Core pipeline and prospective comparisons

The follow-up software is implemented and checked with controlled fixtures. No new campaign, actual-route replay, optimizer search, calibration or validation has run. The previous 72-case comparison remains the latest campaign evidence. The recording app is unchanged.

An IMU (inertial measurement unit) measures turning and acceleration. The new research tools simulate its behavior along a fixed airline route and ask which turn times preserve useful Earth-model comparisons despite wind and instrument errors.

Implemented work

• A shared aircraft-motion mask checks integrated IMU turns. Aircraft maneuvers or unverifiable GPS make later orientation unresolved. Geometry search adds a configurable two-minute buffer.
• Optional design calculations check all model anchors across crab/wind states, forward-angle uncertainty and mount-epoch perturbations. The weakest retention/information sets the score. Geometry search uses beam 16 and known assumed axes, with no simulator or nonlinear fit.
• Wind, mixed bias, correlated noise and temperature are composable with independent named RNG streams. Wind changes assumed aircraft heading/airspeed while preserving the recorded ground path.
• Optional direct pair profiles use their own convergence, nonlinear bootstrap and prior gates, and the pair difference before the global science cutoff. Integrity and selection checks remain. Their thresholds/pooling method differ from older comparisons; mixed methods are rejected.
• Diagnostics retain the original session, available turn inputs/mount matrices/bins/fit rows and analysis or failure. Unreached stages have no fabricated arrays.

The default three-model rule and numerical cutoff are preserved. The envelope is a finite development grid, not a guarantee for every nuisance state or reconstructed orientation. No final empirical decision threshold or real-IMU evidence has been added.

Seven archived tracks

Five windows have two frozen replay specifications each. Two remain coverage-blocked.

The table and preparation artifacts below describe the earlier source freeze. A subsequent CI repair smooths the bank-change estimate over the same 30-second window as the GPS course estimate, preventing position noise from falsely blocking clean IMU turns. The maneuver masks, timing controls and prospective freezes must be refreshed before new runs; this historical table is not a recommendation under the revised implementation.

These are preparation-only checks on an assumed interpolated path, including its course and climb variation. They do not establish actual aircraft motion between fixes or an optimized schedule.

The observed-fix mode retains original observations and unknown accuracies. The simulated frequent-GPS mode assumes a smooth path between nearby fixes and generates simulated GPS and IMU readings. Estimated positions and intervals over 90 seconds split support. Neither bridges long gaps. Source timezone/altitude reference remain unknown; the clock is synthetic and reported altitude is assumed geometric height. Empirical policies are refused pending domain enforcement. This does not enable an operational external-GPS fallback.

Unrun proposal and cost

The proposed matrix has 1,440 replay and 720 timing-control evaluations: 2,160 total, using paired fresh seeds 600600–600602, three truths, four nuisance scenarios, two crab fits and anchor/envelope-profile candidates. These are repeated fits of shared simulated recordings, not independent flights. Bootstrap is zero; this cannot establish error tails.

Timing controls include the original 75/90-minute schedules, the second 90-minute turn shifted by ±5/±10 minutes, and the earlier correction. The ±10-minute single-turn shifts violate unchanged spacing and are blocked. The remaining controls each fail at least one buffered maneuver check; unsafe controls must abstain and cannot become recommendations. Per-route search should precede choosing an informative replay schedule.

Historical analysis cost was 7.18 seconds per evaluation. New envelope/profile runtime and diagnostic storage are unmeasured. Search bounds total 2,580,930 SVD evaluations per objective across the five routes, with four objectives proposed; no searches have run. A separately authorized bounded timing/storage pilot and search budget are needed. Authorized attempts: zero.

Files and commands

proposal.json freezes implementation/configuration/policy/two-thread environment and source hashes. optimizer-estimates.json retains maneuver checks and cost bounds. trajectories/ contains the ten hashed replay specifications. Historical evidence is unchanged. Preparation tool analysis/tests/prepare_core_pipeline.py runs no flight/search/SVD scoring and refuses overwrite.

After compute approval, research.py accepts --geometry observed --trajectory SPEC, --design-mode envelope --pairwise-method profile, and --diagnostics-root DIRECTORY. Requested profile bootstrap requires --bootstrap-refit nonlinear. optimize_turns.py supports --track-plan, --track-id, --comparison, --maneuver-buffer-s and --estimate-only; the old complete-pipeline search requires --mode pipeline. Use the existing virtual environment and at most two total CPU threads. A manifest is not run permission.

177 focused checks passed in 5.97 seconds, covering encoded replay, fixed ground path/gaps, unsafe turns, diagnostics, optimizer ordering/caching, profile fitting/eligibility/calibration compatibility/pooling and prior/bootstrap failure abstention. These fixtures are software checks. The physical wind/airspeed candidate has subsequently been implemented and checked; its matched campaign comparison remains unrun. The optional magnetic-ambiguity comparison has also been implemented and checked; its campaign comparison remains unrun. Operational domain enforcement and sharded execution remain deferred. Default magnetic exclusions remain. Final-domain approval, fresh calibration and independent validation still require separately authorized work.