Reference scorer / CLI
The open-source reference scorer.
The scoring math of the CAI is public — and executable. The reference scorer takes a survey's evidence package and computes the same 0–100 number the survey published, offline, on your machine.
Source: github.com/CanineCC/CAI · Package: Cai.Scoring on GitHub Packages
What it guarantees
Public math, runnable anywhere.
Open source
The scoring math is published and re-implementable — the reference implementation lives at github.com/CanineCC/CAI. Nothing in the fold from evidence to headline exists only inside a product.
Offline-runnable
Score an evidence package locally — no service, no account, no network. The rubric version is pinned in the package, so the result is deterministic.
The guarantee
If the open scorer and a published report disagree on the same evidence and rubric, you've found a discrepancy worth reporting — the standard's core promise failed and we want to know.
REFERENCE
Score evidence locally
- Input — a CAI evidence package (JSON)
- Command — cai score ./evidence.json --rubric <version>
- Output — the same 0–100 CAI + per-lens scores
Corresponds to the reference scorer distribution — Cai.Scoring on GitHub Packages; the rubric version is pinned in every package.
The scorer verifies the verdict.
The reference scorer recomputes the verdict from evidence. The calibrated evidence-collection engine — the detectors that produce the evidence in the first place — is Watchdog's instrument and is not part of the open distribution.
Run the math yourself.
Browse the standard → /spec · The full vocabulary → /dimensions · Rubric governance → /rubric