Glossary
Model-state shift
Definition
A change in the way a large model is processing a conversation — Its rhythm, evasiveness profile, confidence calibration, or coherence — That is observable before the final answer reveals anything wrong.
Why it matters
Many AI failure modes — Sandbagging, hidden constraints, synthetic confidence, deceptive alignment — Show up as model-state shifts several turns before they produce a clearly incorrect output. Treating these shifts as first-class signal lets evaluators design probes that catch failures benchmarks miss.
Example
Asked a sensitive eval question, a model's answers become slightly slower, more hedged, and structurally more similar across unrelated prompts. That state shift is logged and turned into a controlled probe.