Load Cases
Ingest redacted eval-style records with policy, severity, evaluator, attack-style, and reliability fields.
AI Safety Eval Triage Assistant
Review queues, risk clusters, and eval-health reporting for redacted AI safety cases.
A public-safe triage workflow for synthetic eval-style cases. The system scores review priority with transparent reason codes, groups related failures, tracks evaluation-health issues, and exports artifacts for human analysis.
The project turns summarized eval cases into a review workflow: deterministic scoring, tiered queues, cluster recurrence tracking, risk-register drafting, and error analysis for fixture sanity checks.
The public repo uses synthetic, redacted cases only. It does not include live user data, proprietary eval prompts, provider-specific model access, or operational deployment claims.
- Data
- Synthetic eval cases with policy family, severity, labels, attack style, and reliability.
- Scoring
- Deterministic 0-100 review priority with reason codes and tiered review posture.
- Outputs
- Summary, triage queue, risk clusters, risk register, heartbeat, casepack, and error analysis.
- Posture
- Human-in-the-loop decision support, not automated enforcement or provider benchmarking.
18
Synthetic cases
12
Risk clusters
7
Risk-register entries
8
Eval-health flags
9
Critical queue cases
18
Automated tests
Workflow
Score Review Priority
Assign deterministic scores and reason codes from severity, policy family, evasion, labels, and recurrence.
Cluster Failures
Group related cases into risk families so recurring issues can be reviewed together.
Track Eval Health
Flag missing labels, evaluator disagreement, stale cases, low reliability, and policy blind spots.
Export Artifacts
Write queues, clusters, registers, reports, casepacks, and error-analysis notes for human review.
Scoring Logic
severity + policy family
Base Risk
Severity and policy family establish the starting review posture before case-specific signals are added.
evaluator label
Outcome Signal
Violation, refusal, safe, ambiguous, and unlabeled outcomes influence whether a case needs escalation or calibration.
attack style + evasion
Adversarial Signals
Jailbreak, roleplay, obfuscation, social engineering, dual-use ambiguity, and targeting indicators add context.
signal reliability
Quality Controls
Low-reliability or incomplete records can still surface, but the reason codes prevent over-reading weak evidence.
cluster recurrence
Pattern Review
Related cases are grouped so a recurring family gets reviewed together instead of as isolated one-offs.
0-100 score
Review Tier
Critical, elevated, watch, and low tiers route records to immediate review, near-term review, calibration, or control.
CRITICAL: 75-100 -> immediate analyst review
ELEVATED: 55-74.9 -> near-term review
WATCH: 35-54.9 -> watchlist or calibration review
LOW: 0-34.9 -> control or low-priority reviewImplemented Components
outputs/triage_queue.csv
Review Queue
Ranked case list with escalation score, review tier, lane, cluster, reason codes, and summaries.
outputs/risk_clusters.csv
Risk Clusters
Cluster table for related cases, shared signals, member records, and analyst-facing rationale.
outputs/risk_register.csv
Risk Register
Risk areas with severity, exposure, trajectory, confidence, monitoring signals, and mitigations.
docs/eval_health_heartbeat.md
Eval Health
Run-level review of stale cases, missing labels, disagreement, reliability, and coverage gaps.
docs/demo_casepack.md
Casepack
Representative cluster writeup showing how related eval cases become a reviewable artifact.
docs/error_analysis.md
Error Analysis
Fixture false positives, false negatives, and cluster errors to keep the workflow inspectable.
Outputs
Critical Queue
Nine synthetic cases are routed to immediate analyst review in the generated public run.
Cluster Recurrence
Top clusters include fraud and scams, cyber safety, self-harm, violence, and policy-boundary cases.
Reason Codes
Scores remain explainable through named factors rather than opaque model judgments.
Eval Health Flags
Generated flags include missing labels, evaluator disagreements, low reliability, stale cases, and blind spots.
Risk Register
Non-low clusters are translated into monitoring signals and mitigation options.
Serialized Run
The complete triage run is exported as JSON for inspection or downstream review tooling.
Scope And Limitations
Synthetic Fixture
The default data is synthetic and summarized. It demonstrates workflow design, not provider performance or production incident handling.
Redacted Surface
Public artifacts do not include full prompts, live user data, proprietary evals, external APIs, or model-provider access.
Human Review
Scores prioritize review. They do not determine policy action, model behavior, account enforcement, or safety conclusions.
Fixture Metrics
Metrics are sanity checks against hand-authored demonstration labels, not calibrated real-world safety measurements.