| name | job-match-standards |
| description | How real ATS (Workday/Greenhouse/Lever/Taleo/iCIMS/Ashby) parse, score and rank in 2026; evidence-based matching; and the legal limits on automated screening (EU AI Act high-risk, NYC LL144, EEOC Title VII, GDPR Art. 22). Load for changes under commands/match_resume.rs, cover_letter.rs, validate/, documents/embed. |
ATS scoring & job-match standards (reality, not myth)
External best-practices for ATS scoring, JD analysis, and resume↔job matching. Load with author-contract (job-match-author) / token-efficiency (job-match-expert). Pairs with docs/knowledge/matching-algorithm.md (the scoring kernel).
How real ATS work (verified 2026-06)
- No universal "ATS score." Each platform scores differently; a single portable percentage is marketing fiction. Present our number as a guidance estimate with evidence, never as the employer's verdict. https://www.hireflow.net/blog/workday-vs-greenhouse-vs-lever-which-parses-best
- Greenhouse — structured scorecards + Boolean over parsed fields; AI Talent Matching added Feb 2026. Lever — full-text relevance + Gem semantic JD understanding (not exact-keyword). Workday — weights job-title/seniority match heavily (mismatched title tanks the score). Taleo — strict literal keyword match. iCIMS — ML semantic match. Ashby — Boolean search; 0–100 Match Score + reason bullets only via AI add-ons.
- Recruiter Boolean/keyword search is still the dominant filter — candidates surface via search, not just auto-rank.
- AI/LLM screening — ~65% of US enterprise employers use AI-assisted screening (2025); LLM layers now score career-narrative fit + achievement quality. https://incruiter.com/blog/ai-in-recruitment-2026-trends-stats-what-works/
Matching best-practices (what our scorer should do)
- Extract JD requirements and classify hard (must-have/knockout) vs nice-to-have; treat knockout/screening questions as gating, not weighted.
- Normalize keywords + synonyms (title/skill aliases, seniority mapping) — helps both literal (Taleo) and semantic (iCIMS/Lever) parsers.
- Evidence-based scoring — credit skills backed by experience/context, not raw frequency; never reward keyword stuffing (semantic + AI-content detection penalize it). https://www.jobscan.co/blog/can-ats-detect-ai-resume/
- Explainable output — per-requirement match + reason bullets; be honest the number is our estimate.
- Invalidate derived caches on input change — when a posting's text changes (e.g. the full description is resolved on open), drop its cached embedding + any text-hash-keyed score, and invalidate the renderer query that reads that posting. Otherwise the next score reuses the stale snippet embedding and the UI keeps showing the truncated text (#486).
⚠️ 2026 legal / AI constraints on automated screening — flag prominently
- EU AI Act: recruitment AI that sources/scores/ranks/shortlists CVs→JDs is high-risk (Annex III). The legally binding high-risk deadline under Art. 113 is still 2 Aug 2026; a provisional May-2026 "Digital Omnibus" political agreement would defer it to 2 Dec 2027 but is not yet adopted in the Official Journal — until formally enacted, treat 2 Aug 2026 as the binding date and advise preparing for it. Obligations: risk mgmt, human oversight, transparency, logging, conformity assessment. (Prohibited-practices + AI-literacy duties already in force since 2 Feb 2025.) https://www.gibsondunn.com/eu-ai-act-omnibus-agreement-postponed-high-risk-deadlines-and-other-key-changes/
- NYC Local Law 144: automated employment-decision tools need an independent bias audit within the prior 12 months, published, with 10-business-day candidate notice. https://rules.cityofnewyork.us/rule/automated-employment-decision-tools-2/
- EEOC (US): withdrew its 2023 AI guidance (2025-01-27), but Title VII disparate-impact liability still applies (unintentional bias counts); four-fifths/adverse-impact validation + human oversight expected.
- GDPR Art. 22: no decision based solely on automated processing with significant effect — a glance at an AI shortlist is not "meaningful" human involvement; candidates get human review + contest rights + a right to meaningful information. https://gdprinfo.eu/gdpr-article-22-explained-automated-decision-making-profiling-and-your-rights
Myths & mistakes — do NOT encode these
- ❌ "75% of resumes are auto-rejected by ATS" — debunked; traces to a 2012 sales pitch, no primary source. https://jobcannon.io/blog/ai-resume-statistics-2026
- ❌ "One ATS score works everywhere" — vendor logic differs (Workday title-weighted, Lever/iCIMS semantic, Taleo literal).
- ❌ "Keyword stuffing beats the bot" — semantic + AI-detection layers penalize it.
- ❌ "All ATS keyword-match like Taleo" — over-tuning for literal match misleads users.
- ❌ "ATS read everything" — scanned/image PDFs + graphics-heavy layouts break legacy parsers.
- ❌ "Our match % = the employer's decision" — present as a guidance estimate with caveats.
