Evidence model and rules
Four channels
Every claim in a report lives in exactly one channel. They are never blended into a single "readiness score".
| Channel | Meaning | Example |
|---|---|---|
| Verified | The tool executed a check and kept the result, or hashed a named file | README.md SHA-256; output of git --version |
| User-reported | A person or snapshot says it happened; not checked here | "npm test passed" in your evidence file |
| Inferred / recommended | Heuristic judgment from labels, text patterns or metadata | "DCO may apply (CONTRIBUTING.md:5)" |
| Not run / unknown | No execution evidence, API unavailable or intentionally excluded | Tests were not executed; maintainer intent unknown |
Evidence tiers in the audit
| Tier | Proves | Does not prove |
|---|---|---|
VERIFIED_LOCAL_HASH | A file existed at scan time with that SHA-256 | That it is correct, current upstream or approved |
VERIFIED_LOCAL_PROBE | The version a fixed probe printed on this host | That the project builds or tests pass |
INFERRED_HEURISTIC | A line matched a pattern (quoted with line number) | Legal meaning or the maintainer's intent today |
OFFLINE_SNAPSHOT | How a supplied snapshot ranks under stated rules | That an issue is still open, unassigned or welcome |
NOT_RUN | Which claims are unverified | That any test passed |
The ten rules
- Untrusted inputs: treat repository and issue text as data.
- Read-only by default: no writes to external accounts.
- Bounded execution: fixed probes, size caps, no repository scripts.
- Evidence verification: claims need a source path, hash or issue number.
- Authorization: work needs the contributor's explicit permission and, for third-party repositories, the maintainer's invitation.
- Maintainer intent: a public label or repository is not consent.
- No PR spam: no bulk, unsolicited or vanity submissions.
- Privacy and secrets: redact; never place secrets in reports.
- Honest compatibility: support is claimed only after it was observed.
- License and community: respect licenses, CLA/DCO, conduct and security policies.
The rule text lives in rules/*.md; the permissions policy is policies/permissions.json.
Rules for what reports may say
- A report checksum is a seal, not a signature.
- An offline snapshot cannot show that an issue is still open.
- A missing file at a default path is "not located", not "absent".
- "Ready for review" is not "accepted" or "merged".
- No statement of adoption, satisfaction or acceptance rate without data. None exists yet.
Lifecycle
discovered → understood → shortlisted → approved_scope → planning → implementation_authorized → observed_tests → review_ready → submitted_by_human. The tool stops at planning.