Locate the terms that make saved prediction ranges wide

Use the same six completed generated long-record cases. Verify their sealed completion, result, final training region, future envelope and bound hashes. Do not load packet histories, original recordings or future observations.

For north, east and height, reconstruct the saved prediction width with exact fractions. Apply the original coordinate scale so every reported width is in metres. Separate the affine support range, position and lever reconstruction allowances, arithmetic allowance and outward rounding. Require these terms to reconstruct the saved width and preserve the model, processing assumption, source binding and original result. Report both the share of the complete width and the share of the unshared allowance; these have different denominators.

Limits: six identities, three arithmetic width checks per identity, one worker, one numerical thread and 120 cooperative seconds. Charge all 18 checks before execution; no retries or allowance expansion. Use the shared study lifecycle and indexed journals. There are no LP proposals, new support solves, interval-map evaluations, trajectory integrations, aircraft reads or optimizer starts.

This is a saved-certificate arithmetic audit. It does not independently validate the supplied likelihood, source intervals, physical assumptions or prediction coverage, and it cannot change the failed usefulness labels. Its purpose is to choose the next diagnostic before repeating an expensive integration.

Earlier allowance audit · Static coupling result · Prediction requirements · Reproducibility.