Baselines and CI

A baseline is a JSON snapshot of the findings of one tree. judge compares later trees with it without reading Git history, so it works the same on a laptop, in CI and on an unpacked archive.

Save a baseline

cargo judge --save-baseline
# baseline saved: .judge/baseline.json (N findings)

Compare

cargo judge compare .judge/baseline.json

The report lists unchanged, resolved and introduced findings and a verdict. Introduced advisory findings are listed separately and have no verdict effect. The baseline compare example shows a real run: judge 0.7.0 compared with a baseline saved right after its own source reorganisation.

For a pull-request comment, render the delta as Markdown:

cargo judge compare .judge/baseline.json --format markdown

The verdict

The verdict fails when the comparison introduces gating findings with warning or failure severity. Heuristic (advisory) findings never fail it.

Exit code	Meaning
0	Verdict passes.
1	Verdict fails on introduced gating findings.
2	Error: unreadable or unsupported baseline, broken judge.toml, analyzer failure. No verdict is given on incomplete analysis.

In CI

Commit the baseline (or restore it as a CI artifact) and let a job compare against it. A GitHub Actions step could look like this:

- name: judge
  run: |
    cargo install cargo-judge --locked
    cargo judge compare .judge/baseline.json

Refresh the baseline deliberately, with cargo judge --save-baseline, when you accept the new state.

Health score trends

cargo judge health --score prints a score from 0 to 100 and a letter grade (A ≥ 90, B ≥ 80, C ≥ 70, D ≥ 60, F below). Deductions are severity-weighted and normalised by authored-LOC density; advisory findings are not scored.

• The score is a configurable trend index, not an objective quality ranking. The delta against a baseline is the message, not the absolute number.
• A trend is shown only with --baseline PATH, and only when the baseline used the same score formula version and the same crate profiles. Anything else is reported as not comparable.
• Without a basis for a score, for example no authored lines of code, the score is reported as unavailable and judge exits with 2 — never a fake perfect score.

Per-crate weighting profiles are opt-in in judge.toml:

[[crate_profile]]
name = "parsers"
crates = ["my-parser"]
deduction_multiplier = 0.5