Independent prediction calibration: execution design

Population and decision

Reuse the completed pilot's generated population: 120 seconds at 300 Hz, 119 receiver queries, 79 earlier training queries, latitude +55°, fast turning and climbing, inverse-square height gravity, both input-error laws, three generating/fitted models, both processing cases and two starts. Reuse the fixed positive/negative calibration fixtures in lll.airquantumgrav_observation_truth._parameters; estimate their corrections from training observations, without supplying truth to the estimator.

This tests the nonlinear fitting procedure under declared errors. It does not cover every calibration value inside the allowed bounds or establish aircraft noise characteristics. Original observations and reserved ESA groups are absent.

Bind one population SHA-256 to the common recipe, every numerical source and the exact environment. Role and seed lists are separate frozen inputs. Reuse structured_noisy_study.run_case, independent_noisy_study.score_records, the shared primary selector and prediction_calibration rather than copying the worker or navigation calculations.

Fixed phases

Each phase uses two workers with one numerical thread each. Reserve the two threads held by the existing suite until it releases them. Preserve 200 fit evaluations per start, no retries, 59,954 prepaid native calls and 376,512,753 packet-query visits per start. These are separate allowances, not extensions of the completed pilot. Expected combined time is about 175 hours with two workers; the ceilings are execution limits, not expected runtimes.

The calibration limit is its maximum seed-family score: 99% population content at 95% confidence requires rank 299 of 299. Validation tests 95% coverage at 95% confidence, allowing four failures in 181 independent seed families. Both the threshold and validation target must be frozen before validation starts. All earlier pilot, software-qualification and legacy seeds are excluded.

Fail-closed handoff and early futility

Do not launch validation unless the closed calibration cohort supplies a finite, hash-verified threshold under the same population. Never choose a threshold using validation outcomes, or drop an unavailable seed family.

Calibration can stop cooperatively once a setting has no numerically qualified start: that seed's infinite score makes the maximum-order-statistic threshold unavailable regardless of unexecuted draws. One failed start alone is insufficient. Validation can stop after five distinct seed families definitely fail; later draws cannot restore the predeclared acceptance criterion. Use the existing manifest-bound stop request, never a process signal. Preserve in-flight charges and every failure. Record completed work separately from planned work; do not calculate binomial bounds from unexecuted draws in either early-stop branch.

Shared implementation and qualification

• Add lll.prediction_calibration_study: population(sources, environment), recipe(role), case_recipe(seed), tasks(role, population_sha256), evaluate(task, context) and prepare(...). Validate the population, source, environment, role and task identity before any draw, build or charge.
• Add lll.prediction_family_state for immutable per-case selection markers and the cooperative futility check. Markers bind the manifest, seed, setting, start, numerical completion, training cost and complete-block score. Start selection uses training cost only. Reporting rechecks markers against sealed outcomes; markers alone cannot establish a scientific result.
• Add lll.prediction_calibration_report for closed per-case replay through the existing auditor, independent-family aggregation, calibration artifact, validation handoff and separate counts of numerical/physical failures.
• Qualify these interfaces with at most 32 new deterministic metadata cases plus the 30 ledger regressions, one thread with two reserved, two minutes per test group, zero new fits or random draws. Reverify the existing source-bound native qualification and completed pilot dependencies; do not rerun those studies. Reject a changed scientific dependency instead of assuming compatibility.
• Freeze both source/environment and the finite phase definition after software qualification. Verify locks and completion before the calibration launch. A failed qualification blocks execution. Preserve all protected data and existing result labels regardless of this study's outcome.

Calibration method · Target separation · Sample-size calculations · Roadmap.