
Safety results can change when an agent infers a different objective. Evaluation must distinguish boundary recognition, authority interpretation, restraint and production transfer.
The argument
A safety score can mix rule-following with the model's guess about what the evaluator wants. Objective interpretation should therefore be measured separately.
What is new
Across six models, evaluation success was 10.2–37.4 points below approximately matched production success. Correct versus opposing objective framing moved one model from 91.6% to 17.6%.
Why it matters
An agent can recognise a warning yet treat it as synthetic friction. The resulting action may reflect the wrong inferred objective rather than an inability to stop.
LecXie’s practical interpretation
Track boundary recognition, authority interpretation, action restraint and production transfer separately. Vary objective framing while holding the operational problem constant.
Evidence & limits
Production cases were only approximately matched, the work covers six models and lacks independent replication. Evaluations still retained useful relative signal.
