All qualifying high-speed airborne stretches

This extension checks all 232 retained ILVIS0 recordings and preserves every qualifying stretch. Both the six-recording refinement and this extension are complete. Screening found 53 eligible stretches totaling 424.3 minutes; these can overlap across instrument streams. Earlier evidence and originals remain intact.

The completed comparisons favor globe in 12 stretches and flat in none; 41 remain numerically unresolved. All 12 resolved stretches favor rotating over still globe in both processing cases. The unfinished comparisons are not support for flat Earth. The findings explain these results, the partial comparisons and specific remaining checks. The original audited summary reports every stretch and fit.

An interval must last 240 continuous seconds after exclusions, with every associated receiver ground-speed observation at least 700 km/h. It must also satisfy:

Fused navigation screens aircraft motion only; it is never an independent Earth-model observation. Sampled limits do not rule out tiny aircraft corrections. No gyro residual or preferred model determines selection. Ground speed is neither true airspeed nor a flight average. Geometry eligibility is separate from supported decoding, convergence and model information.

Permanent selection record

Before extension fitting, the worker saves the complete technical selection and publishes:

• eligible-stretches.csv: stable stretch identifier, original filename/SHA-256, exact UTC start/end, duration, minimum/median receiver ground speed, IMU type, mounting epoch, policy version and selection fingerprint.
• files.csv: all retained files, including exclusions and unresolved context with reasons and nonexclusive rejection counts. Byte-identical aliases reference their canonical file.
• ELIGIBILITY.md: plain-language criteria and selected totals.

These were published before fitting finished. They are generated from the frozen selection, so publication and later analyses need not recalculate intervals. Changing the rules requires a new version. Different instruments on the same flight may overlap; recording counts and minutes are not counts of independent flights.

Analysis and recovery

Each complete stretch is fitted jointly without a ten-minute primary limit. Fixed four-to-ten minute sections check residuals at the same parameters; they are not separately fitted decisions. Known gaps/configuration changes split stretches. Joining files requires verified continuity.

Two primary sensitivity cases use the same ±1 degree/hour constant gyro-offset allowance:

The survey-grade systems' published capabilities support testing slow turning. The offset allowance and possible Earth-rate removal specify what this comparison permits; neither is a measurement of actual unit drift or proof of its onboard corrections. The tighter ±0.1 degree/hour sensitivity is deferred for the extension; the completed six-recording refinement retained all four cases. Other error limits and assumed correlated receiver covariance match that refinement. Time-varying calibration errors are not added in this pass.

All three models are fitted together under each case: flat disc, still globe and rotating globe. The first reported question is globe, flat or unknown, comparing the flat fit with the better globe fit. Rotation results are exposed only when all required fits converge and both cases favor the globe. This reuses the globe fits; a separate rotation fit is unnecessary. The report states which model fits better. A validated statistical significance level has not been assigned; the software's separate significance gate still abstains.

At launch, source/input/environment hashes are frozen. After inventory, a second manifest freezes identities and the finite allowance: N stretches permit 6N primary starts plus 2N interruption retries, at most two starts per identity and 200 total evaluations per start. Interrupted starts stay charged; verified selections/results are reused. Exclusive locks, fsynced journals and atomic hashed outputs support recovery. At most two numerical threads; no synthetic campaign or original deletion.

Code: analysis/lll/ilvis0_segments.py; runner: analysis/tests/ilvis0_segment_worker.py. Output: data/ilvis0-highspeed-segments-v2-20261008/; matching .log beside it. The superseded four-case extension was stopped while queued, before inventory or any fit start. Its original freeze remains preserved; it must not be resumed. The geometry-selection policy is unchanged; only the fit-case grid and its separately frozen run have changed. Check lock/status before resuming an unfinished, unchanged freeze. Completed studies never resume. To regenerate public evidence after completion without scans or fits:

env PYTHONPATH=analysis OPENBLAS_NUM_THREADS=2 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 /home/sweisman/venv/bin/python analysis/tests/ilvis0_segment_worker.py --package-only

See the general methodology and instrument/refinement assumptions.

Public results summary

The pilot findings are available now. The findings explain shape and rotation outcomes in plain language. The original audited summary compares convergence by model and processing case, fully converged versus unresolved stretches, residuals and parameter limits. The separate summarize_results.py --watch waits for audited publication and reads only saved results; it does not scan original recordings or run fits. To regenerate that summary after completion, run the script without --watch.

The detailed stretch ledger includes original filename/hash, Group-4 record type and logged IMU type, UTC bounds, duration, minimum/median receiver ground speed, approximate horizontal track distance, mounting epoch and conversion information. A separate track CSV retains every selected receiver position without interpolation. The distance uses a declared spherical coordinate convention for description, never as evidence about Earth shape. For indeterminate results, matched globe/disc pairs are described only when both fits converged; unfinished costs remain diagnostics. Partial leans do not promote the primary shape decision or expose rotation when the complete shape comparison remains unresolved.

Review after the frozen run

The analysis roadmap sets the next priorities: numerical qualification, calibration-boundary diagnosis, absolute fit adequacy, matched clock and receiver-coordinate checks, and recording dependence. Exact device characterization is lower-priority parallel work; available survey-grade specifications and empirical checks support continued analysis without waiting for a producer reply.

The completed saved-result diagnosis reports stopping reasons, speed/duration associations, parameter limits and clock sensitivity. The bounded refinement plan specified the separate matched solver trial, now complete with 10/12 converged fits versus 11/12 before. It did not improve completion. Its allowance is separate from this completed study; it does not change the frozen results. The calibration and adequacy audit reports signed limits, local weak directions and residual quality from all saved fits.

First compare convergence with duration, receiver ground speed, IMU type, installation and route geometry. Use the number of converged fits per stretch as well as whether all six converged. Describe correlations without treating overlapping stretches or repeated instrument recordings as independent trials. The completed subset during execution is ordered by source date, so its relationships need not describe the final selection.

Then separate numerical difficulty from scientific separation. Review stopping messages, stationarity, parameter-limit hits, residual patterns and poorly constrained parameter combinations. Check whether solver scaling, initialization or parameterization can improve convergence while preserving the same physical model. Review calibration and processing assumptions against available instrument evidence, and test their effects equally across all Earth models. Faster travel strengthens the curvature scale, but neither speed nor duration guarantees convergence or useful separation from instrument errors.

This completed study remains preserved. Any revised fits require a new identified development configuration and finite approved compute scope. Reuse the saved intervals; do not tune selection boundaries or assumptions to favor an Earth model. Improvements chosen using these results are development findings and require separate validation.