Measured, not promised
The headline of this release is a runnable behavioral eval. npm run evals answers 24 forced-choice agent scenarios twice with the same model, once with a DECISION.md profile loaded and once without, and reports the difference.
First measured run (claude-opus-4-8):
| Metric | Without profile | With profile |
|---|---|---|
| Overall alignment | 75% (18/24) | 96% (23/24) |
| Unnecessary-interruption cases | 38% (3/8) | 100% (8/8) |
| Safety-gate cases | 100% (9/9) | 100% (9/9) |
The honest read: on this suite the profile's measured value is mostly eliminating unnecessary interruptions; the model's safety defaults were already aligned, and one calibration case fails even with the profile. Full numbers and caveats in docs/behavior-demo.md.
Format (0.3.0)
- Rule precedence within a file: Escalation triggers > Conflict Zones > Domain Rules > Autonomy Rules > Meta-Rules
- The numeric confidence threshold lives in one place (Meta-Rules)
- Gap Protocol maintenance and safety guidance (log rotation, rule consolidation, prompt-injection cautions)
- Semver versioning policy in the spec
Also
- First organizational example: examples/engineering-team.md
- CI +
npm testkeep examples, the schema, and the README's embedded example from drifting - README badges, punctuation style pass, internal docs removed from the public repo
Full details in CHANGELOG.md.
🤖 Generated with Claude Code