I build small, deterministic review artefacts. Each one starts from a shortcut verdict — the output passed, the longer answer wins, the code runs — and shows, with a reproducible matrix, exactly where that shortcut fails. The method is the same across all of them: score 0–2 on named dimensions, derive the verdict from the scores, publish the evidence, and state the limits.

Target role: GoMeasure — Prompt Quality Reviewer https://www.gomeasure.ai/talent-network/roles/prompt-quality-reviewer

Every artefact below is live, has a machine-readable matrix.json, and reports its own commit as buildId, so the claim can be checked rather than trusted.

What they share:

• Derived verdicts. The verdict is computed from the dimension scores, not asserted. A dimension at 0 means defective/misleading, so the defect class is mechanical rather than editorial.
• A discrimination matrix. Each case pairs a tempting shortcut verdict with the measured verdict; the "sound" flag is true only when the rubric rejects the shortcut on every case where it is wrong.
• Evidence with provenance. matrix.json is committed to the repo, regenerated in CI, and carries the commit as buildId; the deploy fails if the committed matrix is stale.
• Stated limits. Four cases per rubric, authored meta-scores, and non-exhaustive hidden cases are named in each README rather than hidden.

• Review prompts, constraints, rubrics and criteria — gomeasure-rubric-reviewer reviews the rubric, not the output: five dimensions (measurability, coverage, discrimination, specificity, calibration), each cited to HELM, Diataxis, Concrete Problems in AI Safety, NIST AI RMF, and Cohen's kappa.
• Apply clear, consistent rubrics; explain with concrete evidence — every case carries the evidence for the measured verdict, and gomeasure names the exact replacement criterion in its proposed fix.
• Surface errors, risks and edge cases — edge-case-solution-reviewer separates visible tests passed from hidden constraint cases passed; coding-agent-interaction-reviewer catches the plausible fix that would break in production (e.g. trust proxy: true).
• Recommend improvements — gomeasure emits a concrete replace/add fix per defective rubric; micro1 scores against the posting's own scope (clarity, tone, helpfulness, adherence, substance).
• Calibrate with other reviewers — gomeasure derives calibration from two independent reviewers (0 disagreements → 2, every criterion disputed → 0); micro1 is pairwise by construction.

The index that ties these together: swipswaps-portfolio-showcase · source.

It is a 3D repository index, and it demonstrates the same capability discipline as the instruments: the page picks its API base from the host (github.io → the local HTTPS proxy, localhost → local HTTP), probes /api/health, shows Server: Online / Offline (fallback), and falls back to the GitHub API when no local backend is present. Detect the environment, expose the state, gate the strong path, label the fallback.

Application letters are tailored per employer from these instruments and their live matrix.json; they are not published here.

git clone https://github.com/swipswaps/<artefact>
cd <artefact>
npm ci && npm run review      # prints the discrimination matrix and "sound: yes"
npm run build && npx playwright test

• Each reviewer demonstrates a method on four cases; it is not a benchmark.
• The non-calibration meta-scores are authored, not measured from live raters.
• Hidden cases are not exhaustive; edge-case-solution-reviewer now runs candidates in a bounded child process, but node:vm is still not a security boundary (documented in that repo).
• The review artefacts are English-only.

Source for every claim: the linked repo README and the live matrix.json. If a claim here is not backed by one of those, treat it as a bug in this page.