Fresh prediction calibration with qualified timing checks

This defines a separate finite release. The 48-case optimizer qualification passes all fits, 24 paired-start comparisons and four original input replays. Both earlier calibration campaigns and their unused allowances remain closed.

Question

Can the qualified training-and-prediction method produce a finite error threshold on new generated examples, then meet its fixed coverage target on a separate cohort? These are two-minute simulated measurements with declared noise and calibration settings. Actual aircraft error laws and longer-record coverage remain separate requirements.

Method stays fixed

Retain compensated coordinates, local timing-cell verification, the original noise propagation and physical parameter bounds. Keep both input formats, all three generating candidates, both processing settings and two prescribed starts: 24 fits per independent noise family. Retain every 26/27-column nuisance configuration, the 200-evaluation limit, 620-operation linear-work limit and all numerical and physical forecast screens. No extra starts or retries.

Choose between available starts by training cost alone. Aggregate each complete family before scoring it. Save forecasts before consulting later truth. A missing required setting prevents a finite calibration threshold; it cannot be silently removed. Preserve the existing cooperative failure rules.

Proposed separate cohorts and limits

The estimates scale the measured 65.7-minute qualification run. They are planning estimates, not performance guarantees. The earlier 120/72-hour ceilings belong to closed releases and cannot be extended or reused. This proposed release requires its own reviewed compute allowance before launch.

Calibration seeds: 281474977210656–281474977210954. Validation seeds: 281474977310656–281474977310836. Before freezing, verify separation from every earlier declared calibration, validation, development and reserved family. Exclude earlier cohorts even where their fits were never executed. Preserve the new cohorts without examining validation outcomes during development.

Use four workers with one numerical thread each, at most two native builds per worker and the shared deterministic task and prepayment formulas. Derive and verify exact native-call and packet-visit totals before preparation. Charge work before execution, retain interrupted starts and close unused work when a phase ends. Keep calibration and validation source, configuration and numerical environment identical; freeze both policies before calibration starts.

Allow at least 55 GiB free before calibration and 35 GiB before validation. The release launcher checks this before starting a worker or charging a fit. The completed controls suggest about 15–20 GiB of uncompressed case artifacts for calibration and 9–12 GiB for validation. Publication can temporarily duplicate them. Verify, archive and compact calibration before starting validation. These estimates allow headroom; they do not guarantee storage use for every possible outcome.

Statistical gates stay fixed

• Calibration uses the maximum of 299 independent family scores to target 99% coverage with 95% confidence on the declared generated population.
• Independent validation targets 95% family coverage with 95% confidence: at most four failed families among 181, using the existing exact calculation.
• If both starts fail a required calibration setting, stop cooperatively. Stop validation at its fifth failed family. Keep every failed outcome.
• Validation needs a finite, audited handoff from the entire calibration cohort. A partial cohort, changed method or changed environment cannot supply it.

These targets calibrate generated prediction errors. They do not establish the significance of the minimum-over-globe-models decision on aircraft observations.

Shared implementation and readiness checks

The shared runner uses a counted solver scope that finishes validation before scoring or family selection. The closed-result reporter independently checks method-specific evidence while preserving saved fit labels. A small release policy supplies the separate cohorts, method binding and resource preflight. Source-bound qualification and a fresh execution freeze are required before either phase launches.

Reuse the existing campaign executor, family scoring, closed-result reporter and calibration handoff. Add a small release definition and an explicit timing solver provider, with per-fit linear-work journaling. Preserve the original defaults and archived replay; do not copy a campaign worker.

Require source-bound tests for cohort separation, complete task counts, population/source/environment identity, provider forwarding, charges preserved after failure, cooperative stops and the handoff. Independently audit every returned timing check with the correct parameter count and saved work journal. Bind the completed matrix's exact hashes and unchanged numerical method bodies. Prepare a fresh source/input/environment freeze only after the software passes.

Publish all planned rows and returned outcomes, including failures. Packaging must never refit. Verify receipt hashes and archived replay, compact evidence through the shared store and update the public roadmap at each milestone. Protected ESA groups, original measurements and the 53-stretch findings remain unchanged.

Current roadmap · Next analyses · Previous closed release.