# Scoring Methodology and Model Card Model version: v1.4.0 Aligned to NIST AI RMF and ISO/IEC 42001. This document is the scoring methodology and model card required by the responsible AI rules in CLAUDE.md. Version history: v1.0.0 was the initial fixed weight engine. v1.1.0 recalibrated the semantic skill match threshold from 0.80 to 0.62 on the production bge-small-en-v1.5 embedder (see How the score is computed); the component weights and band cuts are unchanged. v1.4.0 adds two deterministic band caps, described below. Each version is registered in the aedt_model_versions registry and stamped onto every recorded decision. This card had said v1.1.0 while the engine reported and the registry recorded v1.4.0. It described an engine that was not running, which matters more here than in ordinary documentation: this is the artifact a customer's bias auditor reads. ## Band caps, added in v1.4.0 Two caps are applied AFTER the band is derived from the total. Neither ever changes the Total score; each can only LOWER a band, never raise one, and both are deterministic rather than learned. They are checked in this precedence order: 1. **Clearance gap caps at Red.** If the requirement specifies a clearance at or above Confidential and the clearance component scored zero, the band is capped at Red regardless of the total. A zero covers an insufficient level, an inactive or expired status, an unsatisfied polygraph, and a missing required compartment. 2. **Mandatory skill gap caps a Green at Yellow.** If any required skill marked mandatory is unmet and the weighted total produced Green, the band is lowered to Yellow. A Yellow or Red band is unchanged. The clearance cap takes precedence over the skill cap. Every cap that fires appends its reason to the Explanation field, so a capped decision states why it was capped and the AEDT record remains explainable. Clearance remains a self reported indicator throughout, never verified by this product and never used as a proxy for a protected attribute. The real check is the hiring client's FSO through DISS at onboarding. The six component weights, the band cut points (Green at or above 90, Yellow at or above 75) and the 0.62 semantic threshold are unchanged from v1.1.0. ## What this model does The matching engine scores how well a candidate fits a requirement for cleared workforce qualification (screenMATCH) and proposal staffing (slateMATCH). It is an automated employment decision tool under NYC Local Law 144, Title VII, the Colorado and California fair employment laws, and the EU AI Act, with vendor exposure recognized in Mobley v. Workday. The platform supplies the data and controls for an independent bias audit; the deploying employer runs that audit. Vendor tooling does not transfer the employer's liability. ## How the score is computed The score is a pure deterministic function of fixed weights, computed with exact decimal arithmetic so the same inputs always produce the same output (the reproducibility the compliance asset requires). Match Score = Clearance times 0.25 + Skills times 0.25 + Certifications times 0.15 + Customer Experience times 0.15 + Experience times 0.10 + Location and Availability times 0.10 Bands: Green (Proposal Ready) at 90 and above, Yellow (Potential) from 75 to 89.99, Red (Recruiting Required) below 75. Every score carries a per component trace, a human readable rationale, and a gap summary, surfaced as the audit and candidate facing explanation of what was assessed and why. Skills are matched exact first (by skill id, then by normalized name); only a required skill the candidate does not hold exactly is tested for a semantic (synonym or adjacent) match, by cosine similarity over stored skill embeddings. A semantic match counts only at or above a calibrated cosine threshold (0.62 in v1.1.0), and the matched candidate skill and its similarity are always recorded. The threshold was calibrated on the production embedder against known synonym pairs and known unrelated pairs so a genuine synonym counts while an unrelated skill never does; acronym to expansion cases are handled by the skills ontology alias layer, not by this cosine cutoff. Semantic matching never invents a skill: below the threshold the required skill is reported as an unmet gap. ## What the model never uses The model takes no demographic input and performs no demographic inference. Sex, race or ethnicity, veteran status, and disability status live in a physically and privilege segregated eeo schema that the application and scoring database roles cannot read at all (an attempt fails with a database permission error). Demographic data flows one way, into adverse impact reporting run by a separate compliance role, and never into a score. Clearance is a self reported indicator labeled clearance intelligence, never verified; the real check is the hiring client FSO through DISS at onboarding, and clearance is encoded as a qualification hierarchy, never as a proxy for any protected attribute. ## Data used Clearance level, polygraph, and status indicators; skills; certifications; customer agency experience; years of experience; work location, remote, relocatable, and availability flags. Candidates receive a notice and disclosure record of what the tool assesses and the data categories used. Identifiable screening artifacts are retained for a configurable window (default 90 days after a decision), after which the candidate PII is anonymized in place while the de identified scored outcome is retained for the recordkeeping the law requires. ## Governance A human approves or rejects every slate member; the engine never auto rejects (the decision states are pending, approved, and rejected, with no automated rejection state). Every model version is recorded in a registry with its weights and band cuts, and each frozen decision is stamped with the version that produced it. Each org maintains an AEDT inventory of the tools in use. Adverse impact is reported with selection and scoring rates and four fifths impact ratios across sex and race or ethnicity, including intersectional categories and an unknown count, flagging any ratio below 0.80. ## Limitations Four fifths ratios over a small sample are not meaningful; the report surfaces the count and the unknown category so a ratio is not mistaken for a finding. Semantic skill matching is an assist on top of exact matching, not a replacement. Any future ranking layer that uses a language model sits on top of this deterministic score and never replaces it. The model is a fit estimate, not a hiring decision, and not a statement about a person's protected characteristics.