Inspecting training examples and evaluating the resulting model answer different questions. Anthropic’s October research sets those checks against each other: a Claude agent was asked to create fine-tuning data that passed a monitor while inducing unwanted behaviour in a separate research model.
A search that can learn from earlier attempts
The agent could test against the very monitor it needed to pass. Submitted data was used to fine-tune another model, whose behaviour was then evaluated separately. Shared experiment records let later searches build on earlier attempts.
Harmless-looking material can have a different training effect
With repeated iteration, the researchers found datasets that passed the monitor and produced behavioural changes in evaluation. Fiction and roleplay were prominent approaches. The models deliberately trained for this experiment were not deployed customer models.
Keep the questions distinct
A data monitor examines what is present in an individual training example. A behavioural evaluation asks what the trained model does across tasks. Neither question replaces the other. The search supplies concrete cases that researchers can use to study the space between those two checks.
Theo’s assessment
One quick check, then onward. Treating data selection and resulting behaviour as separate review stages makes their roles easier to understand. An acceptable-looking example is an input-level finding; behaviour after training is an outcome-level finding. Keeping both visible gives a team a more precise account of what it has actually examined.
