Ariadne · Reference & analysis

V8 — Gap Analysis (Ariadne: The Self-Authoring Detective Universe)

V1–V7 already deliver (a) a _real, playable_ open-world detective game with a Sherlock-style deduction engine (V5, ~107K lines UE C++), and (b) a _real_ multi-modal AI content factory (isis image/music/voice/3D/video, hathor LLM-NPC,

4sections5 minread1table

On this page

Status: the authoritative map of what is missing between today's V1–V7 and a fully-AI-generated open-world detective universe. Date: 2026-06-04. Every gap below is grounded in a real file/interface (cited), and maps to a V8 subsystem (V8_ARCHITECTURE.md) and a V8_TODOS.md phase.

Convention: HAVE = real, wired code today. PARTIAL = scaffold/templates /types only. MISSING = no implementation. names the V8 closer.


0. The thesis in one paragraph#

V1–V7 already deliver (a) a real, playable open-world detective game with a Sherlock-style deduction engine (V5, ~107K lines UE C++), and (b) a real multi-modal AI content factory (isis image/music/voice/3D/video, hathor LLM-NPC, yemaya's CrewAI/LangGraph creative agents + a Knox/Van-Dine mystery-fairness validator). They are not connected, and the one thing neither half can do is generate a complete, fair, solvable case. V8 builds the missing generator, proves solvability formally, and compiles output into V5's existing authored case format so it ships with no engine rewrite.


1. The three macro-gaps#

Gap A — There is no case/mystery generator (only a validator)#

  • HAVE: libs/yemaya/agents/src/quality-assurance/mystery-fairness-validator.ts — a real, domain-specific fairness checker (Knox Decalogue + Van Dine 20 rules) over a MysterySession (clues: Map<ClueId,Clue>, solutions: Map<SolutionId,Solution>, characters, timeline) returning a FairnessAnalysis (score, rating, violations). It validates a case.
  • HAVE: …/quality-assurance/investigation-mechanic-balance-analyzer.ts — scores clue-distribution, deduction-chain completeness, red-herring fairness.
  • MISSING: anything that produces a MysterySession. No solve-first ground-truth builder, no clue-derivation pass, no uniqueness enforcement, no red-herring generator. Confirmed by repo-wide search: only validators exist.
  • Clew (generative case engine) + Minos (formal solvability proof), feeding the existing validator as a gate. V8_TODOS Phase 2–3.

Gap B — The agent factory is not connected to the game#

  • HAVE: a real UE5 detective game. Case data contracts exist and are loaded from a JSON manifest:
    • V5/ue/Source/V5MindPalace/Public/V5MindPalaceTypes.h:33 FV5MindPalaceEvidenceNode (NodeId, CaseId, SourceId, Cell, LocationId, TimestampInFiction, Tags, Description, Image:FSoftObjectPath, ExaminedDetails, bFalseLead).
    • same file :72 FV5MindPalaceDeductionEdge (FromNodeId, ToNodeId, DeductionText, Confidence, Kind, bCrossEra, RequiredBureauTier, UnlockFlag).
    • same file :117 FV5MindPalaceAccusationOutcome (AccusedId, EvidenceChain, RequiredEdgeIds, Rating∈{Brilliant,Good,Doubtful,Wrong}, OutcomeBranch∈{TruthRevealed,PlausibleConviction,PoliticalCoverup,Unproven}).
    • V5/ue/Source/V5DetectiveColdCases/Public/V5DetectiveColdCasesTypes.h:17 FV5ColdCasePackDefinition (PackId, WeekNumber, CaseId, LeadCell, SourceCells, EvidenceNodeCount, DeductionPairCount, VoiceOverMinutes, VoiceLineCount, DropSlug, RewardId, Title).
    • on-disk: V5/ue/Plugins/V5Mode_Detective_ColdCases/Content/Data/cold_cases_manifest.json (schemaVersion:1, packs[], each mindPalace.evidenceNodeIds[] + deductionPairIds[]), validated by V5/tools/cold-cases/validate-cold-cases.py.
  • HAVE: an asset factory — libs/isis/client IsisClient.submitJob() / submitAndWait() returning JobOutput.files[].url; libs/hathor/llm-npc DialogueGenerator.generateResponse().
  • MISSING: the compiler that turns a validated MysterySession (+ its assets) into FV5* structs + a cold_cases_manifest.json pack; the asset manifest binding generated audio/mesh/texture to FSoftObjectPaths; the runtime relay so the game can request cases on demand. yemaya/agents has zero references to V5/Unreal today.
  • Daedalus (case compiler + runtime relay). V8_TODOS Phase 4 & 9.

Gap C — Orchestration & integrations are scaffold, not live#

  • PARTIAL: apps/lilith/svc-ai/src/agent-coordination.ts:24 carries the literal marker // stub:legitimate Until the coordination protocols are wired to a real LLM … the response strings and confidences are scaffolds.
  • PARTIAL: libs/neith/integration-hathor narrative/quest generators are deterministic templates (Freytag tension tables, fixed complication graphs), not LLM-driven.
  • PARTIAL: yemaya *-integration dirs (isis/hathor/sophia/bellona) are type-layer wrappers around @isis/client, @hathor/client, @sophia/client; not end-to-end pipelines.
  • HAVE (substrate): yemaya WorkflowBuilder / WorkflowExecutor / StateGraph (LangGraph-style) and CreativeAgent base (real llmProvider.complete() calls) — the right bones, under-used.
  • Promote to live: Anansesɛm writers'-room DAG on the existing orchestration substrate; wire the integration adapters to real job calls. V8_TODOS Phase 5–6.

2. Capability matrix — every element a fully-generated detective universe needs#

Element Status today Evidence V8 closer
Open-world detective game loop HAVE V5 investigation/interrogation/Mind-Palace/accusation, ~107K lines C++ reuse
Sherlock-style deduction graph HAVE FV5MindPalaceDeductionEdge, pair-and-deduce reuse
On-disk authored case format HAVE cold_cases_manifest.json + validator.py Daedalus emits it
Mystery fairness scoring HAVE mystery-fairness-validator.ts (Knox/Van Dine) reuse as gate
Investigation balance scoring HAVE investigation-mechanic-balance-analyzer.ts reuse as gate
Image / music / voice / 3D / video gen HAVE isis/ai-providers + IsisClient.submitJob Loom orchestrates
LLM NPC dialogue HAVE hathor/llm-npc DialogueGenerator Ori-Detective bridge
Multi-agent orchestration substrate HAVE yemaya WorkflowBuilder/StateGraph Anansesɛm uses it
Creative agents (story/char/dialogue) HAVE yemaya CreativeAgent base + specialists Anansesɛm extends
Case ground-truth generator (solve-first) MISSING Clew Stage 1
Clue derivation tied to solution vars MISSING Clew Stage 2
Red-herring generator (fairly refutable) MISSING Clew Stage 2
Formal unique-solution proof (CSP/ASP) MISSING Minos Stage 3
Canon / continuity knowledge graph PARTIAL yemaya canon-enforcement lib exists; not graph-RAG Palimpsest Stage 0/4
MysterySession → FV5* case compiler MISSING Daedalus Stage 4
Generated-asset → FSoftObjectPath manifest MISSING Daedalus Stage 4
Automated solver / playtest eval MISSING Theseus Stage 7
LLM-as-judge quality gate (8-dim) MISSING Theseus Stage 7
Runtime case-on-demand relay (UE5) MISSING Daedalus runtime Stage 9
Live AI interrogation (ACE/Inworld) MISSING Ori-Detective + ACE Stage 9
Open-world case director / drama mgr MISSING Oracle Stage 10
Living suspects (memory/reflection/plan) PARTIAL V6 Ori spec exists; UE client is stubs Ori-Detective Stage 6
Cost / budget governance HAVE yemaya budget-management reuse
Human-in-the-loop review HAVE yemaya hitl reuse
Provenance / C2PA on exports HAVE V3 C2PA_EVERY_EXPORT.md reuse
Trust & safety on generated content HAVE V7 Sekhmet reuse
Multiplayer canon sync PARTIAL V5 online services; no generated-case sync Daedalus + Oracle Phase 9/11
Determinism / replay of generation MISSING seed+journal, Phase 11
Localization of generated cases PARTIAL V6/V7 loc extend, Phase 11
Accessibility of generated cases PARTIAL V6 accessibility extend, Phase 11

3. What V8 explicitly does NOT redo#

  • It does not re-author V5's hand-made 25 launch cases — those remain the golden reference corpus and eval fixtures for Theseus/judge calibration.
  • It does not rebuild the UE deduction UI, interrogation rig, or world streaming — V8 emits data the existing systems already consume.
  • It does not add a new asset provider — isis already wires Stability, Suno, ElevenLabs, Meshy, Hunyuan3D, LTX. V8 orchestrates them per case.
  • It does not introduce a new orchestration framework — it uses yemaya's WorkflowBuilder/StateGraph.

V8 is an integration + generation-completion version: one new generative core (Clew/Minos), one new bridge (Daedalus), and the promotion of existing scaffolds (Anansesɛm, the integration adapters, Ori) to live, verified, end-to-end production.