V2 Autonomous Content Agents — Verification Audit — 2026-07-13

Question audited: are all the agents that autonomously create content for V2/ SOTA and production-ready?

Method: main-loop verification (no agent fan-out, per session limits). Inventory was drawn from V2/architecture/esports-companion-and-ai-services.md and V2/features/community-store-support-and-ai-services.md, then every unit was read in current source, the mandatory adversarial stub grep battery plus the silent-stub scan was run over all of it, delegation chains were followed into the shared substrates (@iris/conversation-orchestration, @iris/conversation-providers-anthropic, @psyche/avatar-lipsync, @iris/voice call-translation, @hathor/llm-npc, @isis/visual-dubbing), and every suite plus tsc --noEmit was executed on this box today.

Verdict#

Yes at the code level, with the same honestly-drawn deploy boundary the architecture docs themselves declare — after this session closed one real fail-silent defect. The V2 autonomous content plane is real, adversarially clean, rollback-sealed in both TypeScript and C++, and fully green in tests and typecheck. One genuine fabrication path was found and fixed (the realtime translation bridge accepted the @iris/voice local_default passthrough provider, which invents bracket-tagged pseudo-translations with a fake 0.7 confidence, and reported them as successful deliveries). Everything else absent is a named, fail-loud seam: provider credentials, plan→executor bindings, and the spec-only gRPC tier.

1. The agent inventory (what "the agents" are)#

2. Verified in current source (this session)#

• The rollback envelope is enforced, not described. TS: fromRollbackFrame is rejected by both the V2 cue validator and the Iris stream validator; assertValidLiveRequest throws on any live hint under rollback; snapshots serialize a sorted-key JSON vector with a real sha256 hash. C++: EvaluateAIDirectorDecision hard-falls to the static rule-of-the-day for ranked/tournament/multiplayer/opt-out before consulting any profile; FV2AICommentaryCueContract::IsValidContract rejects non-Calliope commentators, missing fallback seeds, and non-Iris package bindings.
• Fallback is deterministic, labeled, and never fabricated as success. The commentary stream returns source: 'pre-recorded-bank' plus an explicit fallbackReason (missing-orchestrator / latency-budget-exceeded / empty-response / error), selected by a stable hash over (cueKind,ruleset,commentatorId,seed) under a Promise.race latency budget.
• The generation core is a real provider chain. The orchestrator seam reaches @iris/conversation-providers-anthropic (shared LLM service over an OpenRouter-compatible HTTP client with typed errors) with @iris/conversation-providers-local as the on-cap route; @hathor/llm-npc ships a provider chain with circuit breakers, latency tracking, and fail-loud not configured errors for ACE/Inworld/Convai.
• The render math is computed, not asserted. Face animation invokes createPhonemeSequence → phonemeSequenceToVisemes → generateLipSyncFrames (3.6k-line substrate with real coarticulation blending) in-process; ducking computes intensity-based gain with density and loudness compensation — the spec now pins the hand-checked value (climax −10, density −1.5, loudness −3 → −14.5 dB, linear 10^(−14.5/20) ≈ 0.18837).
• Persona governance is cook-blocking. assertV2CommentatorPersonasReadyForCook rejects revoked voice clones and missing signoff scopes; the manifest carries consent ids, likeness licenses, similarity thresholds, watermarking and abuse monitoring flags per artist.

3. Defects found this session (all fixed in this commit)#

1. HIGH — silent pseudo-translation path (iris-realtime-translation.ts): the bridge's provider argument was optional; omitting it silently engaged the @iris/voice local_default passthrough, which returns [lang] text with a fabricated 0.7 confidence as a successful delivery — invisible to the session's reason codes, and the specs always injected a mock so tests never saw it. Fix: provider is now required (V2IrisRealtimeTranslationError, code missing_translation_provider), and any local_default delivery is flagged with an explicit reason code. Two regression tests added.
2. LOW — cwd-dependent test (calliope-commentator-personas): the spec resolved the repo-root-relative manifest against process.cwd(), so the suite failed 0/4 when run from the service directory. Fixed with an import.meta.url repo-root anchor.
3. LOW — un-runnable fact-checking suite (V2/tools/esports): the vitest alias map lacked @oshun/ai, whose package entry points at an unbuilt dist/, so veritas-esports-fact-checking.spec.ts failed at import. Fixed by aliasing @oshun/ai to source; 3/3 suites now green.
4. LOW — under-pinned assertion (euterpe-commentary-ducking spec): gain assertions were range-only (a hardcoded −7 would have passed); now pinned to the known-correct −14.5 dB / linear value.

4. Adversarial stub scan — zero actionable hits#

The mandatory grep battery (excuse comments, placeholder/TODO/dummy/fake vocabulary) returned zero hits across all 15 services plus V2/tools/esports. The silent-stub scan (Math.random in deterministic paths, fabricated-success catch blocks, ignored _-params) surfaced only honest code: typed error rethrows in the NPC bridge, ENOENT→cache-miss in the dialogue cache, URL-validation throws in MOTD, and a hostname fallback in fact-checking. The one genuine fail-silent path found (translation local_default) is fixed above.

5. Test + typecheck evidence (executed on this box, 2026-07-13)#

6. The honest boundary (what "prod ready" does NOT include)#

These are deliberately drawn seams, all fail-loud, most declared by the architecture docs themselves:

1. Generation providers are deploy-bound. Live commentary text, persona TTS, and face rendering come from the provider tier (per deps§67: Anthropic Claude for play-by-play, ElevenLabs for TTS, Isis ai-video for auto-edit). Absent credentials, the plane degrades to the deterministic fallback bank or silence with explicit fallbackReason — never fabricated output.
2. The gRPC wire tier — CLOSED 2026-07-14. The three AI-services contracts now ship at libs/proto/src/oshun/v2/{esports,director,ai_commentary}/ with @v2/grpc-services binding them to the real engines (veritas fact-checked Crown Points credit, rollback-rejecting Director, fallback- labeled Commentary streaming); 16 in-process round-trip tests. The other V2 proto packages (online, live, progression, community, racing) remain designed-only.
3. Isis plan→executor binding is a named seam. The lipsync, cosmetic, and venue-capture packages are plan contracts: plans never claim rendered output, and rendered/generated payload builders demand a real asset id, URI, and content hash. The dubbing/texturing engines exist as substrates (@isis/visual-dubbing et al.); the dispatch wiring is the deploy step.
4. UE C++ verified by source-read only on this box (no UE toolchain on this Mac; the compiled UE 5.5 build lives on the Linux sandbox). The TS mirrors of the same contracts are fully test-executed.
5. Editorial content stays human-authored by policy. MOTD, patch notes, community spotlights, and procedural fighter dialogue (seasonScope: 'post-launch-dlc', manualReviewRequired: true) are delivery-plane, not autonomous generation — matching the EU AI Act disclosure posture ("AI commentary is machine-generated presentation content.", cross-ref §80.4).

7. Bottom line#

• SOTA: yes — the V2 plane composes the same substrate stack the V1 audit verified as SOTA, and adds the parts that are genuinely hard in a fighting game: a rollback envelope sealed in both languages, a real-time latency budget with a deterministic provider-free fallback bank, cook-blocking voice and likeness governance, and post-match publication held behind claims-vs-replay-hash fact checking.
• Prod-ready: yes at the library + wiring level — every unit compiles, typechecks, and passes its suites today; zero result-faking stubs remain (the one found was fixed and regression-tested this session); every absent capability is an explicit fail-loud seam owned by a named deploy step.