Environment-inference dataset
Exhibit A records each explicit executor prepare attempt as private,
environment-attempt/v1 research data. The recorder is a transparent executor
decorator: it returns the same environment handle and re-raises the same setup error.
It does not make dependency guesses or alter silence decisions.
Each record contains the engine version, repository identity and commit, executor,
bounded repository-shape markers, selected strategy, duration, success/failure, and a
truncated setup diagnostic. It never copies lockfile contents, source, credentials, or
arbitrary repository files. CLI investigations place records next to their Case output
under environment-attempts/; study commands use the same private research boundary.
Aggregate empirical recipes with:
cd engine
python3 -m exhibit_a.cli environment-summary \
.exhibit-a/environment-attempts \
--out .exhibit-a/research/environment-summary.json
The environment-summary/v1 output groups attempts by repository shape and strategy
and reports successes / attempts. This is an observed rate, not a promise that the
recipe will work on a new repository. Publishing requires a separate privacy review:
repository identities and build diagnostics are private by default.