conversion/

README

Semantic-disposition conversion evaluation

This evaluation checks one small skill conversion. It uses this operator:

[ C = clauses(S) ]

[ \Pi(S)={(c,d,T,r)\mid c\in C} ]

Each source clause gets one disposition. T lists its destinations. r gives the reason for an exclusion.

How the projection works

1. Split the source SKILL.md into clauses.
2. Assign one disposition to each clause.
3. Map each retained clause to a reachable code or model-facing destination.
4. Give a reason for each excluded clause.
5. Pass the source and projection into flow extraction, writing, and repair.

Source clause	Disposition	Required destination
Run tests and stop on failure.	control	Code
Prefer changed-code evidence.	guidance	SKILL.md or agent_task
Ask for approval and explain why.	both	Code and model-facing guidance
Yarn is only release history.	excluded	No destination; give a reason

Run the paid evaluation:

npm run eval:conversion

The command starts exactly two fresh Codex sessions. Both use gpt-5.6-sol with medium reasoning. The first session runs the real bootstrapped builder. The second session judges the generated result and a control-only negative candidate. Raw transcripts stay under ignored evals/runs/.

Run deterministic checks:

npm run test:conversion -- --force

This is advisory evidence for one four-clause fixture. It does not cover other skills or conversions.