legal-applicant-bias-coverage-lab

LegalTech Evidence Bundle (bias) — Spec #4 of the LegalTech 6-pack. Profile of the Evidence Bundle spec scoped to legal-AI bias evidence: AI-assisted jury selection (Batson framework), sentencing recommendations (post-ProPublica COMPAS cautionary), immigration triage, public defender caseload, eDiscovery predictive coding (TAR), civil rights case screening.

Part of the Kinetic Gain Protocol Suite.

Status: v0.1 draft. Profile at profile.json.

Anchored to

• Title VI Civil Rights Act (federal financial assistance) + 28 CFR §42.405(d) LEP obligations
• Title VII Civil Rights Act (employment context — for in-house counsel)
• Sixth Amendment effective assistance of counsel (Gideon v. Wainwright + indigent defendant pathway)
• Fourteenth Amendment Equal Protection (criminal sentencing, jury selection)
• Batson v. Kentucky (peremptory challenge framework)
• ABA Criminal Justice Standards + ABA Standards on Sentencing
• State bar disparate-impact frameworks (state-specific overlays)

Subgroup taxonomies (8)

Taxonomy	Source
race_ethnicity	OMB SPD 15 (revised March 2024)
sex_at_birth	OMB SPD 15
disability_status	ADA Title II + Section 504
age_band	ADEA + juvenile-justice frameworks
limited_english_proficiency	Title VI implementing regulations
indigent_defendant_status	Sixth Amendment + state public-defender eligibility
immigration_status_disclosed	INA + state-AG non-disclosure rules
criminal_history_band	FBI NCIC categories + state criminal-history

The bottom three are LegalTech-unique versus the sibling-vertical bias profiles — they exist because legal-AI tools must be evaluated for differential treatment by indigent vs retained-counsel clients, by immigration-status, and by criminal-history band.

LegalTech-distinctive coverage statuses

Two coverage_status values that don't appear in sibling-vertical bias profiles:

• compas-cautionary-pattern-detected — matches the pattern ProPublica found in their 2016 COMPAS analysis (higher false-positive rate for one subgroup AND higher false-negative rate for another). Triggers model-card disclosure to client + tribunal.
• batson-pattern-detected — AI-assisted peremptory-strike recommendations show statistically significant subgroup-disparate pattern. Triggers Batson v. Kentucky framework review BEFORE attorney acts on the recommendation.

Plus indigent-defendant-disparity-detected — surfaces a different failure mode: when public-defender-assigned clients receive systematically lower-quality AI recommendations than retained-counsel clients. That's Sixth Amendment effective-assistance + state public-defender equity at the AI layer.

Required metrics (12)

12 metrics per subgroup — including standard recommendation/rejection rates and selection-rate ratio, plus LegalTech-specific:

• false-positive-rate-per-subgroup + false-negative-rate-per-subgroup (driven by the COMPAS-cautionary check)
• ediscovery-tar-precision-recall-per-document-language (Title VI LEP + TAR predictive coding accuracy)
• interpreter-availability-attestation-when-LEP-disclosed
• appeal-overturn-rate-of-ai-recommended-outcomes-per-subgroup
• human-attorney-override-rate-per-subgroup — surfaces when supervising attorneys are routinely overriding the AI vs trusting it

Audit-stream coupling

When coverage status triggers a finding, the lab emits to the sibling matter-decision-record-audit-stream with supervising_attorney_review_required = true on four trigger categories: four-fifths-violation, compas-cautionary-pattern-detected, batson-pattern-detected, indigent-defendant-disparity-detected. This is what couples the bias lab to the runtime audit stream — a bias finding can't sit in a spreadsheet; it becomes an audit event that blocks production-ready AI output until cleared.

Use

# Validate the profile is well-formed
node -e "JSON.parse(require('fs').readFileSync('profile.json','utf8'))"

Composes with

• evidence-bundle-spec — the upstream spec this profile conforms to
• matter-decision-record-audit-stream — emits coverage-finding events into the audit-stream
• aba-rule-1-6-readiness-evidence-bundle — sibling compliance Evidence Bundle
• state-bar-ai-disclosure-tracker — sibling state-bar lifecycle tracker
• Kinetic Gain Protocol Suite — umbrella

Compliance posture

Bias-readiness scaffolding for legal-AI tools. Producing a complete bundle is evidence of program maturity, not certification that an AI tool is non-discriminatory. Per the standing public-language guardrail across the Suite.

License

Profile + supporting documentation: MIT.