V4 Autonomous Content Agents — Audit (2026-07-14)

Audit of every system in V4/ that autonomously creates player-facing content, following the same method and severity bar as the V2 (2026-07-13) and V3 (2026-07-14) autonomous-content audits: read every public function and the private functions they delegate to, run the adversarial stub-indicator scan, check that quality/consent/review gates can actually fail, and verify test coherence (a test that passes against a fabrication is itself a finding).

Scope — what "autonomous content agent" means in V4#

V4 is the multi-cell UE5 game. Its autonomous content surface is:

Out of scope: Workshop/Contracts-Author editors (user-generated content tools, not autonomous), web CMS content, esports tooling.

Verdict#

Generation logic: real and strong. Trust claims: were fabricated in the two commentary pipelines — now fixed fail-loud. The AI Director, Procgen, roguelike campaign, crowd, and schedule systems are genuine domain code with tests that assert exact values and gates that can fail. The commentary pipelines had the same fabrication class the V2/V3 audits found elsewhere: consent/review/watermark/queue claims manufactured from policy constants instead of evidence. As in V3, everything that "generates" prose/audio is either honest deterministic template logic (now labeled as such) or a provider seam (Calliope/Psyche) that is planned but not executed in-engine — V4 is NOT a SOTA autonomous generator end-to-end; it is a well-built orchestration/gating layer around declared provider seams.

HIGH findings (fixed this audit)#

H1 — Spectator live commentary fabricated bias review + watermark#

UV4LiveAICommentaryBroadcastService::GenerateLiveCommentaryLines set bBiasReviewed from the policy's own constants (bReviewPlayerGender && bReviewPlayerRegion && ... — always true), bWatermarked = true unconditionally (no watermark exists), and derived bReadyForBroadcast from both — so every generated line claimed human bias review and provenance watermarking that never happened, while the policy itself declares bBlockUnreviewedBroadcast = true. The struct default bWatermarked = true was fail-open. SpectatorModeSpec asserted the fabricated flags.

Fix (fail-closed evidence seams): generated lines now start bBiasReviewed = false, bWatermarked = false, bReadyForBroadcast = false (struct default flipped too). New FV4LiveCommentaryBiasReviewRecord and FV4LiveCommentaryWatermarkReceipt types carry evidence; ApplyBiasReview accepts a record only from a queue in Policy.ReviewerQueueIds, with a named reviewer, all policy-required dimensions reviewed, and an explicit approval; ApplyWatermarkReceipt requires watermark + provenance-manifest ids bound to the line; IsLineReadyForBroadcast blocks unreviewed output whenever bBlockUnreviewedBroadcast is set. Spec rewritten: unreviewed lines blocked, approved-review + watermark → ready, unapproved/wrong-queue/anonymous records rejected, watermark-without-review still blocked.

H2 — Replay end-of-match commentary fabricated generation/queue/consent claims#

UV4AIGeneratedCommentaryPipeline::GenerateEndOfMatchCommentary returned bCalliopeGenerated = true (no Calliope stack was invoked — the script is deterministic template text), bPsycheSynthesisQueued = true (nothing was queued; the method is BlueprintPure), bShareable = true, and PsycheVoiceRequest.bConsentChecked defaulted to true in the struct — a fail-open consent claim for AI voice synthesis. ReplayModeSpec asserted the fabricated flags ("End-of-match commentary is Calliope-generated", "Psyche voice synthesis is queued").

Fix: new EV4CommentaryGenerationMethod records how the script was produced; the pure generator now returns GenerationMethod = DeterministicTemplate, bCalliopeGenerated = false, bPsycheSynthesisQueued = false, bShareable = false. AttachCalliopeGeneratedScript flips bCalliopeGenerated only with a Calliope result bound to the request id, the Calliope.V1.Commentary stack, a named model id, and one non-empty script per segment. MarkPsycheSynthesisQueued flips queued/shareable only with a queue receipt (queue id + receipt id) matching the voice request, and only when the consent gate was actually evaluated. bConsentChecked struct default flipped to false; the generator still sets it true because it genuinely evaluates the player opt-out before building the voice request. Spec rewritten to drive the evidence paths positively and negatively.

MED findings (fixed this audit)#

M1 — Localization nuance "review plan" pre-approved itself#

V4MakeCommentaryLocalizationReview built the 12-locale native-speaker review plan with bNuanceApproved = true and bNoProtectedTraitInference = true (struct defaults also true) — verdicts recorded before any native speaker looked at anything, and ValidateAICommentaryExtended + ReplayModeSpec asserted the pre-approval.

Fix: plan rows now start unapproved (bNuanceApproved = false, bNoProtectedTraitInference = false, defaults flipped); new FV4NativeSpeakerNuanceApproval + RecordNativeSpeakerNuanceApproval require a matching review/locale, a named reviewer, and explicit approved/no-protected-trait verdicts, and store ApprovedByReviewerId. ValidateAICommentaryExtended now checks the plan is honestly pending AND that the approval seam works (named reviewer flips it; anonymous approval is rejected). Tests updated to the same shape.

Findings verified as honest (no fix needed)#

• AI Director — weighted pressure model (0.42 damage / 0.25 deaths / 0.20 detection / 0.13 objective), deadband step resolution, percentile skill/parry calibration nudges, rubber-band perceptibility rejection (0.25 scalar / 120 ms / 1 parry-frame / 45 s caps), PvP hooks compile-disabled via static_assert over the mode policy table. ValidateLaunchCoverage/ValidatePostLaunchRefinement exercise behavior with contrasting performance windows — they can fail. Spec asserts exact steps, scalars, and reasons.
• Procgen — seeded (FRandomStream(Seed + Attempt * 7919)) constraint solver with bounded retries, BFS playability validation through spawn→objectives→extraction waypoints, blocker placement with rollback when it breaks the path, NPC-placement validation with duplicate/coverage checks, 55-chunk launch library with per-cell constraint tags. Daily seed is server-dispatched; note the signature is only checked non-empty client-side (verification is a server-boundary concern).
• Roguelike campaign — real run-state machine (loadout commits, boon grants, boss defeats, meta-progression banking/spending, operator death) over real procgen layouts.
• Crowd/Schedules — relevance-sorted tier demotion under a measured Mass tick budget, exfil flee assignments with real distances; schedule assets and components are data-driven with validation.
• Compliance subsystem — authored policy-config (DPIA cards, jurisdiction notes, region feature flags, residency rules) with validators that check the config against jurisdictional requirements and can fail on edit. These are declared policies — publishing them (legal portal URLs, annual refresh) is an external launch boundary, mirroring V3's honest-red pattern.
• Podcast recap bGenerated — kept: in the §137 manifest schema (generatedPerCell) this is an AI-generated-content classification used for disclosure (paired with bWatermarkRequired/bHumanReviewRequired), not a claim that audio exists. OutputAudio is the planned target path, matching the manifest.

What remains provider-/deploy-boundary (not locally closable)#

• Actual Calliope (LLM) script generation and Psyche voice synthesis — the engine builds requests and now refuses to claim their results without evidence; executing them requires the provider services.
• Real Trust & Safety / native-speaker review queues feeding the new evidence records.
• Watermark/provenance embedding in synthesized audio (receipt seam exists).
• The final-polish gate (scripts/test/final-polish.test.mjs) is honestly red on two pre-existing unchecked external-boundary TODOs (§56.13 voice SFU full implementation, §58.1 platform integrations) — unrelated to the content agents but part of the honest launch picture.

Verification#

• node src/v4-ai-generated-commentary-check.mjs — passed
• node src/v4-ai-generated-commentary-extended-check.mjs — passed
• node src/v4-live-ai-commentary-check.mjs — passed
• node src/v4-ai-director-check.mjs --manifest ...DA_AIDirectorLaunch... — passed
• node src/v4-roguelike-adventure-check.mjs — passed
• node src/check-syntax.mjs — passed (175 files)
• node --test test/*.test.mjs — 8/8 commentary/director suites pass; only pre-existing final-polish red (see above)
• UE C++ compile/automation run is not possible on this macOS box (engine build lives on the Linux box); the edits follow existing UE idioms in the same files and all repo-side verifier needles (class/method/test names) were kept intact.

Adversarial scan record#

Mandatory grep (stub-comment indicators, --include C++/C#) over V4Procgen, V4Gameplay, V4Cinematics, V4Crowd, V4Schedules, V4OnlineServices, V4Mode_Replay, V4Mode_Spectator, V4Mode_Roguelike_TacticalCampaign: zero actionable hits. Silent-stub scan: no unseeded randomness in deterministic paths (all FRandomStream seeded), no catch-and-fabricate blocks, no ignored _-params; the constant-vs-constant gates found are exactly the H1/H2/M1 items above, now evidence-gated.