DECISION.md v0.3.0

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 test keep 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

DECISION.md v0.2.0

Initial versioned release of the DECISION.md open format. Includes the format specification, template, examples, JSON Schema, Markdown linter, structured profile validator, integration guide, behavior demo, eval cases, and open-source contribution hygiene.