Fighting Game · Guides & deep dives

V2 EU AI Act Surface

A guides & deep dives document in the Fighting Game library.

0sections1 minread
Markdown source
On this page

@v2/eu-ai-act-surface is the V2 adapter for EU AI Act shipping AI surfaces. It now publishes a registry for:

  • Adaptive AI Director: V2/docs/ai/model-cards/adaptive-ai-director.md.
  • AI commentary: V2/docs/ai/model-cards/ai-commentary.md.
  • Anti-cheat classifier suite: V2/docs/ai/model-cards/anti-cheat-classifier-suite.md.
  • Generation pipelines: V2/docs/ai/model-cards/generation-pipelines.md.

The Adaptive AI Director surface additionally composes:

  • @psyche/action-safety for classifier transparency and player opt-out.
  • @nous/safety for the hosted Model Card metadata.
  • @themis/accountability for the AI-system-of-record and conformity audit export.

Every registry entry is hosted by @nous/safety and registered by @themis/accountability. @v2/eu-ai-act-surface rejects mismatched Model Card paths between the Nous runtime metadata and Themis AI-system-of-record.

The Unreal V2AdaptiveAI plugin mirrors this contract with FV2AdaptiveAIEUAIActSurface. The surface is disclosure, settings, and audit data only: it is off-rollback, never reaches V2Netcode, and cannot become deterministic gameplay input.

Player opt-out is applied before adaptive classifier use. When enabled, V2 uses the static CPU profile and stops applying live Psyche or Shakti classifier hints for that player.

For the AI Director, the player-facing disclosure is Settings.Gameplay.AIDirector. That row lists recent win rate, session length, repeated rage-quit, repeated same-mistake, declining session time, and declining session frequency before the single-player Director can tune CPU difficulty, encounter pacing, loot pacing, or tutorial prompts. Opt-out locks single-player difficulty to manual and keeps ranked/tournament modes on the static rule-of-the-day.

Verification is owned by V2/ue/Tools/check-v2-eu-ai-act-surface.py, the targeted package tests for @psyche/action-safety, @nous/safety, @themis/accountability, and @v2/eu-ai-act-surface, plus the rollback determinism gate.