Fresh prediction calibration after optimizer qualification

This plan defines a new finite release; it does not start one. The combined optimizer audit qualifies all 48 computational cases, 24 starting-point pairs and four original-input replays. All previous calibration campaigns and their unused allowances remain closed.

Question and fixed method

Can the corrected method produce a finite prediction-error threshold on new generated examples, then meet the fixed coverage target on a separate cohort? These are two-minute simulated measurements with declared noise and calibration settings. They do not establish the error law or confidence level of aircraft observations.

Retain compensated integration, local timing checks, feasible-anchor bounded linear solves, the nonworsening guard, all original physical bounds, all 26 or 27 nuisance columns and the existing numerical and forecast screens. Keep both input formats, three generating candidates, two processing settings and two prescribed starts: 24 fits per independent noise family. The limits remain 200 evaluations and 620 linear operations per fit, with no retries.

Choose between starts by training cost alone. Save forecasts before scoring later truth. Every required setting must produce a usable result; missing settings cannot be silently dropped. Preserve family-level scoring and the existing cooperative failure rules.

Proposed cohorts and compute allowance

Use four workers with one numerical thread each. The estimates scale the 51.0-minute matrix and 4.6-minute completion controls, together covering 48 completed cases. They include interruption overhead and remain approximate; different noise draws can need more evaluations. Total expected runtime is about nine days, with a sixteen-day combined ceiling. Early stopping may end a failed phase much sooner, without supplying a usable threshold.

Proposed calibration seeds: 281474977410656–281474977410954. Proposed validation seeds: 281474977510656–281474977510836. Exclude all earlier declared calibration and validation cohorts, including unexecuted reserved families, plus exposed development controls. Verify the complete exclusion list and phase separation before freezing. Validation outcomes remain unavailable during method development.

Maximum native prepayments are 430,229,904 calls and 2,701,855,515,528 packet visits for calibration; 260,440,176 calls and 1,635,571,399,032 visits for validation. Derive those totals from the unchanged 24-case table and verify them in software tests. Charge before execution. Preserve interrupted charges and failures; close unused allowance when a phase ends. Freeze both phase policies, sources and environment before calibration starts.

This is a proposed new allowance, separate from every closed release. Prepare and qualify the implementation and exact execution freeze before the compute review. Do not launch either phase merely because the preceding controls pass.

Temporary and retained storage

Require at least 55 GiB free before calibration and 35 GiB before validation, checked before any worker or fit charge. The latest closed calibration supplies 348 retained case examples: their average uncompressed size is about 1.83 MB and their largest is about 3.01 MB. Scaling those sizes suggests roughly 12–20 GiB for calibration case artifacts and 7–12 GiB for validation. Publication can temporarily duplicate these files.

The same closed study retains 260.06 MiB of unique compressed payload for its 348 returned cases. Scaling that measurement suggests about 5.2 GiB for calibration and 3.2 GiB for validation, plus metadata, journals and archive overhead. Those are planning estimates, not upper bounds. Original datasets and already retained project evidence are additional storage.

After each phase closes, verify and archive it, then compact recoverable working copies before starting the next phase. Serialize archive writers. Keep original measurements, source snapshots, failed outcomes and journals; do not delete unarchived scientific evidence to meet a disk limit.

Qualification and shared implementation

Reuse the existing campaign executor, family scoring, independent reporter, calibration handoff and counted solver scope. Add a small release policy rather than copying a worker or altering the qualified numerical code.

The completed four-case publication supplies the source/environment baseline already supported by the shared preparer. The separately hash-bound combined audit supplies proof of complete 48-case coverage. Both refer to the same 241 common execution sources. Bind the exact combined report hash, both constituent completion hashes, the completed source map and the numerical environment. Reject any changed shared source or incomplete coverage before preparation or execution. No preparer hook or numerical-method change is needed.

Require source-bound tests for all cohort exclusions, exact task/prepayment counts, unchanged physical/noise inputs, complete combined-gate binding, provider forwarding, work retained after failures, handoff rejection and disk preflight. Calibration and validation must use the same population/source/environment identity. Freeze only after these tests pass.

Statistical gates and interpretation

• Calibration takes the maximum of 299 complete independent family scores to target 99% coverage with 95% confidence on the declared generated population.
• Independent validation targets 95% family coverage with 95% confidence, allowing at most four failed families among 181 under the existing exact rule.
• Stop calibration cooperatively if both starts fail a required setting. Stop validation at its fifth failed family. Retain every returned outcome.
• Validation requires a finite, audited handoff from the entire new calibration cohort. Partial cohorts and changed methods or environments cannot supply it.

Publish every planned row and returned result without refitting during audit. Verify hashes and archived replay, then update the public roadmap. These targets calibrate generated prediction errors; they do not establish significance of the two-stage Earth-model comparison on actual observations. The original 53 aircraft results and protected ESA confirmation groups remain unchanged.

Next analyses · Optimizer qualification · Previous closed calibration.