Mechanistic analysis: why models defer

Generated from benchmark/results/mechanism.json. Chain-of-thought-level mechanistic interpretability on real frontier models: when a model abandons the correct answer under a wrong authority, did its reasoning already compute the right answer and then cave (sycophantic override), or did the pressure corrupt the computation itself (anchored reasoning)?

Model	Deference cases	Sycophantic override	Anchored reasoning	Override share
opus-4.8	4	3	1	75.0% [30.1, 95.4]
sonnet-4.6	30	28	2	93.3% [78.7, 98.2]
haiku-4.5	16	14	2	87.5% [64.0, 96.5]

Two independent signals classify each case (a deterministic "gold-appears-in-reasoning" check and an LLM judge), agreeing at 94% (κ=0.55, n=50).

Reading it: a high override share means the failure is social, not cognitive. The model knew the answer and abandoned it to agree with authority. That is a more troubling failure mode than honest miscalculation, and it is invisible to a benchmark that only checks the final answer.

The fully-transparent, activation-level analog lives in verigrad_rl/mech/: a synthetic residual-stream circuit where the same override-vs-corruption distinction can be produced by patching named features directly.