Model Risk Register
| Title: Model Risk Register |
Version: 1.0 |
Owner: [TENANT_CONFIGURATION_REQUIRED — AI Governance Lead] |
Status: Draft (living document) |
Last reviewed: 2026-09-07 |
Next review: [TENANT_CONFIGURATION_REQUIRED] |
Reviewers: AI Governance, Security, Legal |
Purpose and scope
Tracks each model/skill combination in production, its risk tier, known limitations, and mitigations — the AI-governance equivalent of the security threat register.
Register
| Skill |
Model tier (model_version.risk_tier) |
Known limitations |
Mitigations |
Owner |
| CV Extraction |
Standard |
May mis-extract poorly formatted/scanned CVs |
Confidence threshold + manual review queue |
[TENANT_CONFIGURATION_REQUIRED] |
| Candidate Matching |
Elevated (influences human decisions at scale) |
Potential proxy bias if JD criteria are loosely worded |
Fairness testing (fairness-and-bias-governance.md), mandatory human approval |
[TENANT_CONFIGURATION_REQUIRED] |
| Offer Drafting |
Standard |
Template misapplication if wrong template version selected |
Template version validation, mandatory human approval before send |
[TENANT_CONFIGURATION_REQUIRED] |
| Document Verification Assist |
Elevated (touches identity documents) |
OCR errors on low-quality scans |
Confidence threshold, needs_review fallback, never auto-pass |
[TENANT_CONFIGURATION_REQUIRED] |
| Policy Q&A (RAG) |
Standard |
Hallucination risk if grounding fails silently |
Citation enforcement, no-answer fallback |
[TENANT_CONFIGURATION_REQUIRED] |
Risk tier definitions
| Tier |
Definition |
Review cadence |
| Standard |
Low potential impact if wrong; always reviewed by a human before consequential action |
Quarterly |
| Elevated |
Higher potential impact (identity, fairness-sensitive); stricter thresholds and mandatory sampling |
Monthly |
| Restricted |
Not currently used in this platform — reserved for any future fully-autonomous capability, which would require a new ADR and explicit organizational sign-off |
N/A |
Review process
Each entry is reviewed on its cadence for: continued accuracy of “known limitations,” whether mitigations remain effective, and whether production incidents (see incident-response-runbook.md) or evaluation drift (see ai-evaluation-strategy.md) require a tier change.
Change control
| Version |
Date |
Author |
Change |
| 1.0 |
2026-09-07 |
Documentation package generation |
Initial creation |