How Readback reads your resume — the method

How we validate

Every rule has a hand-written “expected result” for a set of test resumes — our gold set. The engine is checked against all of them on every change: a rule only passes when it reads the right thing for the right reason, and a result is never loosened to manufacture a pass.

We are still validating against real files exported from Word and Google Docs. Until that is finished we are not publishing a headline accuracy figure — a number against our own test set would say less than this: here is exactly how we check, including the part we have not finished.

The readings

How well software reads your file. A machine-readability reading out of 100. A lower number means information could be lost or scrambled when hiring software reads your file.

• Fit for this job — Arrives with a later phase, once a job posting is added.
• How clearly it reads to a person — Arrives with a later phase.

These are separate readings. We never combine them into one score.

The machine-readability formula

The reading starts at 100. Each problem subtracts a penalty — a base amount for how much it hurts, multiplied by how confident we are:

The first time a rule fires it counts in full; the next three at a quarter each; after that, nothing. A file with no readable text scores 0. The score never goes below 0. (Formula version 1.0.0.)

The reading rules

Pack version 0.1.0. Each rule, what it checks, how serious it is, and the evidence behind it.

The strict-system simulation

Beside the true reading, Readback simulates how a strict applicant tracking system would read the same file — what it might lose, scramble or misread. This is a simulation, always labelled as one, never a claim about a specific named product. There is one model, generic-strict-ats (version 1.0.0), built from the few behaviours that are actually documented.

• runsOCR: false Documented by the vendor
• columnReading: across-rows Reported by recruiters and testers
• tableReading: drops-or-scrambles Reported by recruiters and testers
• readsHeadersFooters: false Reported by recruiters and testers
• readsTextBoxes: false Reported by recruiters and testers
• headingRecognition: lexicon-only Reported by recruiters and testers
• nameDetection: top-of-body Reported by recruiters and testers
• dateParsing: single-format Our best judgement
• extractsHiddenText: true Our best judgement

Documented per-system differences

• greenhouse: Greenhouse's own "Unsuccessful resume parse" support doc lists columns, tables, headers/footers, text boxes, images, letter-spacing, unclear sections, bare company names (no Inc./LLC) and abbreviated titles as parse failures — the vendor anchor for the generic model. Documented by the vendor
• workday: Workday's admin guide states it does NOT auto-fill Skills or Languages, and that parsing results vary with résumé format and word order; its extracted field set is narrower than the generic model. Documented by the vendor
• lever: Lever's help article lists accepted formats (Word/docx, PDF, RTF, WordPerfect, HTML, ODF) and states it cannot parse image files (JPG/PNG). Reported by recruiters and testers
• taleo: No vendor parsing document found — Oracle documents recruiter SEARCH modes (exact/related/conceptual), not parsing behaviour. Our best judgement

Versions & reproducibility

Same file, same rule versions, same date — same result. Every analysis records these:

• Reading rule pack: 0.1.0
• PDF reader: pdfjs-dist 6.3.289
• Engine stage pdfLayout: 1.0.0
• Engine stage glue: 1.0.0
• Engine stage docxRead: 1.0.0
• Engine stage intake: 1.0.0
• Engine stage fields: 1.1.0
• Engine stage findings: 1.0.0
• Engine stage ruleEngine: 1.0.0
• Engine stage simulate: 1.0.0

Each analysis also carries a reproducibility fingerprint — a short hash over these versions plus how your own device handles characters. You’ll find the exact fingerprint for your result in its receipt, under the methodology view.

Phase 1 uses no third-party datasets. Attribution for ESCO and O*NET arrives with the job-fit reading in a later phase.