Skip to content

Maintained reference evaluations

This page records an editorial and contract review of three synthetic reference artifacts kept in the repository. Read the scope carefully: this is a source-level review, not evidence that an external agent executed a recipe or that a human approved an agent-produced outcome.

You can repeat every check yourself. Run pnpm demo to repeat each deterministic output-contract check, the disposable installation lifecycle, and an explicit review-pull-request invocation through the repository's fake fixture agent. Run pnpm validate:content to repeat the evidence-reference and cross-file content checks.

Fixture-agent evaluation

The demonstration installs review-pull-request for Generic Markdown in a disposable path that contains spaces. It then explicitly invokes scripts/demo-fixture-agent.ts with the installed workflow, the synthetic input, the committed reference output, and a new output path. The fixture agent proves that the runner reads those inputs and creates a distinct output file without overwriting an existing result. The demonstration then evaluates the produced file against output.schema.json before continuing.

Because the fixture agent deliberately replays a committed reference artifact, this is a runner and contract test. It is not evidence of autonomous reasoning or external-agent execution.

Debug failing CI

Review pull request

Synchronize documentation

Verification boundary

The maintained outputs are synthetic examples authored together with their inputs. Passing the output contract proves only that the required artifacts, populated sections and tables, literals, and evidence-reference rules are satisfied. The content validator additionally checks structural derivability signals, but reviewing whether a claim actually holds remains human work. No external-agent command, agent version, external execution artifact, outcome reviewer, or outcome approval is retained by this record.

No analytics. No cookies. No telemetry.