Eve library

@nous/wm-ops

Other

Revision 1

On this page

Serving operations and evaluation for the Phase 176 world models (Phase 176.8, TS side; the Rust inference runtime lives in @nous/wm-serving).

  • DurableSessionStore — persists a session's durable state tree (memory bank + seed + prompt log) gzip-compressed into object storage (MinIO boundary) with HMAC-signed, verifiable session_id → resumable URL. (176.8.1.2)
  • MultiplayerWorld — two+ Maya users share one Genie world (same seed + memory bank); per-tick inputs merge deterministically and simultaneous prompts resolve by a deterministic order so clients never diverge. (176.8.1.3)
  • GpuAutoscaler — session↔GPU affinity (stickiness), utilisation-based scale up/down, and graceful drain that migrates a node's sessions to the least-loaded survivors without dropping them. (176.8.1.4)
  • BenchmarkHarness — runs a checkpoint across the standard suites (Atari-100k, Crafter, Procgen, DMLab-30, Minecraft-diamond, MineRL BASALT, RT-Bench) and emits human-normalised scorecards with mean/median HNS. (176.8.2.1)
  • PoetEvaluator — the POET co-evolutionary loop (mutate → minimal-criterion
    • novelty filter → optimise → transfer) scoring emergent complexity. (176.8.2.2)
  • WorldModelTrajectoryCollector — the Phase 85 flywheel extension: ingests session transitions, computes a real quality label, and emits @nous/world-model-core episodes for continual retraining. (176.8.2.3)
  • HumanFactorsEvaluator — consistency / memory-fidelity / prompt-adherence / enjoyability scoring from paid raters with an LLM-judge backstop, driving a promotion gate. (176.8.2.4)
text
npx vitest run --root libs/nous/wm-ops