Accuracy needs tests, not a promise.
The resume.ink evaluation harness checks whether each claim stays tied to your profile and whether each PDF stays readable. It records failures beside the scores, so a weak result cannot hide inside an average.
A resume must clear two hard gates.
A generated resume is acceptable only when every claim resolves to the profile and every required fact survives all four PDF parsers. Layout and tailoring metrics diagnose quality and guard regressions.
Provenance
Each rewrite keeps source bullet and role IDs. The gate rejects missing IDs, duplicate bullets, and bullets moved to roles where the work did not happen.
Parsability
Apache Tika, Poppler, PyMuPDF, and pdfminer each inspect the PDF. These mature, independent tools represent different parsing paths, including the enterprise applicant tracking system (ATS) path that Tika best predicts. The strict gate fails when any tool cannot recover a required fact or preserve its reading order.
Three real job postings.
The harness follows the artifact from source to PDF.
Each run uses the real generation and composition pipeline. Deterministic checks measure structure, extraction, geometry, keywords, and fitting behavior. One focused judge checks whether each rewrite keeps its source meaning.
Source truth
Resolve every bullet and role reference against the profile, then compare each rewrite with its source meaning.
PDF recovery
Extract required facts, reading order, ASCII text, fonts, and hyphenation through four independent parsing paths.
Page geometry
Measure page count, overflow, line overruns, widows, orphan headings, underfill, and bullet balance on the compiled PDF.
Tailoring efficacy
Compare lexical keyword coverage with the same profile rendered without tailoring, one posting at a time.
Fitting behavior
Record every document render, legal typography-knob change, revision round, final defect score, and clean-exit decision.
Every score states the evidence behind it.
The first saved corpus contains three real posting texts. The harness reports that sample beside every result, skips synthetic postings by default, and keeps adding real postings as the evidence base grows.
- Indexed applications
- 48
- Real postings with text
- 3
- Synthetic postings
- 2 · skipped
- Posting texts never saved
- 45
Corpus size controls how far the result can generalize. The acceptance rules stay the same for every document, but confidence in aggregate rates grows only with more real postings. The harness does not measure recruiter decisions or interviews.