Historical task source. The Eve task board holds current task state.
Source: derived from V8/V8_features.md, V8/V8_ARCHITECTURE.md,
V8/V8_GAP_ANALYSIS.md, V8/V8_DEPENDENCIES.md, and V8/V8_SOTA_RESEARCH.md,
cross-referenced against V5 (V5/V5_features.md, V5/ue), the Yemaya agent
studio (libs/yemaya/agents), Isis (libs/isis), Hathor
(libs/hathor/llm-npc), V6 Ori, and V7 Sekhmet, as of 2026-06-04.
Purpose#
This file is the source of truth for V8 completion state. V8 — Ariadne — turns V5's authored detective game into a self-authoring universe: an integrated system of SOTA AI agents that generates an endless supply of fair, solvable, fully-realized cases and ships them into the existing V5 game.
Each unchecked [ ] is a coding-agent-actionable task. A task is complete only
when its implementation, tests, and required wiring all exist — not when a
stub is present. The Oshun-repo CLAUDE.md "Zero Tolerance for Stubs" rule
applies verbatim: no Math.random() faking computed results, no hardcoded
returns, no "would require…" placeholders, no simulated runtimes where real ones
belong. The checkbox is the sole source of truth; ✅/done notes in prose
mean nothing.
Cross-Reference Convention#
- Bare
§N→ a section/phase in this file. arch§"<anchor>"→V8_ARCHITECTURE.md.features§"<anchor>"→V8_features.md.gap§N→V8_GAP_ANALYSIS.md.sota§N→V8_SOTA_RESEARCH.md.deps§N→V8_DEPENDENCIES.md.v5src:<path>→ a file underV5/ue/Source.ylib:<name>→libs/yemaya/<name>.
Phase / cut-line convention#
- [P1] GA launch-blocking. [P2] Year-1 post-launch. [P3] Year-2 stretch.
- Default for any
##section is [P1] unless tagged. The "Definition of done" (features§"Definition of done") must be fully[x]before V8 GA.
The eight ReleaseDecision gates (no case reaches a player until ALL pass)#
A minted case must pass, in order: (G1) Minos formal uniqueness, (G2)
Minos deductive-completeness, (G3) Knox/Van-Dine fair-play, (G4) Theseus
in-game solve to Brilliant-eligible path, (G5) judge-panel quality ≥
threshold, (G6) Sekhmet safety clear, (G7) canon-consistency, and
(G8) preregistered human-quality launch evidence. Each gate is a task below;
decideRelease enforces all eight. The separate §12 platform suite retains its
seven content-gate IDs; it is an additional harness, not a substitute numbering
for this decision.
Phase 0 — Foundations, contracts, and scaffold-promotion [P1]#
Goal: stand up the V8 packages and lock the data contracts so all later phases can proceed in parallel. Close
gap§1C(scaffold→live) before depending on it.
- 0.1 Create
ylib:case-contracts(project.json/package.json/tsconfig.json, path-map intsconfig.base.json). DefineCaseSpec,CaseGroundTruth,GroundTruthFact,ConstraintClause,TimelineEvent,Alibi, asset-manifest types (arch§3.1–3.2). Import, do not redefine, the existingMysterySession/Clue/Solutionfrommystery-fairness-validator.ts(arch§3.3). - 0.2 In
case-contracts, define the V5 compile-target schema as TS types mirroringv5src:V5MindPalace/Public/V5MindPalaceTypes.h(FV5MindPalaceEvidenceNode,FV5MindPalaceDeductionEdge,FV5MindPalaceAccusationOutcome) andv5src:V5DetectiveColdCases/Public/V5DetectiveColdCasesTypes.h(FV5ColdCasePackDefinition). Generate them from the headers (parse) so they cannot drift; add a docs-drift CI check. - 0.3 Extend
cold_cases_manifest.jsonschema to v2 (addgeneratedBy,seed,provenanceC2PA,assetManifestId,verifierReport) backward-compatibly; updateV5/tools/cold-cases/validate-cold-cases.pyto accept v1 and v2. - 0.4 Scaffold apps:
apps/v8/loom-service(Nest BFF),apps/v8/daedalus-compiler(CLI),apps/v8/minos-asp-sidecar(container). Wire Nx targets (build/test/lint); verify withnpx tsc --noEmit+npx vitest runper-lib (Nx-bypass rule,CLAUDE.md). - 0.5 Promote scaffold #1: replace the
// stub:legitimatedeterministic responses inapps/lilith/svc-ai/src/agent-coordination.ts:24with real LLM-backed coordination, or formally retire that path in favor of yemayaWorkflowExecutor. Document which orchestrator V8 standardizes on. (gap§1C) — formally retired w.r.t. V8; seeV8_ORCHESTRATION_DECISIONS.md. - 0.6 Promote scaffold #2: audit
libs/neith/integration-hathortemplate generators; either wire them to LLM calls or mark them explicitly as deterministic fallbacks that V8 supersedes. No silent template-as-AI. - 0.7 Promote scaffold #3: convert yemaya
*-integrationadapters (isis/hathor/sophia) from type-layer wrappers to live calls behind a tested interface (gap§1C). Cover each with an integration test that hits a real (or recorded-fixture) backend. - 0.8 Establish the V8 event/journal spine: every pipeline stage
emits a typed event to Redis + a durable journal row (Postgres) keyed by
(caseId, seed, stage)for determinism/replay (arch§5). - 0.9 Define the cost budget envelope per case (tokens + asset-job
ceiling) in
libs/yemaya/budget-management; exposereserve()/spend()/remaining()to all stages.
Phase 1 — Palimpsest: the continuity knowledge graph [P1]#
Closes the "canon/continuity" gap (
gap§2). Implementssota§3(SCORE + graph-RAG).
- 1.1 Create
ylib:canon-graph. Model the graph:Character,Location,Faction,TimelineEvent,PriorCaseOutcome,OpenThread, typed relations. Persist in Postgres + pgvector; optional Neo4j overlay (existinggraphprofile). - 1.2 Seed the graph from V5's authored world bible + the 25 launch
cases (parse
cold_cases_manifest.json+ Vice Squad data) so generated cases reuse the real city, NPCs, and factions (features§"2"). - 1.3 Implement SCORE-style retrieval: dynamic state tracking
(symbolic per-entity state), hierarchical prior-case summaries, hybrid
TF-IDF + embedding retrieval. Return a bounded
CanonContextfor anyCaseSpec. (sota§3) - 1.4 Absorb/replace
libs/yemaya/canon-enforcement: acheckConsistency(draft, canonSnapshot)API that flags contradictions (geography, character facts, timeline). This is G7. - 1.5 Implement canon write-back (
features§"2"): a verified+played case's outcome (who was accused/convicted, threads opened/closed) commits to the graph, snapshot-versioned. Co-op write-back is leader-arbitrated (arch§5). - 1.6 Snapshot-hash a canon slice so
(seed, model-versions, canon-hash)is the reproducibility key (arch§5; ties to 0.8). - 1.7 [P2] Cross-cell canon: let the Mind Palace's cross-era deductions
(
bCrossEra) draw on canon from multiple V5 cells.
Phase 2 — Clew: the generative case engine (the core missing piece) [P1]#
Closes Gap A (
gap§1A). Implements the solve-first, backward-derivation method (sota§1). This is the heart of V8.
- 2.1 Create
ylib:case-engine. Implement theTruth-Weaveragent (extendsCreativeAgent,base-creative.ts:140): givenCaseSpec+CanonContext, propose victim/culprit/means/motive/opportunity; commit asGroundTruthFact[]. - 2.2 Timeline construction + conflict detection: generate every
character's movements across the case window; detect/repair physical
impossibilities (no bilocation; culprit at scene in window). Output
CaseGroundTruth.timeline. - 2.3 Constraint encoding: emit
ConstraintClause[]linking facts→suspects (motive/means/opportunity network) such that the culprit variable is the unknown (arch§3.2). This is the input Minos will solve. - 2.4 Implement the
Clue-Smithagent: for each culprit-distinguishing fact, deriveClue(s) withtype/visibility/importance/pointsTo[]/eliminates[]set against solution variables — the clue inventory tied to solution vars (sota§1). Populate the existingMysterySessionIR. - 2.5 Red-herring generator: produce misleading-but-fair clues;
each red herring MUST have a derivable eliminating clue (set
redHerringFairness). Reject any red herring that cannot be fairly refuted. - 2.6 Alibi generation: innocents get verifiable alibis for the murder window; the culprit's alibi carries exactly one discoverable flaw.
- 2.7 Revelation staging: assign
Clue.introducedAt/visibilityfor fair gradual disclosure (no information withheld from the player that the solution needs). - 2.8 Build the
Solution.deductionPath(Ariadne's thread / the clew) — the ordered chain of deductions from first clue to accusation. - 2.9 Emit a complete, schema-valid
MysterySession+ the hiddenCaseGroundTruth. Unit tests assert structural validity and that every clue maps to a ground-truth fact (no orphan/ungrounded clues — anti-hallucination). - 2.10 [P2] Sub-genre profiles: locked-room, cold-case, conspiracy,
courtroom — parameterize Clew so
CaseSpec.themeTagssteer structure.
Phase 3 — Minos: solvability & fair-play verifier (gate G1–G3) [P1]#
Closes the "no formal solvability proof" gap. Implements
sota§1CSP/ASP uniqueness + the existing fairness validators as a gate.
- 3.1 Create
ylib:case-verifier+apps/v8/minos-asp-sidecar(clingo/ASP,deps§3). CompileCaseGroundTruth.constraints+ player-available clues into an ASP program. - 3.2 G1 — formal uniqueness: prove the culprit is the only model
satisfying player-available clues. Fail if 0 models (unsolvable) or ≥2
(ambiguous). Return the offending model(s) for repair. (
sota§1, AAAI-2007.) - 3.3 G2 — deductive completeness: verify each
deductionPathstep is supported only by clues introduced earlier; no external knowledge, no forward refs. - 3.4 G3 — fair-play: call the existing
analyzeFairness(sessionId)(mystery-fairness-validator.ts:1825, Knox/Van-Dine) + the investigation-balance analyzer; requirerating ∈ {exemplary,fair}, zerobroken/unfairviolations. - 3.5 Repair loop (critique-revise,
sota§2): on any gate fail, return the specific violations to Clew (§2) to regenerate ONLY the offending clues/herrings; bounded retries (default 3), then escalate to HITL (libs/yemaya/agents/src/hitl). - 3.6 Difficulty grading: from the constraint graph (suspect count, clue depth, herring density) compute an a-priori difficulty estimate; reconcile later with Theseus's empirical solve-rate (§9).
- 3.7 Emit a signed
VerifierReport(gates, models, fairness score) into the manifest (0.3) and journal (0.8).
Phase 4 — Anansesɛm: the writers' room (narrative realization) [P1]#
Realizes the verified skeleton into the experienced surface. Uses yemaya orchestration + creative agents (
gap§1Cpromoted in §0.5–0.7).
- 4.1 Create
ylib:case-writers-room. Build the realization DAG withWorkflowBuilder/StateGraph:Showrunner → {StoryDirector, CharacterWriter, DialogueWriter, Cinematographer, SoundDesigner} → Integrator. - 4.2 Grounding: every agent prompt is augmented with
CanonContext(graph-RAG, §1) so realized prose/dialogue cannot contradict canon (sota§3). - 4.3 Interrogation realization: generate dialogue trees that align to V5's 12 facial-tell beats and Truth/Doubt/Lie mechanic; each catchable lie maps to a specific clue (consistency with §2.5/§2.6).
- 4.4 Generate scene descriptions, case briefings, partner banter, and era/cell voice (Vice-Squad 1947, etc.) consistent with the V5 cell.
- 4.5 Localization-ready output: structured text with stable string ids so per-locale generation (§11) can fan out.
- 4.6 Tests: realized text references only entities present in ground
truth + canon (no invented suspects/locations); golden-fixture diff
against an authored V5 case for tone parity.
<!-- Grounding test DONE (
assertGrounded, case-writers-room). The tone-parity diff is externally blocked: V5 ships a manifest skeleton with no authored case prose/VO text as data to diff against. --> Board tag 2026-09-18: the tone-parity half waits on authored V5 case prose as data — V5-002 ofV1_V9_AUTONOMOUS_CONTENT_SOTA_GAP_CLOSURE_TODOS_2026-07-15.md("Convert all 25 intended cases into structured canon") andV5/V5_TODOS.md's Vice Squad cases; three trackers wait on that one corpus.blocked:upstream
Phase 5 — Loom: asset realization fabric (Isis orchestration) [P1]#
Binds generated cases to generated multimedia via the existing Isis client.
- 5.1 Create
ylib:case-assets. From a realized case, compute an asset plan (which env/props/portraits/documents/music/VO/video are needed) keyed to clues, suspects, and scenes. - 5.2 Retrieve-then-generate (
sota§2): query an asset library by semantic tags; reuse matches; only mint new assets for gaps (cost control, ties to 0.9). - 5.3 Orchestrate Isis jobs via
IsisClient.submitBatch/submitAndWait(libs/isis/client): crime-scene env + props (Hunyuan3D/Meshy), suspect portraits + MetaHuman params (Stability/Flux), document/photo evidence (Stability), case music + stings (Suno), full VO for every line (ElevenLabs), optional flashback/CCTV (LTX). - 5.4 Asset binding: map each
JobOutput.files[].urlto a stable asset id +FSoftObjectPathplaceholder for the compiler (§7); produce anAssetManifest. - 5.5 Enforce per-case asset budget (0.9); on over-budget, degrade gracefully (reuse library art, fewer bespoke props) rather than fail (arch§5).
- 5.6 VO coherence: voice-cast each suspect consistently (one ElevenLabs voice id per character across cases) and store the mapping in canon (§1).
- 5.7 [P2] MetaHuman assembly automation: drive parametric MetaHuman creation from generated portrait + persona for non-principal suspects.
Phase 6 — Ori-Detective: living suspects + runtime interrogation [P1]#
Promotes suspects from dialogue trees to generative-agent minds (
sota§2), bridging V6 Ori + Hathor + ACE/Inworld.
- 6.1 Create
ylib:case-suspects. Instantiate a V6 Ori per suspect/witness, seeded with their ground-truth knowledge, secrets, and the lies they may fairly tell. - 6.2 Implement memory-stream / reflection / planning over the case so
interrogation stays consistent across multiple passes (V5 allows
re-interrogation) (
sota§2). - 6.3 Wire Hathor
DialogueGenerator.generateResponse()(libs/hathor/llm-npc) as the dialogue backend; inject world + case state via the (promoted) hathor-integration adapters. - 6.4 Fair-lying guard: a runtime check that an Ori never asserts a lie that no available clue can refute, and cracks when the correct evidence is presented (mirrors §2.5/§3.4). This is the runtime half of fair-play.
- 6.5 ACE/Inworld runtime (
deps§3,sota§4): broker live interrogation sessions; stream audio + visemes to MetaHuman (Audio2Face/NeuroSync). Fail-closed to the offline-generated dialogue tree (§4.3) when creds absent. - 6.6 Latency budget: meet an interactive interrogation turn-latency target; prefetch/stream; cache per-suspect system context.
Phase 7 — Daedalus: the case compiler (the bridge into V5) [P1]#
Closes Gap B (
gap§1B). Turns a verified case into data the shipping game loads.
- 7.1 Create
apps/v8/daedalus-compiler. Map a verifiedMysterySession+AssetManifest→FV5MindPalaceEvidenceNode[](one perClue; bindImageto the generated document/photo asset; setbFalseLeadfor herrings). - 7.2 Map
Solution.deductionPath→FV5MindPalaceDeductionEdge[](FromNodeId/ToNodeId,DeductionText,Confidence,Kind=Generated,bCrossEra,UnlockFlag). - 7.3 Map culprit + plausible wrong accusations →
FV5MindPalaceAccusationOutcome[](EvidenceChain,RequiredEdgeIds,Rating,OutcomeBranch). - 7.4 Emit a
FV5ColdCasePackDefinitionheader + a v2cold_cases_manifest.jsonpack (evidenceNodeIds[], deductionPairIds[], VO counts, provenance, seed). Runvalidate-cold-cases.pyas a build gate — a case that fails the schema does not compile. - 7.5 Package generated assets into a loadable bundle (cook/pak or
runtime-download bundle) and resolve every
FSoftObjectPathso the case loads with zero missing refs. - 7.6 Round-trip test: compile a case, load it in V5's cold-cases plugin path (headless/automation), assert the Mind Palace graph builds and an accusation resolves.
- 7.7 Author-tool parity test: a Daedalus-compiled pack is byte-compatible with a pack produced by V5's human Case Author tool (same schema, same load path).
Phase 8 — V8 UE5 client plugin [P1]#
The thin in-engine layer. Real C++, depending only on existing V5 modules.
- 8.1 Create
V8/ue/Plugins/V8_Ariadne_CaseClient(a UE5 GameFeature/plugin). Depend onV5MindPalace,V5DetectiveColdCases. No provider keys in the client (sota§4). - 8.2 Implement async case fetch (UnrealGenAISupport delegate pattern,
sota§4): request a pack from the loom-service relay (§9), download the asset bundle, register with the cold-cases plugin, surface on the Bureau Case Board UI. - 8.3 Implement the runtime interrogation client: open an ACE/Inworld session for a suspect, drive MetaHuman audio/visemes, send player utterances, receive Ori responses (§6).
- 8.4 Graceful degradation: if the relay or runtime NPC backend is unavailable, fall back to pre-fetched offline cases + authored dialogue trees.
- 8.5 Build/automation verification on the on-box UE (per
CLAUDE.mdUE rules: build + run automation asueagent, check disks first).
Phase 9 — Theseus: automated playtester + eval gates (G4–G5) [P1]#
No case ships unsolved. Implements
sota§1solver agents +sota§5LLM-as-judge.
-
9.1 Create
ylib:case-eval. Build the Theseus solver agent: an autonomous LLM detective that plays the compiled case against V5's actual data (evidence nodes, deduction edges), collects clues, attempts deductions, and accuses. -
9.2 G4 — in-game solvability: assert Theseus reaches the intended culprit via a
Brilliant-eligible path using only player-available data. Fail → repair (§3.5)/HITL. This catches compiler/asset losses that the IR-level Minos proof (§3) cannot. -
9.3 G5 — quality judge panel: LLM-as-judge 8-dimension rubric (coherence, surprise, fairness-feel, pacing, character, prose, voice-fit, difficulty-accuracy), PCA-aggregated (
sota§5); threshold-gate. Report per-dimension scores. -
9.4 Difficulty calibration: run Theseus at varied "skill" to estimate empirical solve-rate; reconcile with Minos's a-priori grade (§3.6); write the calibrated difficulty to the manifest and Oracle's model (§10).
-
9.5 Regression corpus: use V5's 25 authored cases as golden fixtures — Theseus must solve them and the judge must score them highly, anchoring the gates. <!-- The harness is DONE (
runRegressionCorpus, case-eval, green over a generated golden corpus). Externally blocked on the SPECIFIC fixtures: V5's authored cases have no full FV5 deduction graphs as data (manifest skeleton, zero .uasset) for Theseus to play. --> Board tag 2026-09-18: same missing corpus as 4.6 — V5-002 of the V1–V9 ledger.blocked:upstream -
9.6 Wire G1–G5 + G6 (Sekhmet, §11) + G7 (canon, §1.4) + G8 (preregistered human-quality launch evidence) into a single
ReleaseDecisionper case; only all-eight-pass publishes.
Canonical-provider audit note (2026-07-27): released Metis reasoning guidance
now enters V8 through @oshun/v8-metis-detective-learning-bridge, strictly
behind Clew's authoritative minimalProofTreeHint and an exact all-eight-gate
published ReleaseDecision. The bridge exports one already-released solve-path
clue and never exports ground truth, the culprit, the conclusion, the deduction
path, unreleased evidence, or release evidence. It cannot unlock evidence, drive
interrogation, make or score an accusation, generate or repair a case, publish,
mutate canon, place a case, or adapt the player model. This proves governed
player-learning admission, not a rendered V8 client.
Phase 10 — Oracle: the open-world case director [P1]#
Decides which cases appear where/when/for whom. Drama/experience manager (
sota§1).
- 10.1 Create
ylib:case-director. Maintain a per-player model (skill, pacing, preferences, open canon threads). - 10.2 Mint-ahead queue: keep N verified cases warm per player/cell
within budget (0.9); request new
CaseSpecs from the pipeline as the queue drains. - 10.3 Placement: surface cases on the Bureau Case Board, as
player-commissioned cases (Commission Studio,
features§"6"), and as ambient open-world mysteries seeded into the living city (tie to V4/V5 world-state + NPC schedules). - 10.4 Pacing: select difficulty/theme to match the player model; avoid repetition; weave open canon threads (e.g., a freed wrongly-accused suspect returns).
- 10.5 [P2] Seasonal/live-service cadence: weekly generated cold-case drops (replacing V5's authored weekly packs) through the live-service calendar.
Phase 11 — Cross-cutting hardening [P1 unless noted]#
Everything that makes generation safe, reproducible, affordable, and shippable.
- 11.1 Budget governance: enforce per-case + per-player + global
token/asset ceilings via
budget-management; Oracle throttles mint rate; dashboards for spend. - 11.2 HITL review:
libs/yemaya/agents/src/hitlqueues for repair escalations, commissioned cases, and flagged content; reviewers see ground truth + VerifierReport. - 11.3 Provenance (C2PA): every generated asset C2PA-signed (V3
C2PA_EVERY_EXPORT.md); manifest records seed + model/provider versions + job ids. - 11.4 G6 — safety: V7 Sekhmet scans all generated text/art/audio (abuse/CSAM/defamation/IP) before publish; Oracle won't surface unscanned cases.
- 11.5 Determinism/replay: prove
(seed, model-versions, canon-hash)reproduces a case bit-for-bit at the IR level; journal every stage (0.8); add areproduce <caseId>CLI. - 11.6 Multiplayer canon sync: co-op cases share one canon snapshot; outcome write-back leader-arbitrated through V5 online services (arch§5).
- 11.7 [P2] Localization: per-locale text + VO generation (extend V6/V7 loc); keep deduction logic locale-invariant.
- 11.8 [P2] Accessibility: captions for all VO, colorblind-safe evidence cues, difficulty assists, dyslexia-friendly case text (extend V6 accessibility).
- 11.9 Telemetry/evals: stream solve-rates, accusation-rating distributions, abandon-points to Maat-style dashboards; alert on quality drift.
- 11.10 Cost/perf load test: sustained mint throughput at target concurrency within budget and latency; verify the relay (§9-arch) under fleet load (reuse V3 load-validation patterns).
Phase 12 — Release gates & docs-drift [P1]#
- 12.1 Implement the separate seven-gate platform content harness
(G1–G7) as a CI job mirroring V7's
ADVERSARIAL_EVAL_GATES.mdpattern: unsolvable-case gate, ambiguous-solution gate, unfair-clue gate, in-game-unsolved gate, low-quality gate, unsafe-content gate, canon-contradiction gate. Each has an adversarial fixture corpus; the authoritativeReleaseDecisionadditionally requires G8 human-quality launch evidence. - 12.2 Adversarial red-team corpus: hand-craft known-bad cases (unsolvable, ambiguous, unfair, canon-breaking, unsafe) and assert every gate catches its class.
- 12.3 Docs-drift CI (
verify:v8 docs-drift): keepV8_features.md/V8_ARCHITECTURE.md/V8_DEPENDENCIES.md/ this file aligned with shipped behavior and with the V5 header-derived contracts (0.2). - 12.4 End-to-end acceptance (
features§"Definition of done"): a scripted run that, from a single player request, mints → verifies → realizes → asset-generates → compiles → in-game-solves → safety-clears → publishes → places a case, fully reproducibly, with zero human authoring. All eightReleaseDecisiongates and all seven platform content gates green. - 12.5 GA sign-off: every
[P1]task above[x], both gate suites enforced, the regression corpus (9.5) green, and the acceptance run (12.4) reproducible from seed. <!-- The substantive criteria ARE met: the gate suites are enforced (loom-service release harness), the §12.4 acceptance run is reproducible from seed, and a regression corpus is green. GA cannot be honestly signed while §4.6 and §9.5 (both P1) remain blocked on V5-authored content that does not exist as data. 87/90 is the honest ceiling without fabrication. --> Board tag 2026-09-18: follows 4.6 and 9.5.blocked:upstream
Appendix A — Gap → Phase traceability (every gap§ is closed)#
Gap (V8_GAP_ANALYSIS.md) |
Closed by |
|---|---|
| A — no case generator | §2 (Clew), §3 (Minos) |
| B — factory not wired to game | §7 (Daedalus compiler), §8 (UE plugin), §9-arch (relay) |
| C — orchestration/integration scaffold | §0.5–0.7, §4 (Anansesɛm), §6 (Ori-Detective) |
| canon/continuity | §1 (Palimpsest), §11.6 |
| asset realization | §5 (Loom) |
| automated solvability/eval | §9 (Theseus), §3 (Minos) |
| runtime delivery / live NPC | §8, §6.5, §9-arch relay |
| open-world direction | §10 (Oracle) |
| safety/provenance/determinism/budget/loc/a11y | §11 |
| release gating | §12 |
Appendix B — The eight ReleaseDecision gates → tasks#
G1 §3.2 · G2 §3.3 · G3 §3.4 · G4 §9.2 · G5 §9.3 · G6 §11.4 · G7 §1.4 · G8 human-quality launch evidence — all enforced by §9.6. The separate seven-gate platform content suite is enforced by §12.1.