Matchproof: match scores that show the CV line behind them

A match score that says "86%" without a reason does not earn trust, and when a model produces the number directly it drifts between runs for the same pair. Worse, a quote shown as evidence may not be in the CV at all, and a missing work-rights requirement can hide behind a high score. Matchproof builds profile-to-job scoring where the model only labels each requirement and code computes the score.

What it does

• Fixed requirement schema: each row has kind (must or nice), category, label (have, partial, missing, unknown), evidence and gap, versioned as RUBRIC_VERSION.
• Extraction once per job: 5 to 12 requirement rows per job, stored by job id and content hash.
• Cheap pre-filter: code drops hard location, work-rights, seniority or salary mismatches before any per-pair LLM call.
• Quoted labels: no quote means missing or unknown, never have. Code checks every quote against the CV text and downgrades failures.
• Score in code: must weighs 3, nice weighs 1; deal-breakers cap the score at 40 and lead the reasons. Unlisted salary is neutral.
• Free and paid views: free shows the score and top 3 reasons including a gap; paid shows every row. Deal-breakers are never paywalled.
• Golden-set gate: 30 to 100 hand-labelled pairs, binary judges validated against your labels, and CI that blocks a regression.

How it works

1. Schema first, mapped to schema.org JobPosting fields.
2. Extract requirements with JSON schema output and Zod validation.
3. Pre-filter, then label each pair with prompt caching on the profile.
4. Verify evidence, then score with a pure, unit-tested function.
5. Cache by profile, job, rubric version and model; batch overnight re-scores.
6. Label a golden set on a local review page, then gate CI on it.

Frameworks & tools

Claude structured outputs, prompt caching and Message Batches API, Zod, and a local review page checked with Playwright.

Why this beats prompting it yourself

Prompting "score this CV against this job" gives a number you cannot explain or reproduce. Here the same labels always give the same score, every shown quote is checked against the CV, and changes to the prompt or model must pass the golden set.

Use cases

• Job apps ranking jobs against a resume.
• Recruiter screening, flipped direction.
• Grant, tender or course fit.

Known limitations

Not for pure semantic search, writing the CV or pulling job data. You need 30 hand-labelled pairs before tuning. Examples use Claude's API features; other providers need equivalents.