Target Evidence Designer

Use this skill after the generic evidence harness exists. The goal is to replace placeholders with evidence derived from the target project itself.

Workflow

1. Inspect the repo contract:

npx evidence-kit inspect --json
npm run docs:perf:search -- contract benchmark fuzz source
npm run research:list

Read public exports, README/API docs, existing tests, issue/regression fixtures, benchmark scripts, and package boundaries. Do not invent a contract that source or docs do not support.

2. Write a short target evidence map:

• public surfaces that users import or execute,
• behaviors the project documentation promises,
• existing tests that already encode contracts,
• known failures or unsupported cases,
• performance claims that maintainers actually care about,
• external sources worth fetching.

3. Create a target-owned evidence plan with five concrete outputs:

• research/<topic>-sources.json: primary sources to fetch.
• research/<topic>-sources.md: what was inspected and transferred.
• corpus file: seed cases and known failures.
• test/fuzz/*: generated cases plus a stable oracle derived from the target contract.
• benchmarks/*: fixtures tied to specific target performance claims.

4. Add or update a repeatable fetcher:

npx evidence-kit add-source-fetcher --name <topic>
npm run research:fetch -- <topic>

Use git, url, npm, file, or inline source entries. Record URL, commit/hash, local path, and why each source exists.

5. Convert target contracts into executable checks before writing random generators. Each check must name:

• the source of truth for the contract,
• the generated or corpus input shape,
• the expected observation,
• the failure artifact to write on mismatch.

6. Promote corpus seeds deliberately from target evidence:

• documented behavior,
• existing tests,
• fetched source material,
• user-reported failures,
• reduced fuzz failures.

7. Design benchmark fixtures from claims the project will make. Each benchmark row needs a category, fixture, metric, status, and output path under benchmarks/results/*latest.json. Keep bad-result, unsupported, timeout, and over-budget rows visible.

8. Run focused verification:

npm run test:evidence
npm run fuzz
npm run bench:evidence
npm run bench:startup:check
npm run bench:package:gates
npm run docs:perf
npm run docs:perf:search -- <topic>

Rules

• Do not stop at a prose plan when a small executable fixture can be added.
• Do not use external source names as local benchmark names unless the benchmark actually measures that source.
• Do not add runtime dependencies from source mining without benchmark and boundary evidence.
• Keep generated corpora small, readable, and replayable; cache large source material under ignored paths.
• Record rejected ideas with reasons, not silence.