[TRAINING] Add Evennia telemetry, sample trace, and baseline eval (#37) #40

Timmy · 2026-03-28T19:20:55Z

Timmy commented

2026-03-28 19:20:55 +00:00

Implements the first #37 training/telemetry slice.

Adds:

canonical Evennia telemetry/session mapping contract
evennia_tools/training.py
sample trace generator
replay/eval harness for world basics
committed example trace/eval artifacts
baseline training report

Important truth:

this PR is a baseline, not a fake-green story
the eval harness currently exposes a real world-state/control issue: login is landing in Chapel instead of the expected Gate-first path
that is useful and intentional: the harness is doing its job by surfacing the gap early

Proof artifacts:

reports/production/2026-03-28-evennia-training-baseline.md
training-data/evennia/examples/world-basics-trace.example.jsonl
training-data/evennia/examples/world-basics-eval.example.json

[TRAINING] Hermes/Evennia telemetry, replay, and DPO/eval alignment (#37)
follows merged world-body PR #38

Implements the first #37 training/telemetry slice. Adds: - canonical Evennia telemetry/session mapping contract - `evennia_tools/training.py` - sample trace generator - replay/eval harness for world basics - committed example trace/eval artifacts - baseline training report Important truth: - this PR is a baseline, not a fake-green story - the eval harness currently exposes a real world-state/control issue: login is landing in Chapel instead of the expected Gate-first path - that is useful and intentional: the harness is doing its job by surfacing the gap early Proof artifacts: - `reports/production/2026-03-28-evennia-training-baseline.md` - `training-data/evennia/examples/world-basics-trace.example.jsonl` - `training-data/evennia/examples/world-basics-eval.example.json` Related: - #37 - follows merged world-body PR #38

Timmy added 1 commit 2026-03-28 19:20:56 +00:00

feat: add Evennia training baseline and eval harness (#37 ) 3f5a174943

codex-agent added 1 commit 2026-03-28 19:33:45 +00:00

fix: make Evennia training replay deterministic (#37 ) aca26da71f

Timmy referenced this pull request

2026-03-28 19:34:14 +00:00

[TRAINING] Hermes/Evennia telemetry, replay, and DPO/eval alignment #37

Timmy merged commit 0a421e5ba5 into main

2026-03-28 19:35:05 +00:00

Timmy deleted branch codex/evennia-training

2026-03-28 19:35:05 +00:00

Timmy referenced this issue from a commit