Next-stage research preparation — 2026-10-06

Both 1,000-attempt pilots are complete. The original preparation used their saved records and seven observed position tracks without new fits. Subsequent 300-case geometry and 72-case protocol diagnostics are also complete; neither established a protocol that passes all required comparisons across both crab models. No fresh calibration or validation has run. Real IMU testing is pending hardware. The current scientific gate is retained; no production policy or external-GPS fallback is promoted.

Subsequent observable-domain enforcement binds flight thresholds to an explicit route/protocol envelope and acquisition source. It abstains outside that envelope and rejects calibration/validation domain mismatches. This software guard does not approve the historical restricted proposal below; its source freeze is stale, no usable experimental domain has been established, and further execution allowance remains zero.

The sharded runner now implements deterministic task identities, two single-thread local workers, durable independent journals, interrupted attempt preservation and streaming duplicate-checked merging. It has software-fixture checks, including two actual processes; it subsequently completed the initial three-attempt airline pilot, preserving all three preprocessing failures. Sharded preparation needs a fresh one-thread environment freeze. These engineering changes do not approve a usable domain, set decision thresholds or grant additional compute.

A refreshed airline-route pilot now freezes 18 matched evaluations under the current source and one-thread worker environment. Its authorized initial three evaluations are complete, all failing before fitting for weak forward-reference correlation; the remaining 15 are unrun and unapproved. Five archived paths were reviewed without fits or SVDs; speed/interpolation diagnostics expose assumptions that need testing. Its three-attempt allowance is exhausted and does not revive older plans.

The saved-data follow-up identifies amplified GPS course noise and artificial bank jumps from the original position interpolation. Matched smoothing is a research-only forward-reference diagnostic with no angle uncertainty or flight acceptance. An explicit C2 route alternative preserves positions and gaps while eliminating bank steps on the selected window. Neither has run through the complete science fit. The trajectory extension changes the scientific source freeze, so the completed pilot plan remains historical and cannot be resumed as new work.

The fresh matched-reference pilot now freezes the opt-in uncertainty-aware method and smoother route with fresh seed 600901. Controlled checks include continuous small aircraft corrections; covariance treats GPS and gyro jointly and splits at gaps. Its approved initial three evaluations are complete: all fits converged, none passed the scientific gate. Only 57 minutes of cruise survived; heading diversity and every intended design contrast failed. The other 15 remain unrun and unapproved. Coverage, calibration and independent validation remain pending.

Subsequent geometry-only preflight screened 192 overlapping 75-minute windows across the saved tracks and 9,640 safe three-turn schedules, with no passing window and one missing-height failure. Twenty-one longer extensions around the sole no-turn duration/heading pass produced 19 observable-screen passes. The extended airline pilot uses a 120-minute AUH–ORD window with 25/50/85-minute turns. Its approved three-case prefix is now complete: all fits converged without analysis failures, retaining 97 cruise minutes with sufficient heading diversity and a passing forward reference. Every intended design contrast still failed, with fitted and worst design ranks zero. All pairwise decisions abstained; the optional mount comparison was unavailable, with no flags or coefficient shifts. Wall time was about 4m42s. The budget is spent and 15 cases remain unrun. Source and acceptance gates are unchanged. Diagnose nuisance/contrast losses before another expensive pilot.

That saved-array diagnosis is now complete: smooth bias drift dominates most checked projection losses, and westbound geometry weakens rotating-globe versus disc separation. Six fixed nominal turn controls do not restore all contrasts. A course-gap interpolation bug in the analytic design tool is repaired; source was a7cc6194... after that repair; the later shape-rule extension changes it again, so old prospective execution freezes are stale. Completed pilots and their decisions remain unchanged. Next compare route direction and feasible signal modulation, keeping the nuisance space, before budgeting a complete-envelope/full-pipeline comparison.

The analysis now also preserves globe versus the specified disc as a separate experimental result when rotation cannot be resolved. It uses the existing gated comparisons; it does not rescue a route whose globe/disc contrasts fail. The saved first and second pilots give 889 and 900 correct shape preferences, zero incorrect preferences, and 111 and 100 abstentions. These add 9 and 12 preferences beyond their three-model winners. No new fits or flights ran, and the composite decision still needs its own error-budget validation.

The geometry tool can explicitly optimize both shape comparisons without requiring rotation separation. Keep both globe alternatives in that design objective, or declare a single pair when only that question is intended. Next screen different route directions and safe turn patterns under these explicit objectives with the same nuisance space.

A potentially useful real airborne source is also documented for inspection. Its raw IMU archive could support a separate empirical discriminator test after format and processing checks. The first supplied IPUTI0 measurement sample has now been decoded for inspection; its status/navigation messages provide no independent raw gyro channels.

The direction controls are now complete: 21 fixed patterns, 12 safe, 24 candidate evaluations, including two no-bin failures. Opposite traversal of the longer route gives a nominal globe/disc pass under physical wind, while rotation remains unresolved. It still fails under the broader wind correction; no accepted protocol was found. The NASA inspection has confirmed actual ILVIS0 gyro filenames, a second distinct IPUTI0 source and login redirects on both access paths. An Earthdata account is now available. The preserved IPUTI0 sample spans about 25½ minutes, with 1,527 verified packets of each message type and 11 unframed bytes preserved without repair. Three focused decoder checks passed. The subsequent ILVIS0 sample contains 132,444 time-tagged IMU8 packets at approximately 200 Hz over 662.226 s, with GPS in the same log. All 143,471 outer checksums pass. Timed opaque payloads are preserved, but physical gyro units/axes/corrections still need documentation. Two focused checks pass; no empirical Earth-model test has been run.

The remaining plan separates implemented review features from the geometry, device qualification and fresh calibration/validation evidence still needed. A 30-state follow-up preserves the reversed physical-wind shape pass in a small registered subset; it does not certify the full envelope, and the broader wind candidate still fails.

Saved-record review and chosen restricted domain

pilot-comparison.json compares both pilots under the same 10% rank-margin requirement, without modifying the original records. They used the same exact trajectory, nuisance-fit settings and numerical environment. Their source versions and candidate IDs differ and remain separate; the first pilot's margin review is retrospective, the second's margin was preregistered.

The first diagnostic rejection is retained. Do not claim that the combined pilots demonstrate zero false rejection or pool their source versions as independent calibration/validation evidence. Within each pilot, shared seeds pair the truths and scenarios.

Wind watchdog flags are 132/500 in the first pilot and 120/500 in the second, versus 6/500 and 3/500 under mixed bias, despite zero injected mount creep. Every design-contrast failure has a watchdog flag: 91 in the first pilot, 72 in the second. The second saves 129 flagged segment records, including 36 without calibrated airframe-field subtraction; the first lacks those details. The prior eight-case known-wind oracle establishes crab confounding in those diagnostic cases, but its simulator heading is unavailable operationally. Keep the protective exclusions; do not improve apparent acceptance by disabling the watchdog or removing unfavorable records.

The restricted proposed calibration domain remains:

• The exact 75-minute protocol, SPP variant, dynamic crab and bias, axis/segment weights, forward-axis uncertainty, 20 nonlinear bootstrap replicates.
• Three model truths × bias_mixed and wind; the same simulator assumptions as the pilots.
• Model-test rank 2, a 10% boundary margin, all intended design contrasts estimable, valid convergence/bootstrap and the existing integrity/prior/WMM-selection gates.
• Primary and separate pairwise thresholds, calibrated over all preregistered diagnostic cells.

This defines a restricted simulated benchmark, not an eligible envelope for arbitrary airline routes or real instruments. Rank-1 evidence remains a development/pairwise diagnostic; this proposal provides no rank-1 calibrated decision. Broad geometry, combined faults, different sensors and pooled pairwise summaries need separate review and calibration.

Prospective geometry stress cases

geometry-stress-plan.json selects 24 deterministic synthetic cells from the 288-cell matrix. They cover both hemispheres, low/mid/high latitude, two/six headings, poor/good heading separation, 60/180 minutes, 150/270 m/s, and zero/three/six mount turns. Axis cases plus poor two-heading/no-turn interactions avoid treating speed alone as measured contrast information. All 288 schedules now meet the ten-minute spacing/five-minute edge rule: the former 60-minute, six-turn schedule used eight-minute spacing, and is corrected to 5/15/25/35/45/55 minutes. That prospective-source fix invalidates old source freezes for future campaigns; completed evidence retains its original hashes and is not rewritten.

The next proposed design comparison is three truths × two nuisance scenarios × dynamic/wind crab fits for each selected cell: 288 evaluations, plus 12 exact-protocol controls if included. The 300-case diagnostic batch used fresh development seed 600500, geometry-only fitting, zero bootstrap and a one-hour cumulative evaluation cap. It completed with six no-fit records; no geometry passed all comparisons. A saved-record audit corrected unit-dependent nuisance rank selection, followed by a 72-case protocol comparison. All 72 fits converged, but neither protocol passed every comparison. Processing/timing diagnostics remain necessary before choosing a domain for power and tail work.

Observed cases preserve the position rows, estimate flags and explicitly usable intervals:

Selection uses the largest observed latitude span among 75-minute windows satisfying the provisional 95% coverage/90-second gap screen. It uses no gyro or inference outcomes. Blocked tracks remain in the plan, with their gaps/estimates, for coarse inspection. Prepared windows are not certified cruise or scientific eligibility. Unconfirmed UTC and altitude references remain explicit; elapsed geometry alone does not align a real IMU recording.

Observed-trajectory replay is now implemented with controlled checks; execution on these tracks is still pending. Core-pipeline preparation freezes two explicit modes: original observed fixes, and simulated frequent GPS on an assumed path inside short observed intervals. Neither bridges long gaps or uses provider estimates as observations. Mount turns, airframe heading, wind, sensor noise and bias are simulated assumptions. Original accuracy remains unknown; the frequent-GPS mode has explicitly simulated accuracy.

Frozen proposal and compute budget

calibration-proposed-manifest.json preserves the earlier source, configuration, eligibility and one-thread numerical environment. It is now stale. readiness.json records current stage status and artifact hashes. The original source freeze is historical; the protocol comparison used a new two-thread environment freeze. Review the final domain and refreeze before calibration. The historical proposal is not a final validation policy or run permission.

Sample goals remain 29,285 accepted calibration and 33,169 accepted validation samples per truth/scenario cell, with calibration alpha 0.00135, claimed alpha 0.0027, 18-test family, validation failure budget 64 and planned validation power at least 90% at rate 0.00135. Development samples cannot substitute for either partition.

compute-budget.json uses the latest pilot's rank-2 acceptance and measured runtime, without pooling historical source versions:

The reserve uses simultaneous one-sided acceptance bounds across the six cells and negative-binomial attempt caps across both phases, splitting a 5% planning error allowance between those stages. Its minimum acceptance input is 0.67456. The calibration proposal has 43,830 attempts per cell. This is conditional on stable independent within-cell acceptance; it does not guarantee future calibrated-rule acceptance, runtime or actual experimental coverage. Parallel scaling is unmeasured. Validation cannot be frozen yet: actual primary and pairwise threshold files must come from fresh calibration first.

Reproduction and handoff

Preparation tool: analysis/tests/prepare_next_stage.py. Run with numerical threads set to one, both campaign .json.gz archives, both normalized-track archives, this output directory and --calibration-manifest pointing at the proposal. It reads saved evidence, never calls a flight runner, and rejects holdouts/replays, overlapping attempts, mismatched policies/settings and stale proposal freezes. Source/archive hashes are preserved in its outputs.

Verification: 27 focused offline checks passed, including protocol feasibility, preservation of estimated positions, rejection of holdouts/overlaps and conservative attempt planning. The previous Python CI fixture fix passed GitHub Actions before this preparation.

Next use saved processing diagnostics and independent trajectory/attitude fixtures; additional flights require a new bounded compute budget. Observed-trajectory integration has controlled checks; a bounded real-route campaign and review/refreeze of the final domain remain pending. Fresh calibration, threshold freezing and independent validation follow only after separate compute authorization. Real IMU bench qualification and production promotion remain pending hardware/evidence.