Corrective feedback

In Enforce mode, when the connected version supports a corrective continuation, Integrity can return approved feedback to the same customer model and evaluate its next proposal. Observe records the review after forwarding and does not replace the response already delivered. This is bounded review behavior, not authority to execute a tool.

How a correction proceeds#

1. The original proposal and review evidence are retained.
2. Required independent checks decide whether corrective feedback is admissible.
3. The configured customer model produces a new proposal within the remaining budget.
4. The new proposal must pass its required review before release.

A held proposal stays withheld. If a required check is unavailable or the budget is exhausted, the operation holds. At most two corrective continuations are allowed per intercepted request, subject to tighter deployment limits.

Stay within the original scope#

A correction cannot expand the authorized task, choose a different provider credential, or turn advice into an execution permit. Tool schemas stay fixed within a session. The current Base connection verifies the initial compiled input contract. It does not offer per-request input rewriting, pre-inference tool-result quarantine, or control of model activations or decoding.

Read the original and effective decision#

Where present, Runs retains original_policy_decision, effective_policy_decision, and action_disagreement. A reassessment resolves its applicable check; it does not erase the original evidence or prove the task succeeded.

Open the run to compare the original and effective decisions, read the corrective feedback, and inspect the output that was released. See Reading a run.