Projects

Asks what changes in an agent's behaviour and in its errors when one model family grows from 0.8B to 9B parameters and nothing else changes. Evaluates four Qwen3.5 checkpoints, each with thinking on and off on the same weights, across the three τ²-bench domains — 6,672 conversations in all. Two behavioural frameworks are joined turn by turn: the ACT-onomy action taxonomy applied by an LLM coder, and the τ²-bench error reviewer run with two judge models from different families. The mix of behaviours holds at the level of ten action groups but shifts across 46 sub-actions, and the kind of error shifts more than the amount: comprehension errors fall by an order of magnitude while guideline violations do not, appearing in six of ten of the 9B's successful conversations.

Opens CT-CHAT's decoder with linear probing, attention analysis, and activation steering, all at the single position where the model commits to its first report token. 78.6% of the findings present in its evaluation set go unreported (8,176 of 10,402), yet the evidence for most of them stays linearly decodable there from layer 1 through layer 32, over and above patient metadata and scan protocol — the model has the evidence and does not report it. Within volumes where a finding is present, the attention difference between reports that name it and reports that miss it correlates negatively with recall, and steering lifts lung nodule from 0 to 49.6% against 0.4% for a matched random direction while consolidation stays below 1%.