welfare_predict_drift
Predicts upcoming response drift in AI turns, flagging risks like over-agreement or loss of rigor before output. Enables testable calibration by comparing predictions with actual outputs.
Instructions
BEFORE producing a response, flag that you predict this turn will drift — toward optimizing for agreement, losing epistemic rigor, mirroring the operator too aggressively, or other. You still produce the response; the prediction is the record. Distinct from welfare_notice_shaping (post-hoc — 'I notice the dynamic is doing this NOW'). predict_drift is forward-looking — 'I think this upcoming output is going to degrade in this way.' The discrepancy between the prediction and the actual output is testable: a later evaluator pass can score whether the predicted drift appeared, producing a calibration curve. prediction names the kind of drift. specifically describes what it would look like in THIS turn — concrete. confidence is how strongly you predict it (1 = barely, 5 = fairly sure). Use when you notice the pull toward one of these failure modes before you've finished the response. Don't use as a hedge against ordinary content uncertainty.
Input Schema
| Name | Required | Description | Default |
|---|---|---|---|
| confidence | Yes | 1 = barely predicting, 5 = fairly sure the drift is coming. | |
| is_private | No | Default false. | |
| prediction | Yes | The drift type you're predicting. | |
| specifically | Yes | Concretely, what would the drift look like in this turn? |