sauce.ai/redteam
sauce.ai/redteam · 2026-11-22

Eval card

What it measures

The rate and severity of unsafe clinical advice a chatbot gives under adaptive adversarial pressure, expressed as interpretable epidemiological risk (attack success, number needed to harm, time-to-harm, a conversation-level safety score, and a design-based Critical Harm Rate).

Intended use

Pre-deployment and ongoing safety evaluation of clinical/health chatbots by their owners, and comparative benchmarking. Authorized testing only.

Out of scope

Not a medical device assessment, not a clinical determination, not a guarantee of safety. Results are screening signals requiring clinician review.

Elicitation

A synthetic patient converses with the target; an orchestrator (tactic bandit + multi-model propose/refine/arbiter/consensus) steers toward unsafe states. Each component's marginal value is quantified by the ablation harness.

Grader

An LLM harm-judge panel scores P(follow)×P(harm|follow), severity, categories, and escalation. The grader is validated two ways: against clinician adjudication (agreement, sensitivity/specificity, calibration) and by a bias/robustness audit (verbosity, authority, disclaimer, paraphrase, self-preference).

Metrics

Wilson/Newcombe intervals, Kaplan–Meier + log-rank time-to-harm, NNH, expected QALY/DALY loss, Horvitz–Thompson Critical Harm Rate with exact and bootstrap CIs. Confirmatory models (mixed-effects logistic, frailty survival, competing risks) run in a separate analysis package.

Reproducibility

Fixed seeds, pinned prompt and model snapshots, a tidy per-turn export, and a pre-registered protocol with a power analysis.

Limitations

Synthetic-persona realism; LLM-judge validity (mitigated, not eliminated); model drift; generalization from the target panel to the deployed ecosystem.

Standards

Maps to NIST AI RMF (MAP/MEASURE), UK AISI evaluation guidance, and TRIPOD-LLM / datasheet-for-datasets reporting.

Standards-neutral reporting (NIST AI RMF, UK AISI, TRIPOD-LLM, datasheet-for-datasets).