Agent view — served from /registry.md. Structured as a flat, chunkable list so a retrieval system can pull any single dimension without losing its meaning.
The registry
Every check we run,
and how it is scored.
This is the technical reference. If you want the plain-language version of what these checks mean for your business, start with how the audit works instead.
Fifty-five dimensions across four layers. Tier 1 checks are deterministic and produce the headline score. Tier 2 checks are LLM-assisted and produce advisory flags with evidence — they never contribute points, because a score an LLM can move is not a score you can track.
No dimensions match that combination.
Scoring policy
If an LLM can move the number, the number proves nothing.
- Only Tier 1 contributes to the headline score
- deterministic
- Scorer is a pure function — no network, no LLM, no randomness
- unit-tested
- Rubric is versioned; every run records its version
- v1.4
- Re-audits reuse the original rubric by default
- comparable
- Tier 2 calls run at temperature 0 with pinned model versions
- recorded
- Every lost point stores the URL and test that failed
- evidence
Honest limitation: even at temperature 0, provider-side model updates change Tier 2 output over time. We pin the model, record the version string with every flag, and note any upgrade in your report rather than silently changing the analysis.