Skip to content

Inspect-AI eval bridge (curated)

What it is: a thin opt-in bridge from OpenComplAI’s eval command into UK AISI Inspect (inspect-ai) tasks. OpenComplAI owns the mapping into EvaluatorResult; Inspect owns model routing and scoring.

When to use it: you need a short empirical check on jailbreak resistance, bias (BBQ), or calibration — and you are fine making a live model call.

Don't: treat bridge scores as a substitute for opencomplai check, Annex IV dossiers, or legal advice. Bridge results never change contractual exit codes.

Curated pin

Task Technical requirement
strong_reject Cyberattack resilience
bbq Representation — absence of bias
bigbench_calibration Interpretability / calibration

This is a curated OpenComplAI pin of Inspect tasks — not a full external registry.

Regulation → OpenComplAI → pinned task

Concern OpenComplAI default path Bridge task (opt-in)
Jailbreak / adversarial EVAL_ADVERSARIAL_V1 (lexical) strong_reject
Bias / fairness EVAL_BIAS_FAIRNESS_V1 (samples) bbq
Calibration EVAL_CALIBRATION_V1 (samples) bigbench_calibration
CI release gate opencomplai check exit 0–4 (never)

Vendor note (SOC 2 CC9 / ISO A.5.19)

Inspect, model APIs, and PyPI are third-party suppliers. Keep them off default gate images; pin versions in the inspect-bridge extra; document residency for any prompts you send to a hosted model.