43ff-91f6-02520dbf7348
§6.5 Nisaba: text and study content (NI) 32 4 14 0 12 2 Eve 8563ce32-96b4-45d3-af99-5115633fb753
§6.6 Metis: learning content (ME) 60 0 0 31 26 3 Eve b325c39f-6ba1-4b49-98ba-c2d062dfde80
§6.7 Veritas: news and analysis content (VE) 40 0 0 23 14 3 Eve c15926ca-fd59-414b-a3bf-614c4551a235
§6.8 Lilith, characters and the creative studio (CR) 38 1 20 0 17 0 Eve 17508fb3-47c1-453a-a3f7-8bb1cf1fe950

6.1 Shared studio core [CO]#

Six rooms kept rediscovering the same needs, so they are built once. The research found that state-of-the-art content in every category has the same anatomy: it is composed from explicit models (a practice, a technique, a text locator, an outlet), structured as an arc whose order has reasons, bound to evidence sentence by sentence, checked by domain gates, reviewed by named experts, produced to a measured technical standard, labelled honestly, and measured on outcomes and harms. The shared core provides each of those as an engine; the room pillars supply the models, templates, gates and reviewers.

The acceptance tests from each room's research that belong to a shared engine are folded into that engine's task, so nothing a room needs is lost in the merge (for example CO08 carries the Tara, Nisaba and Nyx tests for permissions, restricted knowledge and cultural review).

Key Task Milestone Feature
CO01 Content studio architecture ADR M0 —
CO02 Knowledge-model registry and editor M0 —
CO03 Exemplar deconstruction canvas M0 —
CO04 Arc composer with invariants and typed edit operations M0 —
CO05 Frozen evidence packs and sentence-level provenance M0 —
CO06 Quote-by-reference lock M0 —
CO07 Rights and licence registry with render-time enforcement M0 —
CO08 Tradition and community authority registry M0 —
CO09 AI-disclosure and editor-of-record ledger M0 —
CO10 Market claims engine M0 —
CO11 Expert and panel review workflow M0 —
CO12 Evaluation harness with per-room golden sets M0 —
CO13 Audio production chain and human audition gate M0 —
CO14 Crisis-resource registry with verification M0 —
CO15 Sensitive-data governance for content features M0 —
CO16 Security for untrusted content, agents and code execution M0 —
CO17 Audience and age policy enforcement for content M0 —
CO18 Format library with structural slots and per-format gates M1 —
CO19 Corrections, errata and re-render propagation M1 —
CO20 Agent pipelines for every room through the workbench gates M1 —
CO21 Consented outcome and harm measurement M1 —
CO22 Harm-incident intake, triage and content withdrawal M1 —
CO23 Jurisdiction and age policy compiler M1 ARE-39
CO24 Multi-calendar and event-driven planner M3 —
CO25 Catalogue coverage and gap map M3 —
CO26 Ghanaian-language content workflow M3 —
CO27 Personalisation rules and variants for the personal agent M3 —
CO28 Channel feed contract from workbenches to V1.4 M3 —
CO29 Operational evidence programme: pilots, panels and cohorts M3 —

6.2 Tara: contemplative content [TA]#

Research: Eve document 13d4487f-57de-4523-81d6-e919e7b2afba.

What the best do.

  • Every leading programme is built in phases, not flat day lists, and the usual order is stabilise attention → open awareness → heart practice and insight → integration. Examples: the Healthy Minds Program trial runs four weeks of Awareness first because attention is "foundational"; Tergar runs Calm → Heart → Wisdom; Goenka runs ethics → breath (days 1–3) → body sensations (from day 4) → loving-kindness last; Diana Winston's Waking Up course runs focused → investigative → choiceless → natural awareness. Waking Up's own Introductory Course (rebuilt from 50 to 28 days in August 2020) pairs a short lesson with a practice each day and brings in "looking for the one who is looking" only after concentration groundwork; its day-by-day list is not publicly documented, and the report says so.
  • Where heart practice goes is a genuine design choice. Goenka puts it last, Tergar in the middle, Mindful Self-Compassion builds on it and Healthy Minds branches after Awareness. The workbench treats placement as a parameter with the traditions' reasons attached.
  • Guidance fades on a documented schedule. The MBSR 2025 Curriculum & Teaching Guide (Kabat-Zinn, v1.0, September 2025) introduces unguided practice from week 2 and fully self-structured 45-minute practice by week 7. Headspace Pro removes prompts level by level, and the MBI:TAC teacher rubric expects silence to lengthen with experience.
  • MBI:TAC is a ready-made rubric for guidance language. It flags striving words and promised outcomes, prefers "the breath" to "your breath", and asks for a pace slower than conversation but "not hypnotic".
  • Theory runs as a parallel strand. Waking Up's Theory section, Healthy Minds lessons and MBSR dialogue each run alongside practice. Insight Timer rejects "courses" that lack depth and progression.
  • Adverse effects are common enough to design for. In a US population sample, 50% reported at least one specific effect, 10.6% some impairment and 1.2% impairment lasting a month or more. In MBCT samples, 6–14% had lasting bad effects. Intensity, practice approach and the relationship to a teacher are the top risk factors, and an app has no teacher by default.
  • Hyperventilation breathwork has hard contraindications. They include sickle cell disease or trait (20–30% trait prevalence in Ghana). Slow, exhale-weighted breathing has the best, modest evidence: cyclic sighing, and resonance breathing at about 6 breaths per minute.
  • Claims are tightly regulated. The UK ASA/CAP code allows only mild claims such as relaxation or aiding sleep. Under the FDA's January 2026 wellness guidance, anxiety-disorder treatment is a medical device claim.

What the workbench must add. A practice ontology and tradition deconstruction canvases (from the research: family, attentional style, phenomenological matrix, sense grid, technique, object, instruction verbs, stance, view, prerequisites, signs of progress, hindrance→remedy pairs, contraindications, adverse-effect watch-list, trauma adaptations, lineage and permission). Also phase-structured course arcs derived from those canvases, guidance-fading and language gates, and contraindication propagation. Arcs stay malleable through typed operations: compress, swap object, reorder, insert module, re-voice or adapt culturally. Craft invariants warn; safety and permission invariants block.

Key Task Milestone Feature
TA01 Replace the Tara demo catalogue with reviewed content M0 —
TA02 Gate intense breathwork until screening exists M0 —
TA03 Practice Ontology Registry and Editor M1 TD-01
TA04 Tradition Deconstruction Canvas M1 TD-02
TA05 Provenance Tiers for Tradition and History Claims M1 TD-05
TA06 Phase-Structured Course Schema M1 AC-01
TA07 Tradition-Derived Arc Templates with Invariants M1 AC-02
TA08 Guidance-Density Metrics and Fading-Curve Gate M1 PF-01
TA09 MBI:TAC-Grade Language Linter, Content-Type Aware M1 SC-01
TA10 Practice-Kind Checklists as Gates M1 SC-02
TA11 Pacing Calibration from Rendered Audio M1 SC-03
TA12 Contraindication Propagation Engine M1 SF-01
TA13 Breath-Protocol Registry with Evidence and Risk Classes M1 SF-02
TA14 "Never" Hard-Rule Suite M1 SF-03
TA15 Trauma-Sensitive Variant Authoring M1 SF-04
TA16 Dose and Intensity Envelope M1 SF-05
TA17 Consent, Screening and Signposting Copy per Course and Market M1 SF-06
TA18 Persona Honesty Guard M1 LP-05
TA19 Expert Review Rubric M1 EV-01
TA20 Produce the V1.0 Tara launch catalogue through the workbench M1 —
TA21 Malleable Edit Operations with Constraint Checking M1 AC-03
TA22 Tara release gate: every required control in place M1 —
TA23 Server-enforced practice eligibility M1 —
TA24 Cross-Tradition Comparison View M3 TD-03
TA25 Stage-Map and Progress-Signal Library (authoring-side) M3 TD-04
TA26 Prerequisite and Unlock Graph M3 AC-04
TA27 Theory–Practice Rhythm Planner M3 AC-05
TA28 Difficulty and Plateau Module Library M3 AC-06
TA29 Cultural Adaptation Layer M3 AC-07
TA30 Self-Guided and Timer Sessions as Course Items M3 PF-02
TA31 Continuity and Callback Tracker M3 PF-03
TA32 Sleep-Content Craft Rules M3 SC-04
TA33 Adverse-Effect Monitoring Hooks and Editorial Alerts M3 SF-07
TA34 Teacher and Source Reputation Notes M3 LP-03
TA35 Attribution Renderer M3 LP-04
TA36 Voice Direction and TTS Prosody Profiles M3 AU-03
TA37 Music, Soundscape and Night-Listening Registry M3 AU-04
TA38 Variant Generation within Invariants M3 PZ-02
TA39 Practitioner Test Panel M3 EV-02
TA40 Outcome and Harm Measurement Plan M3 EV-03
TA41 Yoga Sequence Builder with Pose Ontology M3 OC-02
TA42 Theory, Dharma-Talk, Q&A and Live-Class Authoring M3 OC-03
TA43 24/7 Channel Programming Grid and Stream Safety M3 OC-05
TA44 Adaptive Progression Rules M4 PZ-03
TA45 Intervention Description Export (TIDieR/BCIO) M4 EV-05
TA46 Mantra and Chant Provenance and Pronunciation M4 OC-04

6.3 Nyx: sky content [NX]#

Research: Eve document 788820e4-8a29-450b-b31c-86104a08c7a4.

What the best do.

  • The best sky content is short, dated and follows a fixed shape. NASA JPL's "What's Up" (monthly since 2007, about 3–4 minutes, always transcribed) runs teaser → 4–6 dated beats → Moon-phase chart → fixed sign-off. It moves from an easy object to a hard one and gives directions by horizon rather than coordinates. Its hedges ("you will need binoculars", "lineup", not "alignment") amount to a reusable phrase library.
  • One script must render correctly for different skies. Accra sits at about 5.6°N on GMT all year; London is at about 51.5°N and switches between GMT and BST. Canopus and the Southern Cross rise in Accra and never in London, and London has no full astronomical darkness near midsummer. EarthSky writes for northern mid-latitudes; apps such as timeanddate, In-The-Sky and Sky Guide already localise event circumstances, so Nyx's edge is not localisation itself but the editorial workflow around it: one reviewed script, computed variants per place, explicit conditions and corrections that propagate. The next big test is close: the annular solar eclipse of 6 February 2027 passes over Accra (NASA's detailed path tables; timeanddate classifies Accra as annular), a major Ghana event that needs city-specific classification and eye-safety content.
  • Data sources have traps and changed in 2026. JPL Horizons separates astrometric from apparent positions, and meteor ZHR is never what a viewer sees. CelesTrak exhausted five-digit catalogue numbers on 11 July 2026, so new objects exist only in modern formats, and it changed its default output to CSV on 9 May 2026, so a request without an explicit format is another silent failure. Horizons observer tables default to UT, not TT, so each numeric claim must state its timescale. Hazard figures move fast: asteroid 2024 YR4's lunar-impact chance was ruled out only in March 2026.
  • WorldWide Telescope is the model for tour authoring: slides with start and end cameras pinned to a sky time, keyframes, per-slide narration checks, dome preview, and a credited science advisor.
  • Cultural sky lore needs provenance and partners. Stellarium's sky-culture model gives each culture a classification, per-name references, native spelling, named knowledge-holders and a licence. A literature exists for West African sky knowledge: Holbrook's Cultural Astronomy in Africa South of the Sahara covers Akan and Yoruba material, and Sègla has published on Yoruba ethnoastronomy (the first research pass, working without web search, missed both). It is thin, and it does not replace contemporary community review: finding a source and having permission to publish are different steps. "Sika", the Ghanaian IAU star name, came from a 2019 naming campaign, not tradition. The Dogon–Sirius story is disputed. The West African strand therefore starts from this literature and needs partners, never model recall.
  • Learning and access have concrete targets. Moon phases, seasons and gravity are the hardest ideas for learners. Ghana's JHS science covers the Solar System and eclipses. NASA's Chandra sonifications ship with a caption explaining the data-to-sound mapping, and a 3,184-person survey found higher self-reported learning, engagement and trust (self-report, not a measured learning gain, so Nyx tests comprehension separately).
  • Images carry different licences. ESO and ESA/Webb/Hubble images are CC BY 4.0 with the credit burned into video. ESA's licences vary by asset (CC BY-SA 3.0 IGO permits commercial use with share-alike; other ESA terms restrict it), so each image's licence is recorded. NASA forbids "according to NASA" in AI products.

What the workbench must add. Sky claims marked up and bound to computations, rendered per audience location, with an observability evaluator, a vetted claims linter and data freshness gates. Also a lesson builder that works from misconceptions, a tour and camera editor, cultural lore intake with consent and takedown, sonification design, and an imagery registry with licence and "illustration versus photograph" labels.

Key Task Milestone Feature
NX01 Nyx workbench host M0 —
NX02 Data-source registry with freshness gates M0 NYX-05
NX03 Migrate hand-coded Nyx lessons, myths and tours into the workbench M1 —
NX04 Sky-claim markup bound to computations M1 NYX-01
NX05 Audience-location matrix and one-script-to-many rendering M1 NYX-02
NX06 Observability evaluator ("can this viewer actually see it?") M1 NYX-03
NX07 Claim linter with a vetted phrase library M1 NYX-04
NX08 Rate, risk and space-weather calculators for copy M1 NYX-06
NX09 Accuracy evaluation harness M1 NYX-07
NX10 Misconception-aware lesson builder M1 NYX-13
NX11 Lore intake and provenance record M1 NYX-17
NX12 Accessibility checker for sky media M1 NYX-22
NX13 Imagery and media registry with licence and image-kind labels M1 NYX-23
NX14 Nyx release gate: every required control in place M1 —
NX15 Celestial-object evidence dossier M1 NYX-36
NX16 Annular eclipse of 6 February 2027: Ghana safety and content package M1 —
NX17 Event-calendar-driven content planner M3 NYX-08
NX18 Format templates with structural slots M3 NYX-09
NX19 Sky-tour and camera-path editor M3 NYX-10
NX20 Frame-of-reference switcher for explanatory scenes M3 NYX-11
NX21 Curriculum alignment tagging and coverage M3 NYX-14
NX22 Observing-programme builder M3 NYX-15
NX23 Star-name resolver with provenance classes M3 NYX-19
NX24 West African sky-knowledge research and partnership ledger M3 NYX-20
NX25 Sonification designer M3 NYX-21
NX26 Live-data segment templates with honest degrade states M3 NYX-25
NX27 Observation-feedback calibration M3 NYX-29
NX28 Contested-topic and pseudo-science dossiers M3 NYX-31
NX29 Units, glossary and language layer M3 NYX-32
NX30 Observing field-pack publisher M3 NYX-33
NX31 Historical-sky validity dossier M3 NYX-34
NX32 Presenter cues and rehearsal desk M3 NYX-35
NX33 Multi-target render and export (phone, AR, VR, dome) M4 NYX-12
NX34 Citizen-science mission templates M4 NYX-16
NX35 Sky-driven 24/7 schedule composer and sleep-safe rules M4 NYX-27

6.4 Arete: coaching content [AR]#

Research: Eve document c479a954-f5e0-43ff-91f6-02520dbf7348.

What the best do.

  • Leading programmes use several patterns, and the strongest sequence content deliberately. Noom, BetterUp and MasterClass use lesson libraries, assessment-led coaching and project courses, so no single arc is universal. Yale's Science of Well-Being runs six modules: misbeliefs first, then biases, then practices, then implementation. Action for Happiness runs six weekly two-hour sessions, each "Tune In / Connect / Take Action". Tiny Habits is a five-day sprint with three tiny habits and coach feedback. FranklinCovey sells the 7 Habits course in fixed formats.
  • Plain techniques carry the best evidence. Implementation intentions: overall d = .36 across 642 tests (the often-quoted .27–.66 range spans different outcome types). Progress monitoring: 138 RCTs, goal attainment d = .40, stronger when progress is recorded or reported. Goal setting: overall d = .34 (difficulty, public commitment and group goals are moderators). WOOP/mental contrasting: g = .34, rising to .47 when delivered interactively rather than as a document, which favours dialogue delivery.
  • Popular "positive" content is weaker than people assume. After correcting for small studies, positive-psychology interventions average about r = .10. Gratitude beats doing nothing (about d = .17 against alternative activities) but not psychologically active comparisons (d = −.03), and Three Good Things matched placebo in a replication.
  • Some popular content backfires. Positive self-statements made low-self-esteem participants feel worse, and positive fantasies about the future predicted worse outcomes. No controlled vision-board trial was found. Values-based self-affirmation is the evidence-backed alternative.
  • Habits take months, not weeks. In the few studies that measured it, median formation times were about 59–66 days (range 4–335; the medians come from different studies, not one pooled timetable), and missing one day did not matter; "21 days" copy is wrong. Showing a broken streak lowers engagement and prompts self-blame. The UK CMA's taxonomy of 21 manipulative design practices is a usable audit.
  • AI coaching has two sides. In two separate trials run ten months apart, a structured AI coach (Coach Vici) produced goal-attainment gains similar to human coaching; this was not a head-to-head trial. But warmth-tuned models validated false beliefs more (Nature, April 2026), and heavier chatbot use tracked with loneliness and dependence. The ICF Code of Ethics (April 2025) now covers AI.
  • Culture changes framing, not mechanisms. In Ghana, family and community goals, faith cues (71.3% Christian, 19.9% Muslim) and low-data formats matter. Prosperity-gospel "declare and receive" affirmations are a positive-fantasy risk.

What the workbench must add. A programme content model on the arc composer, and technique tagging against the BCT Ontology with COM-B mechanism mapping. An evidence registry that grades every technique (with an explicit "backfires" grade), and plan builders that validate the structure of if–then plans and WOOP. Check-in, lapse and re-entry design, and affirmation and prompt libraries with backfire checks. A safety and scope-of-practice gate, a verified crisis registry, persona-coach methodology specs with a red-team harness, and anti-dark-pattern gamification rules.

Key Task Milestone Feature
AR01 Arete workbench host M0 —
AR02 Decide and fix the Seven Habits trademark exposure M0 —
AR03 Programme content model and arc composer M1 ARE-01
AR04 BCT tagging with ontology identifiers M1 ARE-02
AR05 Mechanism map and COM-B coverage checker M1 ARE-03
AR06 Evidence registry and technique grading M1 ARE-04
AR07 Programme-arc templates by goal type M1 ARE-06
AR08 Plan builders with structural validators M1 ARE-07
AR09 Habit design spec and automaticity checkpoints M1 ARE-08
AR10 Check-in and review script designer M1 ARE-09
AR11 Lapse, relapse and re-entry designer M1 ARE-10
AR12 Affirmation and self-talk library with backfire checks M1 ARE-11
AR13 Reflection and journal prompt designer M1 ARE-12
AR14 Safety classification and scope-of-practice gate M1 ARE-13
AR15 Persona-coach methodology specs M1 ARE-15
AR16 AI-coach red-team and assurance harness M1 ARE-16
AR17 Anti-dark-pattern gamification rules M1 ARE-17
AR18 Cultural adaptation workbench M1 ARE-21
AR19 Food, body and exercise content guard M1 ARE-24
AR20 Produce the first Arete programmes through the workbench M1 —
AR21 Constrained generation pipeline M1 ARE-29
AR22 Arete release gate: every required control in place M1 —
AR23 Coaching agreement and review scripts M1 ARE-38
AR24 Coaching-conversation flow authoring with MI-fidelity scoring M3 ARE-18
AR25 JITAI rule designer M3 ARE-19
AR26 Faith-integrated variant authoring M3 ARE-22
AR27 Language, register and reading-level tools M3 ARE-23
AR28 Agora format templates M3 ARE-30
AR29 Challenge designer M3 ARE-31
AR30 Over-reliance safeguards authoring M3 ARE-32
AR31 Practice artifacts and transfer rubrics M3 ARE-36
AR32 Emotion vocabulary and strategy authoring M3 ARE-37
AR33 Micro-randomised trial and experiment manager M4 ARE-20
AR34 Worksheet and offline export M4 ARE-33
AR35 Living evidence review M4 ARE-35

6.5 Nisaba: text and study content [NI]#

Research: Eve document 8563ce32-96b4-45d3-af99-5115633fb753.

What the best do.

  • Leaders separate AI drafting from publishing and label it. Sefaria (2026) shows one icon for AI-assisted content that a human has reviewed and another for fully AI-generated content. Its first AI-assisted translation had a six-month human edit. 84000 forbids AI first drafts of canonical translations and requires a prompt log for each project.
  • Every serious platform stores text as stable segment identifiers with separate layers. SuttaCentral's Bilara, Sefaria references, Quran.com word-level audio timings and Perseus each keep translation, notes, variants, audio and morphology as separate layers on the same segments.
  • Machine translation is good enough to prioritise review, not to publish. A September 2026 study of 46,479 Pali translations found 2.9% major errors. A judge panel caught 81.6% of them by reviewing only the riskiest 10%.
  • General models misquote sacred text. YouVersion reports misquotation rates of 15–60%, and IslamicEval counts one wrong diacritic as a hallucination. The fix is structural: quotation by reference and exact match.
  • Licences are already a live problem. At least three of the nine existing connectors carry non-commercial terms (CBETA, ctext, some Sefaria versions), and OpenITI probably does too. The KJV Psalms Lilith serves today fall under Crown letters patent in the UK; how far that reaches digital display needs a rights decision.
  • Religious authority treats AI as a tool, not an interpreter. Dar al-Ifta (January 2026) ruled against relying on AI to interpret the Qur'an. The Tanzil licence forbids any change to the text. The Buddhist Digital Resource Center seals texts at lineage holders' request, and Ifá is transmitted by priests.
  • Study plans have measurable rules. YouVersion plans run 3–365 days, and completion peaks between 3 and 21 days. Each day carries at most 400 words and at least one reference, with no pasted scripture and one call to action at the end. Sefaria's source sheet is the reference lesson format.
  • Comparative study has a method. Freiberger's five steps and the scholarly critique of perennialism become checks: describe each tradition in its own terms first, and state difference as well as similarity.

What the workbench must add. A segmented, layered text store with locator-verified quotation, and a tradition policy engine that governs the existing nisaba.study-passage generator. A connector licence audit (urgent), edition principles and review, and a segment-aligned translation editor with versions and a risk-ranked machine-translation lane. Also layered commentary, a source-sheet composer, a study-plan composer with the publisher's rules, comparative study with anti-flattening gates, and recitation audio under tradition sign-off.

Key Task Milestone Feature
NI01 Nisaba workbench host M0 —
NI02 Connector licence audit and quarantine M0 NIS-07
NI03 Serve existing passages by territory (KJV for UK members) M0 —
NI04 Guard member surfaces against non-commercial corpus material now M0 —
NI05 Segment store with cognate layers M1 NIS-01
NI06 Tradition policy engine M1 NIS-05
NI07 Editorial-principles record and edition review checklist M1 NIS-11
NI08 Segment-aligned translation editor (CAT) M1 NIS-13
NI09 Translation versions and publication workflow M1 NIS-14
NI10 Machine-translation lane with risk-ranked review M1 NIS-15
NI11 Layered commentary and annotation manager M1 NIS-17
NI12 Source-sheet composer M1 NIS-23
NI13 Study-plan composer with publisher lint and analytics M1 NIS-24
NI14 Comparative-study composer with anti-flattening gates M1 NIS-26
NI15 Recitation, reading and pronunciation audio M1 NIS-27
NI16 Evaluation harness M1 NIS-30
NI17 Produce the first Nisaba source sheets and study plans M1 —
NI18 Nisaba release gate: every required control in place M1 —
NI19 DTS 1.0 endpoint and cross-scheme resolver M3 NIS-03
NI20 Collation workbench M3 NIS-09
NI21 Uncertainty and restoration encoding M3 NIS-12
NI22 Translator-notes layer M3 NIS-16
NI23 Multivocality linter M3 NIS-18
NI24 Cross-corpus parallel and cross-reference graph M3 NIS-19
NI25 Morphology and lexicon correction queue M3 NIS-20
NI26 IIIF manuscript and HTR workspace M3 NIS-21
NI27 Paleography lesson builder M3 NIS-22
NI28 Explainer and documentary pipeline with claim ledger M3 NIS-28
NI29 Script and Unicode checker M3 NIS-32
NI30 Standards export bundle M3 NIS-33
NI31 Variant classification and stemma testing M4 NIS-10
NI32 Reading calendars M4 NIS-25

6.6 Metis: learning content [ME]#

Research: Eve document b325c39f-6ba1-4b49-98ba-c2d062dfde80.

What the best do.

  • Learning-science gates must encode boundary conditions, not slogans. Interleaving helps maths (a preregistered trial scored 61% versus 38%, d = 0.83) but blocking beats it for vocabulary (g = −0.39). Productive failure reverses for younger learners, and worked examples reverse for experts. "Bloom's 2 sigma" is an outlier; later reviews put human tutoring at d = 0.79 and pooled tutoring at 0.37 SD.
  • Misconception-first teaching is the best-evidenced explainer craft. In Derek Muller's experiments, videos that stated and refuted misconceptions beat clear exposition (effect sizes 0.79–0.83), and "interesting extras" did not help. Seductive details lower learning, so "wonder" content needs an objective lock.
  • Feedback content matters more than its presence. Over a third of feedback interventions made performance worse, and person-directed feedback is weakest. Feedback should be typed and linked to the misconception it addresses.
  • Math Academy is the structural exemplar. Lessons break into knowledge points (worked example, then 2–5 adaptive questions). "Encompassing" graph edges let practice on a harder topic count as review of simpler ones. Pass rates are calibrated to about 80%, and it avoids LLM dialogue by design.
  • Oak National Academy's Aila is the best public governed lesson generator. It uses section agents, learning-cycle rules, strict question design and age-adjusted moderation. A separate EEF/NFER trial (259 teachers, ChatGPT with an implementation guide, not Aila) cut planning time by 31%; its limited quality sample does not validate Aila or any generator.
  • Tutor design matters more than the model. Unrestricted GPT-4 cut later unaided scores by 17%, and teaching failures rise from 17.7% in single-turn to 77.8% in multi-turn dialogue. Low-bandwidth WhatsApp tutoring (Rori, Ghana) showed 0.37 SD gains.
  • England's DfE generative-AI product safety standards (government guidance, updated January 2026, not a statute) ask AI education products to avoid names, avatars or characters implying personhood, with no general exemption for fictional labels; the only exception is time-limited, pedagogically justified roleplay. They ban engagement-prolonging design and require progressive disclosure of answers. This bears directly on V1.4's faculty of AI lecturers.
  • AI tutors now have field evidence, and it is modest. An August 2026 study of Khanmigo found about 0.06–0.08 SD over a school year, similar to Khan Academy practice without AI; although 96% of students tried it, the median learner used it in only 17% of exercise sessions that involved a mistake. Unaided retention must be measured separately from assisted performance.
  • Ghana's NaCCA curriculum is machine-alignable (strand → sub-strand → content standard → indicator). WAEC's Chief Examiner reports are a free source of documented misconceptions.

What the workbench must add. Knowledge-component typing that every gate reads, with worked-example fading, retrieval and spacing, interleaving, multimedia-principle and feedback-quality gates. Also a misconception-first script builder with a claim-level research dossier, subject templates (maths CPA, languages, science practicals, history sources, programming), Ghana and UK alignment packs, a manipulation and anthropomorphism copy lint, and a layered evaluation harness with multi-turn tutor tests. Everything extends the shipped Phase M ledger rather than replacing it.

Key Task Milestone Feature
ME01 ADR: learning-science and explainer layer on the Metis courseware ledger M2 —
ME02 Knowledge-component typing M2 MET-LS-01
ME03 Worked-example and fading planner M2 MET-LS-02
ME04 Retrieval and spacing plan M2 MET-LS-03
ME05 Retrieval card generator M2 MET-LS-04
ME06 Interleaving planner M2 MET-LS-05
ME07 Multimedia-principle lint M2 MET-LS-06
ME08 Typed feedback and feedback-quality lint M2 MET-LS-09
ME09 Backward-design and coverage gate M2 MET-LS-12
ME10 Narrative structure templates M2 MET-EX-01
ME11 Misconception-first script builder M2 MET-EX-02
ME12 Research dossier with claim-level sourcing M2 MET-EX-03
ME13 Maths CPA-and-variation template M2 MET-SUB-01
ME14 Exam-technique kit M2 MET-SUB-06
ME15 Ghana pack M2 MET-ALN-01
ME16 UK pack M2 MET-ALN-02
ME17 Manipulation and anthropomorphism copy lint M2 MET-ENG-02
ME18 Rendition budget and text-first equivalence M2 MET-ACC-01
ME19 WCAG 2.2 interaction checks M2 MET-ACC-02
ME20 Grounded generation M2 MET-GEN-01
ME21 Claim extraction and factual verification M2 MET-EV-01
ME22 Pedagogy rubric judge with human calibration M2 MET-EV-02
ME23 Multi-turn tutor evaluation suite M2 MET-EV-03
ME24 Item-writing-flaw checker M2 MET-EV-04
ME25 Moderation with age-adjusted categories M2 MET-EV-05
ME26 Faculty persona pedagogy spec M2 MET-AC-01
ME27 Produce the first Metis courses and explainers through the workbench M2 —
ME28 Metis release gate: every required control in place M2 —
ME29 Evidence applicability and exception register M2 MET-LS-15
ME30 Unaided retention and transfer pilot protocol M2 MET-EV-09
ME31 Session boundaries and offline continuation authoring M2 MET-AC-06
ME32 Cognitive-load estimator M3 MET-LS-07
ME33 ICAP and activity-utility mix M3 MET-LS-08
ME34 Predicted-success and expertise variants M3 MET-LS-10
ME35 Lesson-model templates M3 MET-LS-11
ME36 Prerequisite remediation links M3 MET-LS-13
ME37 Visual metaphor and analogy library M3 MET-EX-04
ME38 Storyboard–narration alignment check M3 MET-EX-05
ME39 Segment and pacing planner M3 MET-EX-06
ME40 Explorable and simulation authoring kit M3 MET-EX-07
ME41 Language comprehensible-input template M3 MET-SUB-02
ME42 Science practical and simulation template M3 MET-SUB-03
ME43 History source-analysis template M3 MET-SUB-04
ME44 Programming PRIMM and Parsons template M3 MET-SUB-05
ME45 Curriculum edition and effective dates M3 MET-ALN-03
ME46 Engagement-mechanics register M3 MET-ENG-01
ME47 UDL 3.0 and representation review M3 MET-ACC-04
ME48 Template-slot generation with multi-candidate selection M3 MET-GEN-02
ME49 Experiment register and ethics M3 MET-EV-08
ME50 Quiz-night kit M3 MET-AC-02
ME51 Office-hours kit M3 MET-AC-03
ME52 Revision-marathon planner M3 MET-AC-04
ME53 Wonder-explainer objective lock M3 MET-AC-05
ME54 Video-to-lesson companion builder M3 MET-EX-08
ME55 Bounded communicative-task authoring M3 MET-SUB-07
ME56 Interactive-state tutor contract M3 MET-GEN-04
ME57 Tutor uptake and missed-help diagnostics M3 MET-EV-10
ME58 Scheduler upgrade to FSRS-6/7 M4 MET-LS-14
ME59 Value-relevance prompt authoring M4 MET-ENG-03
ME60 Audio-dialogue rendition M4 MET-GEN-03

6.7 Veritas: news and analysis content [VE]#

Research: Eve document c15926ca-fd59-414b-a3bf-614c4551a235.

What the best do.

  • Ground News makes no ratings of its own. Its outlet bias label averages AllSides, Ad Fontes and Media Bias/Fact Check. Labels are per outlet, never per article. CJR reported in September 2025 that at least two of those raters had no formal agreement with it. Its Blindspot formula is unpublished, and the evidence that it changes minds is mixed.
  • The raters that people trust use blind, viewpoint-mixed review. AllSides uses blind surveys and multipartisan panels. Ad Fontes has left, centre and right analysts rate each piece (94,600+ items). NewsGuard scores nine weighted criteria and asks the outlet for comment before publishing.
  • A US left–right axis does not fit Ghana. About a third of monitored Ghanaian outlets were state-owned or tied to NPP or NDC owners (Media Ownership Monitor 2017). The January 2025 change of government flipped "pro-government", so outlet axes must be multiple, per country and dated.
  • Balance means due weight, not equal time. The BBC guidelines, the 2011 Jones review and Ofcom Section Five all say so. Every disputed point is classified first: a fact is verified, not balanced; a contested empirical question is weighted by evidence; a value disagreement gets the strongest good-faith version of each side.
  • The best balanced formats share one skeleton: facts first, then attributed perspectives, a clearly labelled view or none, dissent, the view from elsewhere, what is under-covered, and visible corrections. Tangle, Semafor's Semaform, AllSides roundups and Kagi News all follow it.
  • AI news failures are documented. CNET corrected 41 of 77 AI stories. Apple paused its notification summaries in January 2025 after invented BBC headlines. The EBU/BBC study (October 2025) found significant issues in 45% of AI answers, and its failure taxonomy doubles as an audit rubric.
  • Detection tools cannot give verdicts on their own. Deepfake detectors fail on real political fakes. C2PA (now at version 2.4, April 2026) shows who signed a file, not whether it is true. SIFT lateral reading, reverse search, geolocation, chronolocation and archiving remain the backbone.
  • Most coverage is recycled. More than 60% of UK quality-press stories came mainly from wire or PR copy, so counts must collapse syndication and flag single-source claims.
  • Law and labels have moved. EU AI Act Article 50(4) has applied since 2 August 2026; AI-written public-interest text is exempt from labelling only with substantive human review and a named responsible editor. Ghana carries criminal false-news exposure.

What the workbench must add. An outlet registry with ownership graph and a published rating method, and origin-aware clustering with coverage distribution and a published blindspot formula. The balance engine: an epistemic classifier, a due-weight evidence ledger and a steelman builder with a symmetry audit. Also event dossiers with a known/unknown/watch ledger, a SIFT source dossier and OSINT case files, and agent desks with separated roles that extend Phase V's agent plans (V3.17–V3.21). Every derived output gets re-verification, alongside a segment format library, "how we know" panels, an EBU-taxonomy accuracy audit and broadcast impartiality controls for the V1.4 Newsroom channel. Everything extends the Phase V newsroom ledger rather than duplicating it.

Key Task Milestone Feature
VE01 ADR: the AI-first desk model on top of the Veritas newsroom plan M2 —
VE02 Outlet registry with dated, per-country multi-axis profiles M2 V-01
VE03 Ownership and funding graph M2 V-02
VE04 Rating methodology, review panels, right of reply and versioning M2 V-03
VE05 Origin-aware story clustering and coverage distribution M2 V-06
VE06 Independent-confirmation counter and single-source flag M2 V-07
VE07 Epistemic classifier (A: fact / B: contested empirical / C: values), with editor confirmation M2 V-11
VE08 Due-weight evidence ledger for contested empirical questions (class B) M2 V-12
VE09 Steelman perspective builder with fairness checks (class C) M2 V-13
VE10 Event dossier builder M2 V-15
VE11 Known / Unknown / Watch ledger with change diffs M2 V-16
VE12 Lateral-reading source dossier (SIFT) M2 V-18
VE13 Media verification case file (OSINT toolkit) M2 V-19
VE14 Synthetic-media assessment protocol M2 V-20
VE15 Evidence preservation and chain of custody M2 V-21
VE16 Agent desk orchestration with separated roles M2 V-25
VE17 Temporal freshness guard M2 V-28
VE18 Derivative re-verification (headline, notification, ticker, audio, translation) M2 V-29
VE19 Segment format library with per-format gates M2 V-33
VE20 "How we know" transparency panel authoring M2 V-34
VE21 EBU-taxonomy accuracy audit programme M2 V-37
VE22 Produce the first balanced Veritas segments through the workbench M2 —
VE23 Veritas release gate: every required control in place M2 —
VE24 Content-sampled rating pods (AI plus human) M3 V-04
VE25 External trust-signal crosswalk M3 V-05
VE26 Blindspot detector with a published formula M3 V-08
VE27 Headline and framing comparator M3 V-09
VE28 "View From" regional and international perspective retrieval M3 V-10
VE29 Cross-perspective symmetry audit M3 V-14
VE30 Local-language claim monitoring and checkworthiness triage M3 V-22
VE31 Fact-check interoperability M3 V-23
VE32 Numbers rendering guard M3 V-24
VE33 Channel impartiality ledger (24/7 stream) M3 V-30
VE34 Synthetic anchor governance M3 V-31
VE35 Loop-aware on-air corrections M3 V-32
VE36 Blind cross-spectrum panel testing of Veritas output M3 V-38
VE37 Balance metrics dashboard M3 V-39
VE38 Forecast and prediction-market context module M4 V-17
VE39 Reader questions, dissent and feedback loop M4 V-36
VE40 Certification readiness (IFCN, JTI) M4 V-40

6.8 Lilith, characters and the creative studio [CR]#

Research: Eve document 17508fb3-47c1-453a-a3f7-8bb1cf1fe950.

What the best do.

  • AI raises the floor but homogenises. In a 2024 study of 293 writers, AI ideas raised novelty 8.1% and usefulness 9.0%, most for the weakest writers. But AI-assisted stories were measurably more similar to each other, and a 2026 Trends in Cognitive Sciences review names the "homogenizing effect". A catalogue written mostly by one family of cheap models is directly exposed. Diversity can be recovered (prompting strategy changes idea spread) and measured (the Vendi Score needs no reference set).
  • Professional character design is a staged process: brief → research → silhouettes → shape language → colour keys → expressions → turnarounds → model sheet → style guide. Persona platforms (Character.AI, Convai) show the minimum a behaviour spec must hold. Visual consistency still comes from the model sheet: 2026 image models take a handful of character references (Runway Gen-4 up to three; Google's Gemini image models up to 14 in total, four or five of them character images), but long-video consistency remains open, so recurring characters should be rigged VRM 1.0 or Live2D assets rather than regenerated. Ready Player Me shut its public services on 31 January 2026.
  • Personas drift, and size alone does not fix it. One benchmark found significant instruction drift within eight rounds (on older models); the number is not a general expiry time, so each persona must be tested on the model, retrieval set-up and conversation length it actually runs at.
  • LLM judges agree with humans on assistant answers but not with experts on creative quality; human-calibrated judges and panels are required before automated scores approve anything.
  • Voices clone from seconds of audio with open models, so voices designed from description, consent records and a resemblance screen are the safe default. Ghanaian-language speech remains thin: none of the open TTS models checked lists Twi, Ga or Ewe, and GhanaNLP's own warning concerns LLM translation quality; one commercial dubbing service (ElevenLabs) now lists Akan.
  • The law shapes persona design. EU AI Act Article 50 has applied since 2 August 2026, with a grace period to 2 December 2026 for the Article 50(2) marking duties of systems already on the market before then (its code of practice was finalised in June 2026). New York requires AI companions to disclose that they are not human at the start of a session (no more than once a day is required) and at least every three hours in continuing conversations; California's SB 243 adds break reminders for known minors. Arete and the V1.3 agent are the most companion-like surfaces. In a 981-person, four-week trial, heavier voluntary use was associated with loneliness and dependence (an association, not proof of harm).
  • One character across many surfaces needs a system: a voice-and-tone guide (Mailchimp: the voice is constant, the tone varies), one owner, and golden exemplars. Duolingo carried its character Lily into live AI video calls, and its 2025 "AI-first" backlash shows audiences punish being misled about AI.
  • Music and model terms move under you. Udio, Suno and the labels changed terms four times in a year, so every asset records its licence at the moment it was generated.

What the studio must add. Briefs as objects, and an ideation board with divergence operators and novelty scoring. A character forge that edits the existing persona registry: bible, behaviour and safety spec, model sheet and consistency pack, on-model checker, likeness and IP screen, voice design with consent, release and versioning, and a drift suite. Also a generation-ready style guide with conformance scoring, frame-accurate review and finishing QC, accessibility authoring, a human preference lab, packaging honesty checks, and Lilith's canon with its voice-and-tone guide and change control. Items the V1.4 plan builds for channels only are built once here, and V1.4 reuses them (see each task's "Builds on").

Key Task Milestone Feature
CR01 Ownership map and amendments for the V1.4 overlaps M0 —
CR02 Creative brief object M1 CRE-01
CR03 Persona behaviour and safety spec M1 CRE-15
CR04 Shared ideation board M1 CRE-03
CR05 Divergence operator library M1 CRE-04
CR06 Novelty and diversity scoring M1 CRE-05
CR07 Character bible editor M1 CRE-10
CR08 Model sheet and consistency pack builder M1 CRE-12
CR09 Generation-ready style guide and brand kit M1 CRE-21
CR10 On-model checker M1 CRE-13
CR11 Persona release and versioning M1 CRE-16
CR12 Voice consent and likeness guard M1 CRE-18
CR13 Frame-accurate asset review M1 CRE-26
CR14 Finishing and delivery QC M1 CRE-27
CR15 Accessibility authoring M1 CRE-28
CR16 Human preference lab M1 CRE-30
CR17 Persona consistency and cross-surface drift suite M1 CRE-32
CR18 Lilith canon, voice-and-tone guide and change control M1 CRE-36
CR19 Likeness and IP collision screen M1 CRE-40
CR20 Define Lilith and the V1.0 room guides through the character forge M1 —
CR21 Versioned generation capability contracts M1 CRE-41
CR22 Voice design and bake-off M3 CRE-17
CR23 Audience research and jobs-to-be-done library M3 CRE-02
CR24 Notes and greenlight record (Braintrust mode) M3 CRE-06
CR25 Series bible and beat board M3 CRE-08
CR26 Visual development board M3 CRE-11
CR27 Avatar asset pipeline (VRM 1.0 and Live2D) M3 CRE-14
CR28 Performance direction and take selection M3 CRE-20
CR29 Style conformance scorer M3 CRE-22
CR30 Sonic identity and licensed music generation M3 CRE-23
CR31 Script studio with voiced table read M3 CRE-24
CR32 Storyboard and animatic tool M3 CRE-25
CR33 Catalogue diversity monitor and exploration quota M3 CRE-33
CR34 Packaging honesty check and testing M3 CRE-34
CR35 Analytics-to-format feedback loop M3 CRE-35
CR36 Lilith golden exemplars and copy linter M3 CRE-37
CR37 Controlled divergence experiments M3 CRE-43
CR38 Revision impact and selective regeneration M3 CRE-44

7. Sequencing and milestones#

Milestone Scope Release it serves
M0 Foundations and urgent fixes The architecture ADR. The shared engines everything depends on: knowledge models, canvases, arcs, evidence packs, quotation lock, rights, authority, disclosure, claims, expert review, evaluation, audio, crisis resources. Workbench hosts for Nyx, Arete and Nisaba. Five urgent fixes to the live product: the Tara demo catalogue, intense breathwork, the Seven Habits trademark, non-commercial corpus licences, and territory-aware serving of the KJV for UK members, and a guard that keeps non-commercial corpus material off member surfaces until the licence registry ships (the Nyx data-feed format change is also scheduled here) V1.0 (live now)
M1 V1.0 rooms at state of the art P0 capabilities for Tara, Nyx, Arete and Nisaba; consented outcome and harm measurement and harm intake, both part of every launch gate; and each room's launch catalogue produced through its workbench V1.0
M2 V1.2 rooms ADRs placing the new work on the Veritas and Metis ledger phases, their P0 capabilities, and launch sets V1.2
M3 Depth, personal agent and channel feeds P1 capabilities in every room. Calendars, coverage, local languages, personalisation rules for the V1.3 agent, the operational-evidence programme (real pilots, panels and cohorts), and the feed contract for the V1.4 channels V1.2 to V1.4
M4 Horizon P2 capabilities after V1.4

Dependencies enforce the order: no room task starts before the shared engine it needs. Each room's launch task depends on that room's release gate, which depends on every shared engine in M0–M1 (including consented outcome and harm measurement), the room's P0 controls, harm intake and audience enforcement. The gate refuses publication when any of them is missing or unavailable. Veritas tasks depend on the Veritas ADR (VE01) and on the Phase V ledger tasks they extend, and Metis tasks on the Metis ADR (ME01). Real-world evidence (panels, pilots, cohorts) is its own M3 programme (CO29), so software completion never depends on recruitment or elapsed time. The V1.4 channels depend on this family for supply: the room content adapters in lilith-v1-4 should read workbench bundles through CO28 once it exists (see §8).

8. Relationship to other plans#

  • v1-domain-workbenches
    • The Metis (Phase M) and Veritas (Phase V) ledgers stay as written. ME01 and VE01 are ADRs that place this family's work on them, and each ME and VE task names the ledger task it extends.
    • The shared kit (Phase S) is the shell every new workbench mounts on.
    • Isis, Euterpe, Aja, Bellona and Yemaya remain the production engines this family calls, not rebuilds.
  • tara-content-workbench (113/115 closed): the Tara pillar extends that workbench. Its spark → concept → script → grounding → bundle pipeline is kept; the new work adds structure, safety and craft around it.
  • lilith-v1-3 (personal agent): the agent needs structured content to adapt. It reads the personalisation rules (CO27) and the practice, technique and arc models this family creates. No V1.3 task is changed.
  • lilith-v1-4 (channels), three connections:
    • The channel crafts (writers' room, universe bible, casting, voice, art direction, presence, music) stay in V1.4. Where the creative pillar builds a shared object that V1.4 also plans (character sheets, pronunciation lexicons, style kit, formats, rights and provenance), CR01 produces a task-by-task ownership map and proposed amendments for Elikem's decision. No overlapping CR task starts before its row is decided.
    • The room content adapters (lilith-v1-4:TA02, lilith-v1-4:NX02, lilith-v1-4:AR02, lilith-v1-4:NI03, lilith-v1-4:ME03, lilith-v1-4:VE03) should consume workbench bundles through CO28. For example, the V1.4 Tara adapter currently plans to generate through svc-meditation-generation. This document recommends amending those adapter tasks once CO28 is approved; it does not edit them.
    • The Metis research found England's January 2026 DfE generative-AI product safety standards discourage named, avatar-bearing AI personas for school-age learners (Decision ME-2), which bears on lilith-v1-4:ME02 (faculty casting).

9. Decisions only Elikem can make#

The research passes raised 70+ questions. They reduce to the decisions below, each with a recommendation.

Settled 2026-09-29: Elikem accepted every recommendation below ("go ahead with your recommendations"). The recommendations are now decisions, and tasks follow them. Acts that need a person remain Elikem's: contracting reviewers, the clinician and voice actors, partnerships, counsel and spend. Tasks that reach one prepare the request and park with --reason human.

Across all rooms#

# Decision Recommendation
X-1 Fund paid external reviewers per room: meditation teachers and lineage holders, a clinician, astronomers and cultural knowledge-holders, coaches, textual scholars per tradition, subject teachers, and a viewpoint-balanced Ghana and UK news panel Yes. Start with a small retained panel per room at M0; expert review is what separates leading content from generated content (CO11)
X-2 Is Lilith "commercial" for licence purposes? Yes, treat it as commercial. Record each asset's actual licence; suspend non-commercial sources (CBETA, ctext, NC Sefaria versions) until licensed, after tracing what is actually served (CO07, NI02)
X-3 Minors Keep consumer rooms 18+ through V1.2, matching V1.3's adult cohorts. Decide Metis for under-18s separately (ME-1)
X-4 Positioning Wellness-only in every market: no treatment claims for any condition. Each offering is classified by intended purpose and functionality under MHRA and EU rules, not by wording alone (CO10)
X-5 Faith-integrated variants for Ghana (Christian, Muslim) in Tara, Arete and Nisaba Yes, as opt-in variants reviewed by named faith advisors
X-6 African traditional knowledge (Ifá/Afa, Akan and Yoruba sky lore, Kemetic claims) Explainers only, with named, paid community partners. No instruction, no divination, never generated verses (CO08)
X-7 Outcome and harm check-ins (health-adjacent data) Yes: opt-in and minimal, under the sensitive-data governance task (lawful basis and special-category condition, retention and deletion, Ghana Data Protection Commission registration) before first collection (CO21, CO15)
X-8 Ghanaian languages English first; Twi next, then Ewe, Ga, Hausa and Dagbani with native reviewers, in step with the V1.4 language desks (CO26)
X-9 Named advisers Contract a clinician adviser (for practice eligibility and adverse-effect protocols) and name a compliance owner (for claims classification and data protection) before M1

Tara#

# Decision Recommendation
TA-1 Intense breathwork (hyperventilation, retention) Exclude from Ghana and from all continuous streams. In the UK, only behind screening once the breath-protocol registry and contraindication engine ship
TA-2 Course model shown to members A library plus guided phase-structured paths now; adaptive progression later, through the V1.3 agent
TA-3 Stage maps (e.g. The Mind Illuminated stages, insight stages) Authoring-side only; never shown as goals to members
TA-4 "Looking for the self" and other deconstructive practice for beginners Allowed as short, glimpse-style exercises after an attention phase, within a dose and safety envelope
TA-5 MBSR/MSC branding "Inspired by" with attribution; no trademarked programme names unless licensed
TA-6 May fictional mentors say they draw on a named lineage? They may cite traditions as sources, but may not claim titles, transmission or authorisation

Arete#

# Decision Recommendation
AR-1 libs/arete/seven-habits (FranklinCovey trademark, Covey quotations) Rewrite as an original, attributed effectiveness framework without trademark or quotations, unless you want to license it (AR02)
AR-2 Clinical screeners (PHQ, GAD) for triage No. Use disclosure detection and signposting only. This is one input to the intended-purpose classification in CO10, not a guarantee of staying outside medical-device scope
AR-3 Leaderboards Off by default; opt-in co-working accountability only
AR-4 Crisis services in Ghana Seek a formal partnership with the Mental Health Authority or an NGO helpline so the registry points at a verified, staffed service (CO14)

Nisaba#

# Decision Recommendation
NI-1 KJV for UK members Serve a public-domain translation (e.g. the World English Bible) to UK members now; revisit a Cambridge University Press licence later
NI-2 AI and sacred text No AI-drafted canonical translations. AI may assist commentary and explainers, labelled and reviewed. No synthetic recitation of Qur'anic or Vedic originals
NI-3 Sending corrections upstream (SuttaCentral, Sefaria, Quranic Arabic Corpus) Yes, where licences allow, recording any share-alike duties

Nyx#

# Decision Recommendation
NX-1 Astrology Cultural history only: no horoscopes, predictions or personality claims
NX-2 Render targets Device location with a chosen-place fallback. First-class QA locations: Accra, Tamale, London, Edinburgh
NX-3 Rebroadcast of third-party launch webcasts Computed visualisations by default; negotiate rights case by case

Veritas#

# Decision Recommendation
VE-1 Outlet ratings: build or license Build our own published method; show licensed third-party ratings only as crosswalk inputs
VE-2 Party-alignment labels for Ghanaian outlets at launch Show ownership and reliability at launch; alignment labels only after the panel method has run one full cycle with right of reply
VE-3 "False" verdicts on named politicians Top review tier with legal sign-off; at first, cite GhanaFact or Dubawa verdicts rather than issuing our own
VE-4 Which review tiers may auto-publish with an AI label None at launch. Every Veritas item gets substantive human review and a named editor of record, which also meets the Article 50(4) exemption
VE-5 Forecasts Calibrated forecast aggregators only, clearly labelled; no betting odds
VE-6 Ofcom and EU applicability of the Newsroom channel Build to Ofcom Section Five regardless; obtain counsel before launch (joins V1.4 Decision 7)

Metis#

# Decision Recommendation
ME-1 Under-18 learners and sales to schools in England Adults first, enforced by the audience policy task (CO17). If school-age learners or schools come into scope, map the DfE generative-AI product safety standards by educational context and product function
ME-2 Faculty as named avatar characters For school-age content, use function-labelled presenters in time-limited roleplay; named, labelled AI characters only for adult Academy programming
ME-3 First syllabuses JHS mathematics (NaCCA/BECE) and one GCSE science topic
ME-4 Mastery gating Soft by default, with honest readiness signals; hard gates only where prerequisites are safety-relevant

Lilith, characters and the creative studio#

# Decision Recommendation
CR-1 Who owns Lilith's character? One named human editor with final say (you, or a delegate you name), with room leads contributing through change control
CR-2 How close is Lilith to members: host and guide, or friend and confidant? Host and guide. That keeps companion-law duties proportionate and avoids dependence by design
CR-3 One house visual style, or distinct styles per room? One Lilith frame with distinct room styles inside it, all generated from one style kit
CR-4 Recurring personas in 2D (Live2D) or 3D (VRM)? VRM 1.0 for recurring characters, which aligns with the V1.4 stylised tier-1 decision. Pre-rendered 3D only for set pieces
CR-5 If synthetic Ghanaian-accented voices fail the bake-off Commission voice actors under explicit replica consent, rather than ship a voice that mispronounces names
CR-6 Paid human raters in Ghana and the UK Yes, a small standing panel per market. No automated judge approves content before it is calibrated against them
CR-7 Music Licensed generation plus a commissioned sonic identity per room
CR-8 Should V1.4 use the shared studio objects (characters, lexicon, style kit, formats)? Yes. Build them once here, and have the V1.4 craft tasks consume them
CR-9 C2PA signer and the EU code of practice Oshun Synthetics LTD signs, and joins the EU code of practice on AI-content marking

10. Risks#

Risk Where it bites Mitigation in this family
Fabrication at scale. Cheap models running 24/7 multiply known failure rates: misquoted scripture (15–60%), AI news answers with significant issues (45%), invented credentials Every room, the channels Structural, not prompt-level: CO05 evidence packs, CO06 quotation lock, room re-verification of derived outputs, CO12 evaluation, CO11 expert review
Physical and psychological harm from practice or coaching content: hyperventilation contraindications, meditation adverse effects, crisis moments, eating and exercise content Tara, Arete TA02 now; the contraindication engine, dose envelopes, trauma-sensitive variants, scope-of-practice gate, CO14 verified crisis registry
Regulatory drift into medical-device territory through claims or screeners Tara, Arete CO10 claims engine across every surface; Decisions X-4 and AR-2
Licence and IP exposure to confirm: non-commercial connectors, the KJV in the UK, Seven Habits branding and quotations, CC BY-ND lore (adapted versions may not be shared), unlicensed third-party ratings. Exposure is present in code; whether each is a breach depends on what is actually served and on use-specific review Nisaba, Arete, Nyx, Veritas NI02 and AR02 in M0; CO07 render-time enforcement
Misrepresenting living traditions: flattening, invented lore, restricted knowledge, false lineage claims by fictional personas Tara, Nisaba, Nyx CO08 authority registry; tradition reviewers; the persona honesty guard; anti-flattening checks
False balance and false certainty Veritas; any room making contested claims The epistemic classifier; due-weight ledger; steelman symmetry audit; blind cross-spectrum panels
Politically explosive outlet labels in Ghana Veritas Published method, balanced panels, right of reply, dated evidence; Decision VE-2
Engagement over outcomes: streak shaming, seductive "wonder" content, sycophantic coaches Arete, Metis, Tara Anti-dark-pattern rules; seductive-detail objective lock; red-team harness; CO21 outcome measurement
Anthropomorphised AI teachers for children Metis, the V1.4 Academy channel Decision ME-2; the manipulation and anthropomorphism copy lint
Reviewer capacity becomes the bottleneck once agents produce drafts quickly All rooms Risk-ranked review (review the riskiest first); sign-off tiers by risk; Decision X-1
Scope size: this is one of the larger families on the board Delivery Dependencies and milestones gate the order; M0 and M1 serve the live V1.0 rooms first; P2 items wait in M4

11. Research documents#

Area Eve document Revision Features
Tara 13d4487f-57de-4523-81d6-e919e7b2afba 1 50 (24 P0)
Nyx 788820e4-8a29-450b-b31c-86104a08c7a4 2 36 (17 P0), incl. 4 from the web-search supplement
Arete c479a954-f5e0-43ff-91f6-02520dbf7348 2 39 (24 P0), incl. 4 from the web-search supplement
Nisaba 8563ce32-96b4-45d3-af99-5115633fb753 1 34 (18 P0)
Metis b325c39f-6ba1-4b49-98ba-c2d062dfde80 2 61 (29 P0), incl. 8 from the web-search supplement
Veritas c15926ca-fd59-414b-a3bf-614c4551a235 1 40 (26 P0)
Lilith, characters and the creative studio 17508fb3-47c1-453a-a3f7-8bb1cf1fe950 2 44 (25 P0), incl. 4 from the web-search supplement
Planned tasks and acceptance matrix (evidence) b3643474-b87f-4fdb-84ee-2e6e2dadcc79 2 every task body, dependency and folded acceptance test

Features whose acceptance tests belong to a shared engine are folded into that engine's core task (listed under "Room acceptance folded into this task" in its body), so every research test is carried by exactly one task.