Where the prediction ranges widen

Inspect the six completed generated-data cases from the full-history feasibility test. Verify the completion, summary, result and certificate hashes first. Compare the saved position and lever-arm allowances with their saved total, axis by axis. Keep every original width and failed usefulness result unchanged.

This is a saved-data diagnostic: six identities, one numerical thread, a 120-second cooperative limit, zero native calls, zero bounds and zero fits. There are no retries. It cannot qualify prediction accuracy, observational uncertainty or a scientific detection.

The evaluator lives in the shared analysis library. Preparation, execution, accounting and publication use lll.studies.