Run executable rubric-spec adapter checks before claiming schema compatibility.
eval-conformance-suite is for adapter maintainers and evaluation-platform integrators. Its differentiator is an executable, framework-neutral contract: it imports a target Python adapter module and runs 18 synthetic cases covering scale types, anchors, weights, tie-break rules, judge prompt contracts, provenance, validation errors, lint findings, deterministic serialization, and round-trip field preservation.
runwrites a JSON report with the target module, overall pass state, and every case name, boolean result, and detail string.badgewrites SVG text to stdout for a caller-supplied label.
The badge command does not run conformance or read a report; it currently renders a passing badge unconditionally. Generate and publish a badge only after independently confirming that the run JSON passed.
Checks run in the local Python process and make no built-in network requests or model calls. The --against value is imported with importlib, and the suite calls that module's to_spec and from_spec functions. Running against an untrusted adapter executes untrusted Python code.
python -m pip install eval-conformance-suite==0.1.2rubric-spec>=0.1.1 is installed as a dependency. For development from a clone:
python -m pip install -e .eval-conformance-suite run \
--against rubric_spec.adapters.inspect_ai \
> conformance.json
eval-conformance-suite badge \
--against rubric_spec.adapters.inspect_ai \
> rubric-spec-conformance.svgThe bundled CI also exercises the PromptFoo, DeepEval, LangSmith, and EvalKit adapters shipped by rubric-spec.
See docs/architecture.md and the synthetic report template in examples/conformance_report_template.md.
Registry status verified July 13, 2026: version 0.1.2 is published on PyPI and tagged v0.1.2 in the public repository. The project is alpha software. No framework certification, partnership, or adoption claim is made.
Passing means the adapter satisfied this package's synthetic rubric-spec cases. It is not a framework benchmark, security audit, production-runtime test, or independent certification.
Run the suite against the trusted adapter module you intend to support, inspect all 18 case details, fix every failure, and publish a badge only after the JSON report passes.