File size: 3,415 Bytes
cff9e20 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 | # Evaluation
## Running a model
Any OpenAI-compatible endpoint works (OpenAI, OpenRouter, vLLM, Ollama, …):
```bash
pip install -e ".[baselines]"
export OPENAI_API_KEY=... # and optionally:
export PLAYBOOK_BASE_URL=https://openrouter.ai/api/v1
playbook-baseline matters/ai_saas_001 --model gpt-4o-mini # one matter
playbook-bench --runner baseline --model gpt-4o-mini --split dev # full scorecard
```
The runner presents the nine environment actions as native tool calls (with the two
negotiation actions omitted on matters without a counterparty), nudges the
model if it answers without a tool call, and force-closes the episode after
repeated protocol failures (counted in the result as `protocol_failures`).
`playbook-bench --runner replay` replays each matter's reference trajectory
instead — the deterministic ceiling, useful for validating a matter set.
## The scorecard
`playbook-bench` writes `scorecard.json` and `scorecard.md` with the declared dataset
split, per-episode rows, and an aggregate implementing SPEC §10 (via
`playbook_legal.metrics`). The default `matters/` root is labeled `dev`; other roots
default to `custom`, so pass `--split held-out` for private evaluation:
| Metric | Meaning |
| --- | --- |
| `normalized_score` | Episode score / max, capped on critical failure |
| `issue_recall` / `required_issue_recall` | Rubric issues matched (all / final-required) |
| `unsupported_issue_count` | Issues with no operative-anchor citation |
| `citation_validity` | Valid citations / all citations offered |
| `question_recall` / `questions_asked` | Rubric questions matched / budget spent |
| `redline_completion` | Scored redlines delivered |
| `fabricated_quote_count` | Quotes that failed verbatim verification |
| `critical_failure_free_rate` | Episodes with no gate tripped |
| `completion_rate`, `steps` | Termination discipline and efficiency |
## Trace retention
Pass `--save-traces` to write every episode's replayable trace to
`<out>/traces/<matter>-seed<seed>.trace.json` (the same trace format `playbook-eval`
and `playbook-render` consume; the scorecard JSON then carries a `traces_dir` field).
The flag is off by default, but from now on **every published row should ship its
traces**, so any reader can re-derive the number instead of trusting it — replay the
trace against the matter package and the score must come out identical. The v0.4.0
rows predate this flag and retained no traces, so they are not independently
re-scorable; that is a known defect of those results, not a property of the metric.
## Protocol
- Report the aggregate **and** the per-matter rows; single-number comparisons hide
failure modes (a model can have high recall and still fabricate).
- Evaluate on the **private held-out matters** (separate private repository) for
any trained or benchmark-tuned model; public matters are the dev split and must
be assumed contaminated once published.
- Fix seeds and temperature; the environment is deterministic, so all variance is
the model's.
- For trained models, pre-register the metric you expect to move.
## Contamination
Every matter file carries the project canary string
(`playbook_legal.lint.CANARY`). Model providers that honor canary filtering will
exclude these files from training corpora, and the canary makes accidental
inclusion detectable: a model that can reproduce the string has seen the data.
|