timmy-home/reports/production/2026-03-28-evennia-training-baseline.md

Evennia Training Proof — 2026-03-28

Issue:

• #37 Hermes/Evennia telemetry, replay, and DPO/eval alignment

What this slice adds:

• canonical telemetry contract for the Evennia lane
• session-id sidecar mapping path
• sample trace generator
• deterministic replay/eval harness for world basics
• committed example trace/eval artifacts

Committed example artifacts:

• training-data/evennia/examples/world-basics-trace.example.jsonl
• training-data/evennia/examples/world-basics-eval.example.json

Final result:

• replay/eval now starts from a deterministic Gate anchor using a dedicated eval account (TimmyEval)
• sample trace generation succeeds
• world-basics eval passes cleanly
• orientation: pass
• navigation: pass
• object inspection: pass

Canonical mapping:

• Hermes session id is the join key
• world events write to ~/.timmy/training-data/evennia/YYYYMMDD/<session_id>.jsonl
• sidecar mapping file writes to ~/.timmy/training-data/evennia/YYYYMMDD/<session_id>.meta.json
• the bridge binds the session id through mcp_evennia_bind_session

Why this matters:

• world interaction no longer disappears into an opaque side channel
• we now have a path from Hermes transcript -> Evennia event log -> replay/eval
• this complements rather than replaces NLE/MiniHack
• the persistent-world lane now has a real green baseline, not just an aspiration