Ethics & Explainability

Live platform · interactive sandbox.You’re exploring as a guest — every module is visible. Use View as to see the platform tailored to a Chief AI Officer, RISC data scientist, OIG auditor, or leadership. All data is synthetic, and AI features run on Claude in this demo and on your AskSage / Microsoft AI Foundry gateways in production.

Deliverable 4b · Ethics & Integrity Controls

Ethics & Explainability

Bias mitigation, model explainability, and fairness assessment — operationalized as testable procedures and wired into release gates, not stated as principles. Aligned to NIST AI RMF MEASURE and the GAO Performance principle.

Bias mitigation

Pre-, in-, post-processing and in-production controls with a defined fairness schedule.

Explainability

SHAP / LIME attributions on every high-stakes decision, paired with audit trails.

Fairness assessment

Disparate-impact testing against the 4/5ths rule with documented procedures.

Portfolio · ethics & explainability across all 12 systems

Each system carries the fairness lens and explainability method its type demands: disparate-impact testing for models that touch people or firms, source provenance for generative models, saliency maps for vision, and action logs for agentic systems. Parity is the minimum subgroup ratio against the 4/5ths (0.80) rule; a dash marks systems that are not person-affecting.

Contractor Fraud Risk Scorer

UC-001 · Anomaly Detection

0.83Watch

Fairness lens: Vendor size & 8(a)/WOSB/SDVOSB status (proxied by invoice cadence + relationship graph)

Explainability: SHAP on the 0–100 risk score · Invoice-frequency anomaly vs. peer-cohort baseline

Relationship and invoice features can proxy for firm size, over-burdening disadvantaged businesses. Quarterly disparate-impact audit by SBA designation; mandatory analyst review on any score ≥ 72.

Healthcare Provider Fraud Detector

UC-002 · Classification ML

0.88Pass

Fairness lens: Provider specialty, geography, and patient-panel demographics

Explainability: SHAP against peer-cohort baselines · Procedure-code clustering above specialty norm

False positives harm innocent providers. Validated against adjudicated cases at 0.87 precision; peer-cohort baselines refreshed each OPM plan year.

Hotline Complaint Auto-Triage

UC-003 · NLP

0.79Watch

Fairness lens: Language register / dialect and submission channel (phone, web, mail)

Explainability: Token-level attribution (integrated gradients) · Presence of formal complaint vocabulary

Below the 4/5ths line on mail-transcribed and non-standard-English submissions. Hard-coded keyword override escalates serious tips regardless of score; corpus expanded for dialectal coverage.

Mail-Theft Image Classifier

UC-004 · Computer Vision

0.74Watch

Fairness lens: Capture-condition parity (daylight vs. low-light) across facilities

Explainability: Grad-CAM saliency maps over the flagged frame · Motion signature at a collection-box aperture

Night-time precision falls to 61%, exposing innocent employees to scrutiny. Held for low-light augmentation; two independent analysts review every clip before any referral.

Audit Report Summarization Assistant

UC-005 · GenAI / LLM

N/A

Fairness lens: Not person-affecting — integrity, not disparate impact, is the concern

Explainability: Source-grounded citations (each sentence pinned to a workpaper paragraph) · Extractive provenance to the underlying finding

Confabulation (NIST 600-1) is the live risk. Inline citations enable rapid fact-checking; no AI-drafted text enters a final report without auditor attestation.

Congressional Request Intake Classifier

UC-006 · Classification ML

0.94Pass

Fairness lens: Topic and office routing — minimal protected-group exposure

Explainability: Feature importance over request terms · Subject-matter keywords and committee of origin

Misrouting a sensitive request is the main harm. Paralegal confirms every routing; hard-coded override sends oversight-of-investigation requests to Deputy Counsel.

Internal Knowledge Search (AskOIG)

UC-007 · GenAI / LLM (RAG)

N/A

Fairness lens: Not person-affecting — data privacy is the concern

Explainability: Retrieval citations (every answer links its source document) · Top-ranked passages from the permitted corpus

Risk is inadvertent retrieval of restricted material. Index excludes AMBER+ documents; role-based access enforced at query time so the boundary holds even on a misconfiguration.

Audit Data Anomaly Pipeline

UC-008 · Anomaly Detection

0.86Pass

Fairness lens: Employee group and transaction type (payroll, procurement, travel)

Explainability: Per-feature anomaly attribution · Deviation from role-and-period peer baseline

Flags are advisory inputs to audit planning, never findings without independent corroboration. Pipelines exceeding a 40% false-positive rate auto-pause for recalibration.

Cybersecurity Threat Triage Agent

UC-009 · Agentic

N/A

Fairness lens: Not person-affecting — accountability for autonomous action is the concern

Explainability: Immutable action-justification log + decision chain (agentic explainability) · Threat-intel correlation that triggered each step

Explainability here is the audit trail, not feature attribution. Action allowlist caps the agent to read-only queries; a human-approval gate governs anything consequential pending red-team.

Carrier Route Theft Forecaster

UC-010 · Forecasting

0.71Fail

Fairness lens: Race / national origin via ZIP, route, and origin geographic proxies

Explainability: SHAP on the 30-day route forecast · Historical incident density (being replaced with mail-volume-normalized rate)

Below 4/5ths: geographic features proxy for protected characteristics, concentrating enforcement. Held at validation pending feature replacement, demographic-parity testing across census-tract quintiles, and a Counsel civil-liberties review.

FEHB Claims Document Classifier

UC-011 · Classification ML

0.90Pass

Fairness lens: Document type and provider; PHI handling

Explainability: Classification-confidence attribution · Form structure and header signatures

PHI exposure is the risk. Confidence below 0.80 routes to manual review; outputs confined to access-controlled case folders under the OPM data-sharing agreement.

Predictive Workpaper Risk Scorer

UC-012 · Classification ML (retired)

Fail

Fairness lens: Audit-methodology type (penalized novel approaches)

Explainability: Feature importance (which surfaced the bias) · Deviation from historical documentation norms

A feedback loop: it penalized first-of-kind audit methods because they differed from training-data norms, chilling innovation. Rejected and retired — the framework exercising its authority to say no.

Fairness assessment · live

Hotline Complaint Prioritization

Selection rate parity (80% rule)

0.80
Min parity ratio
SubgroupSelectionParityn
Region A61%1.0018,420
Region B57%0.9314,870
Region C53%0.879,340
Region D50%0.826,810
Region E49%0.804,220

After applying class-weight rebalancing and re-calibrating the regional complaint-volume normalization factor in Q4 FY2025, all five regional subgroups now meet or exceed the 4/5ths (0.80) parity threshold. Region E, previously the lowest-performing subgroup at 0.74 pre-mitigation, reached exactly 0.80 post-mitigation and will be monitored monthly for regression.

Explainability · SHAP

Contractor Fraud Risk Score

Feature contributions for a single prediction — red pushes risk up, green pulls it down.

High risk (0.87)

Top drivers (plain language)

  • 1 Invoice totals exceeded the contract ceiling by 31% across the last three billing periods, which is the single strongest predictor of inflated or fabricated charges in this model.
  • 2 Fourteen change orders were submitted in 90 days — more than four times the median for comparable contracts — suggesting scope-creep manipulation or unauthorized work authorization.
  • 3 A single subcontractor accounted for 89% of subcontracted spend, raising concerns about pass-through billing arrangements and lack of competitive sourcing.

Required human sign-off

Reviewed by SA-4412 on 2026-06-02. SHAP output consistent with independent document review. Elevated to full investigation per SOP-INV-07. Case ID: OIG-2026-CF-0883.

Bias-mitigation pipeline

1
Pre-processingRebalanced regional complaint volumes

Resampling, reweighting, and representation augmentation so the training data reflects the population.

2
In-processingClass-weight rebalancing

Fairness constraints in training — demographic parity, equalized odds, calibration across subgroups.

3
Post-processingPer-subgroup threshold calibration

Threshold adjustment and reject-option review where disparities persist after training.

4
In-productionMonthly parity monitoring

Continuous disparate-impact testing with the 4/5ths rule and monthly regression checks.

Release-gate checkpoints

No model reaches production until every gate is green. These checkpoints are wired into the AI release pipeline.

  • Bias & disparate-impact audit passes the 4/5ths rule across all defined subgroups
  • Explainability (SHAP/LIME) available for every high-stakes prediction
  • Model card documents data, intended use, limitations, and TLP classification
  • Human-oversight model defined and override logging configured
  • Adversarial red-team evaluation complete for High-Impact systems
  • OIG Counsel sign-off for any protected-characteristic proxy variable