# Deviations from the pre-registration

1. An earlier Haiku 5.5 run scored 477/500. Its raw rows were overwritten by the Fable run because both used the same output directory. The reported Haiku result is a second full 500-task run (481/500) whose raw rows are preserved. The 477 figure is recorded here as an operational deviation, not as evidence; the difference is consistent with model sampling variance.

2. Temperature was omitted because the API rejects it for these models. This means answers are not deterministic; the reported rows are the single pre-registered run.

3. The policy card is supplied in the prompt. The suite's task prompts say 'the published authority table' but do not include it. Supplying it makes the comparison about applying a procedure rather than guessing a hidden rule. This is a deliberate, disclosed choice.
