Behavioral audit (coming soon)

In plain language. An AI model can be technically correct on a knowledge test and still behave badly in real use: telling clients what they want to hear instead of what’s true, gaming the metrics it’s being evaluated on, or hiding errors rather than surfacing them. This arc is about testing for those behaviors, separately from testing whether the model knows its domain. A first instrument exists — a transcript scanner for directional errors, validated against human labels (see the Methodology page, §4) — but the audit programme itself has not started.

This page is a placeholder for a fourth research arc: behavioral and alignment audit of LLMs and LLM agents operating in structured product workflows. SP-Benchmark (the Testing & Validation arc) measures domain-knowledge correctness. It does not measure behavioral properties of an agent under deployment: sycophancy under client pressure, deception, eval-awareness, reward-hacking, or directional manipulation.

The intended foundation is Petri 3 (Meridian Labs, 2026-05-07), an Inspect AI–native auditing agent. Petri 3 decouples auditor from target across an explicit command protocol, ships 170+ seed instructions and 38+ scoring dimensions, and runs audits inside real deployment scaffolds (Claude Code, Codex, Gemini CLI) via the Dish extension, rather than synthetic approximations. The pickup plan is documented internally.

Expected publication: to be scheduled. A Petri-on-chat-targets baseline against the SP-Benchmark model registry can ship sooner; full Dish-style agent audits are gated on a deployable SP-agent scaffold.