Acquiring and decoding the airborne IMU archive

The completed high-speed analysis favors globe in 12 stretches and flat in none; all 12 also favor the globe candidate containing the standard background rotation field. Another 41 comparisons lack finished calculations. See the findings and source tracks. This is a substantial empirical finding. Unfinished calculations do not erase it or provide equal support for flat Earth. A validated statistical significance level has not yet been assigned, and shared flights or instruments must be accounted for before counting the stretches as independent replications.

The historical fit labels “rotating globe” and “still globe” mean a globe with the specified background angular rate and a globe with zero background rate. They identify the predictions being compared, not the physical cause of a gyro signal. The archive does not compare every possible stationary-Earth explanation with a nonzero inertial field. See measurement and interpretation.

The complete-stretch scope selects every supported interval lasting at least four minutes with receiver ground speed at least 700 km/h throughout. A versioned inventory records files, exact UTC times, speed and exclusions before fitting. Whole stretches are primary joint fits; shorter sections check consistency with the same parameters. These criteria do not establish model separation. See the shared methodology.

NASA's ILVIS0 archive contains Applanix recordings from survey aircraft. An inertial measurement unit (IMU) measures motion using gyroscopes and accelerometers. These .013 files record small changes in angle and velocity; dividing an increment by its measured integration interval gives a rate-equivalent value. They also contain GPS receiver sentences and the instrument's fused navigation solution. The latter combines measurements with a navigation model and is not independent evidence for Earth's shape or rotation.

Measurement provenance and independence

The archived measurements and the finished navigation solution have different roles:

The Applanix V6 interface specification, document PUBS-ICD-003259 Rev. 9.0, section 3.3.2.1, printed page 64, says that Group 10001 packages the primary receiver's output irrespective of message boundaries. Concatenating its payloads reconstructs that receiver stream. The repository's packet decoder extracts those payloads; the observation extractor reconstructs and checks receiver sentences without substituting Group 1 navigation observations. The original inspector also preserves the concatenated bytes and original sentence identifiers where those artifacts were produced.

This supports separate IMU and receiver observation sources and avoids treating the fused INS answer as independent evidence. It does not establish statistical independence of their errors, absence of all receiver-side aiding, or identical behavior across the archive's historical configurations. The cited document describes V6; earlier files also need their recorded version and observed layout checks.

GPS positions are navigation outputs, not direct measurements of surface curvature. Their latitude, longitude and height use geodetic conventions. WGS84 provides GPS's reference frame and an ellipsoidal geometry; see NGA's description. The present comparison applies a specified globe or disc metric to receiver coordinates and integrates the recorded IMU increments. It therefore tests explicit observation models. Planned gyro-channel, horizontal-only and height/geoid controls will isolate which measurements and conventions produce the separation.

The project makes no verified scientific-first or model-free-proof claim. Physical measurements of Earth's figure predate WGS84. Snellius's Eratosthenes Batavus (1617) is an early triangulation-based measurement of a meridian arc and Earth's size; an English translation is available at Ars Astronomica. NOAA's history of geodesy describes measured survey baselines, angles and meridian arcs. The contribution here is an openly reproducible comparison using professional airborne IMU increments, receiver observations, explicit alternatives and preserved failures. A narrower publication novelty claim requires a dedicated literature review.

The planned per-file provenance supplement includes numeric Group 10001 receiver-type codes, packet counts, firmware/ICD strings, receiver-stream hashes, and original NMEA talker/sentence identifiers with checksum status, configuration changes and binary-only or missing-NMEA outcomes. A talker identifier alone does not establish the receiver model. The packet decoder already returns the receiver-type code, but the current observation summaries do not retain it or every original sentence identifier. Those fields remain recoverable from retained originals. The supplement will be separately versioned, preserving the frozen inputs and original result labels. The interface reference above documents the stream's role.

Instrument quality

These are professional airborne survey systems. The validated reference identifies POS AV 510, IMU8 and 200 Hz. Applanix's historical FAQ describes specially selected LN200ROM units for the 510; an LN-200-family identification is therefore likely, rather than proven by the type byte. That family uses three fiber-optic gyroscopes and three silicon accelerometers. Manufacturer description

Applanix publishes 0.02°/√hour noise and 0.10°/hour attitude drift for POS AV 510 relative system performance. The drift figure is not a verified raw-gyro bias specification for this recording. Published post-processed attitude accuracy is likewise fused navigation performance. POS AV specifications This instrument class is suitable for investigating slow turning at degrees per hour. Survey-grade hardware is a strength of the analysis; unfinished optimization is currently the main reason comparisons cannot be reported. The archive also contains other IMU types and POS AV 610 configurations. Their identities and conversion checks are kept separate, rather than assigning the POS AV 510's exact specifications to every file.

Current modeling order: compare the globe family with the flat-disc candidate first; report rotating-versus-still diagnostics only if that shape comparison consistently favors the globe under every tested assumption. The completed six-file sensitivity check uses constant gyro-offset allowances of ±0.1 and ±1 degree/hour. These are calibration hypotheses, not measured drift or guaranteed limits. The historical ±20 degree/hour allowance was a broad stress test. Constant offset, random noise and time-varying drift are different errors. The refinement record explains the numerical checks, public performance context and completed pilot. The completed all-stretch extension uses two primary cases at ±1 degree/hour, with Earth-rate removal fixed at zero or fitted; its ±0.1 sensitivity is deferred. All three models are fitted together, with shape interpreted first and rotation reported only after a resolved globe result. Reports state which model fits better; a validated statistical significance level has not yet been assigned.

The qualified recovery findings add twenty-nine completed globe preferences to the original twelve: 41 globe, zero flat, 12 unresolved. Completed findings include IMU8 and IMU21. These are recorded instrument/data configurations, not unique serial numbers or separate experimental protocols. Their units and conversion scales are checked separately; both follow the same selection and model-comparison rules. All forty-one resolved stretches also favor the specified background-field globe, without identifying that field's cause.

The supplied 14 April 2009 ATM sample now passes an empirical physical-decoder check. Its six signed integers represent velocity X/Y/Z followed by angle X/Y/Z. The recovered scales agree with 2^-14 metres/second per count and 2^-18 radians per count. The check estimates scales before comparing with those values, tests every axis/sign mapping, and holds out successive halves of the recording. Accelerometer comparison accounts for gravity in the aircraft frame. No fitted validation offset is removed from exported data. See the measured results.

The public Applanix interface document supplies structural reference information, rather than an authoritative scale table for this older unit. The observed angle and velocity increments are distinct from the finished navigation output. "Raw" describes that measurement stream, not a guarantee that the electronics applied no calibration or filtering. Ordinary noise filtering can preserve slow turning; automatic zeroing, high-pass filtering or explicit Earth-rate subtraction can remove it. A successful fit does not establish which corrections were used. The manufacturer's filtering explanation illustrates the general distinction; it does not identify an Applanix filter.

The remaining instrument questions concern bias/scale corrections, Earth-rate compensation, coning/sculling, exact filtering, automatic zeroing, lever arms, mounting and physical latency. They are investigated through recorded settings, matched processing cases and producer documentation. They do not make professional survey data equivalent to uncharacterized phone data, and the archive results do not establish consumer-IMU performance.

What the pipeline keeps

The catalog contains 326 .013 downloads, with roughly 3.97 GiB of reported data, dated 14 April 2009 through 20 September 2017. CMR sizes are rounded MiB, so HTTP lengths and SHA-256 checks verify downloads. Filename tokens explicitly identify three ATM and ten LVIS entries; the others remain unclassified. Instruments are never equated by flight date. Only actual HTTPS download paths ending .013 are selected. JPS and .031 are outside scope.

flowchart TD
    A[Catalog every .013 file] --> B[Check a small cross-date batch]
    B --> C[Validate each layout, scale and axis convention]
    C --> D[Screen GPS and aircraft motion]
    D --> E[Keep useful level-flight data]
    D --> F[Record and discard files with no useful level window]
    D --> G[Record unresolved files and keep their originals]
    E --> H[Assess independence and Earth-shape feasibility]

Acquisition uses a geometry-only rule, fixed before model-dependent gyro analysis. A useful window needs at least 60 uninterrupted seconds after ten-second boundary buffers, speed at least 50 m/s, vertical speed at most 1.5 m/s, course change at most 0.05 degrees/second, roll at most five degrees, and valid synchronized GPS/navigation with no context gaps above two seconds. A no-level rejection also requires at least 95% valid motion/GPS context; incomplete coverage and unsupported time mappings remain unresolved. This is a retention rule, not scientific eligibility. The known sample climbs and maneuvers and supplies no qualifying window under it.

Each catalog item has a stable task ID and durable ledger status: pending, retained, discarded with no level window, unresolved, invalid input, failed, or authentication required. Discarded files retain hashes, URLs, validation/motion reports and reasons. All unresolved originals are kept, following the latest retention instruction. Only newly acquired files inside the pipeline's marked directory can be removed; supplied originals and historical evidence are untouched. Retained originals are compressed without changing their bytes. Qualifying IMU windows preserve raw integers and physical values.

Commands and recovery

Use the existing Python environment. Run from the repository root. Future work shares a four-thread total budget through the numerical launch guard; historical freezes and the active recovery keep their recorded two-thread settings. These commands are ordinary acquisition/decoding, not simulation campaigns.

env PYTHONPATH=analysis OPENBLAS_NUM_THREADS=2 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 /home/sweisman/venv/bin/python -m lll.ilvis0_acquisition catalog --output /tmp/ilvis0-catalog --quiet

The catalog command writes a complete CSV and metadata snapshot before any download. Omit --quiet to display filenames, dates, URLs and reported sizes. --start and --end restrict dates. The repository snapshot is in docs/ilvis0-physical-validation-20261007/catalog/.

env PYTHONPATH=analysis OPENBLAS_NUM_THREADS=2 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 /home/sweisman/venv/bin/python -m lll.ilvis0_acquisition batch --catalog docs/ilvis0-physical-validation-20261007/catalog/catalog.json --output data/ilvis0-ready --reference /home/sweisman/Downloads/ILVIS0_gyro_54935_atm_applanix_14Apr09.013

The first batch selects at most six files by explicit stream tokens and date diversity. Earthdata's standard .netrc entry is supported; --prompt requests credentials securely in a local terminal. For a single setup usable across unattended resumes, generate an Earthdata user token and run this locally:

env PYTHONPATH=analysis /home/sweisman/venv/bin/python -m lll.ilvis0_acquisition configure-auth

It uses a hidden token prompt and exclusively creates a private, mode0600 .netrc. An existing credential file is never overwritten. The only intentional credential storage is this explicitly configured standard Earthdata file, outside reports and the repository. Never paste credentials into a chat, source file or manifest. TLS verification stays enabled. Authentication failures pause the batch instead of trying the entire archive. Individual transient failures retry and leave recoverable partial files. The download command can also select one unambiguous --filename from a catalog.

Rerun the identical batch command to resume. Verified completed tasks are skipped, append-only records are fsynced, incomplete final appends are preserved before repair, and a process lock prevents duplicate workers. status.json reports progress; inventory.json tracks every catalog item. The manifest freezes sources, screen policy, runtime and catalog. After a source change, use a new marked corpus directory rather than mixing results. Once at least two dates independently pass compatibility, --full makes all catalog items available to the same single-file-at-a-time worker. Each file is checked independently.

The URLs are HTTPS links to protected data. S3 also requires authentication and an authorized AWS environment in us-west-2; NSIDC directs local-computer users to HTTPS. See its access guide. No cloud instance, AWS credential inspection, paid compute or S3 dependency is needed here.

If a catalog file was already downloaded through a browser, --input-dir ~/Downloads checks that exact catalog filename without listing unrelated files. --local-only imports available files without authentication or network and leaves others pending. Supplied originals are never deleted by that import path.

env PYTHONPATH=analysis OPENBLAS_NUM_THREADS=2 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 /home/sweisman/venv/bin/python -m lll.ilvis0 decode --input /home/sweisman/Downloads/ILVIS0_gyro_54935_atm_applanix_14Apr09.013 --output /tmp/ilvis0-decoded
env PYTHONPATH=analysis OPENBLAS_NUM_THREADS=2 OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 /home/sweisman/venv/bin/python -m lll.ilvis0 screen --output /tmp/ilvis0-decoded

Decode writes a gzip CSV of raw/physical IMU increments, fused navigation CSV, reconstructed GPS bytes, checksum-tested GPS sentences/fixes, a compressed original and provenance reports. Unsupported configurations retain raw values without guessed physical conversion. Invalid outer framing, checksums or truncation fail; the parser never repairs or resynchronizes. Timestamp reversals, rate changes, missing intervals and status problems are flagged. A gap's elapsed time is never assigned to a single IMU increment. There is no science interpolation.

Flight speed and measured performance

The signal caused by travelling over a curved surface grows approximately in proportion to ground speed. As an illustrative spherical calculation, speeds of 100, 150 and 200 metres per second produce local-level tilt rates of about 3.2, 4.9 and 6.5 degrees per hour. The actual calculation uses latitude, altitude and north/east velocity, and projection onto the mounted IMU axes matters. The standard background rotation term has a separate latitude/orientation dependence.

The local follow-up reports these predicted scales for every geometry-selected window alongside speed range, duration and heading coverage. It validates physical units separately for each instrument configuration, then characterizes quantization and observed gyro variation after averaging over one, ten and sixty seconds. That variation includes aircraft motion and vibration; it is not automatically the IMU's intrinsic noise. Small random noise does not remove an unknown constant bias or compensate for poor heading coverage. No significance threshold is inferred from manufacturer specifications.

The follow-up uses only retained originals, without new downloads, and records a separate source/environment freeze in data/ilvis0-followup-20261007/. Its launcher is analysis/tests/ilvis0_followup_worker.py; completed per-file reports allow interruption-safe resumption. The original screening ledger remains unchanged.

That follow-up has completed: 80 candidate files, 35 newly classified no-level files and 153 still blocked by navigation alignment. Four candidate files pass physical units checks, providing about 32 minutes of qualifying data. Candidate window speeds range from about 140 to 256 m/s, with predicted horizontal transport of 4.5–8.3 degrees/hour. Measured variation among minute-averaged body-axis readings remains comparable or larger in several axes, because aircraft motion is included. See the performance and compatibility results.

A separate IMU21 and alignment assessment has completed over all 233 preserved originals. Of 98 IMU21 files, 71 pass per-file physical checks. Three of six exact configurations reproduce both selected dates; requiring both that configuration check and a file's own check yields 24 exported candidate files and about 168 minutes. Including the earlier IMU8 candidates, 28 files provide about 200 minutes across potentially overlapping instrument streams. The IMU21 scales are 2^-28 radians per count and 0.3048*2^-21 metres/second per count; these are empirically supported, not a manufacturer table.

GPS-only screening finds potential stretches in 84 of the 153 unresolved files. Internal navigation uncertainty can be considerably smaller than the broad fine-alignment flag's bound, but its accuracy is not independently demonstrated. Those stretches do not prove level attitude or usable gyro orientation. All 153 remain scientifically unresolved; all 233 originals remain preserved, and the scientific acceptance rules are unchanged.

The subsequent keep/discard review resolved storage decisions for the 84 GPS-promising, alignment-blocked files: keep 83 and remove one verified byte-identical duplicate. All 83 keepers have at least a minute of uninterrupted raw IMU data inside the GPS stretches; 22 also have a minute compatible with fused motion limits. Those diagnostic spans do not prove independent orientation or become scientific windows. The other 61 remain useful for investigation, with level-flight recovery still uncertain. Their completed motion examination reproduced all earlier instantaneous checks exactly. In 53 files, measuring course rate over 2–12 seconds permits a diagnostic minute while keeping roll, climb and speed limits unchanged. Almost 98.5% of the original course-rate excursions lasted less than a second. Three other files also have roll/climb interruptions, one has disagreement between navigation and receiver track, and four yield no minute at the tested resolutions. All 61 remain kept; neither averaging nor fused-navigation agreement demonstrates independent orientation or proves that short corrections are noise. Scientific eligibility remains unchanged.

A six-file IMU6 cross-date study supports a gyro increment scale of 0.4 arcsecond/count. Two of three configurations reproduce the earlier axis mapping on a second date. The third changes mapping; within-file chronological checks support the scale on all six, without erasing the failed cross-date result. No complete accelerometer conversion passes, and these are relative mappings to fused navigation rather than independent mounting measurements. No physical decoder is applied across the IMU6 corpus. The motion-timescale controls detected all five substantial recorded ground-track changes at each tested interval, but showed that a real one-second course correction can be hidden by a 12-second regression. This supports further motion modeling, not changing the acceptance gate merely to recover more windows.

The installation/clock follow-up now explains the axis change with logged mounting-angle candidates. Header intervals fluctuate around 5 milliseconds; treating those fluctuations as sensor integration changes produces most of the earlier vertical acceleration mismatch. An explicit fixed-period hypothesis supports the 3.38e-5 m/s/count velocity increment scale in all six files, reducing vertical disagreement by 76–595 times. One gyro half still narrowly fails the unchanged scale tolerance. Both rate normalizations remain recorded; no automatic conversion is enabled until integration-clock semantics are established. Legacy setting offsets remain inferred, and fused navigation agreement does not establish processing independence. Removing the duplicate reclaimed about 6.66 MiB. Current storage is 232 originals: 80 earlier candidates and 152 alignment-blocked files. Earlier evidence keeps its historical counts; future selection must honor the new cleanup journal and canonical-file mapping.

A subsequent continuity check covered all 326 files. Three overlapping pairs are identical copies; one verified partial overlap adds 16.7 seconds but still provides no qualifying window. No candidate level-flight stretch can currently be lengthened across files. Separated recordings are kept separate, and simultaneous different IMUs are never joined into one measurement.

What remains before an Earth-shape comparison

Level-flight footage or instrument quality does not by itself establish an independent shape measurement. We still need to understand recorded gyro corrections and reconstruct orientation/motion without treating fused navigation as independent truth. Permanently mounted survey equipment lacks the participant's deliberate IMU reversals; these data need their own nuisance and identifiability analysis.

The authorized follow-up is a separate, bounded empirical feasibility/comparison study: at most six geometry-selected retained files, at most 24 optimizer starts including interrupted starts, and 200 evaluations per start. That initial plan required suitable motion geometry and observations separate from the finished navigation solution. If those prerequisites fail, the output is a documented blocker, rather than a forced fit. No synthetic threshold transfers to these data, and no calibrated model winner can be claimed. The original reference is unsuitable under the acquisition window rule. The six-file trial found 272 seconds of qualifying flight in the 16 April 2009 ATM file, with independently supported physical decoding, and 354 seconds in a 2015 POS AV 610 file whose physical decoding was unresolved in that initial trial. Three further trial files were unresolved and preserved. The initial archive-wide pass produced 18 retained files, 58 confirmed no-level rejections and 250 unresolved cases. The completed timestamp follow-up then brought the current totals to 80 candidate files, 93 discarded files and 153 unresolved originals at that stage. The later duplicate removal leaves 152 unresolved originals and 232 originals overall. The subsequent IMU21 assessment raised supported physical candidates to 28, totaling about 200 minutes across instrument streams. Other configurations still failed some checks. No Earth-model fit had run at that stage. Subsequent studies completed the comparisons summarized at the top of this guide. The initial screening record and completed follow-up preserve that sequence.

The route from decoded measurements to the proposed experiment's principle

The shared principle is to ask whether the IMU's measured rotation contains the small, predictable contributions from a background angular-rate field and motion over its surface, after allowing for aircraft motion and instrument error. These archival recordings can address that question, but do not inherit the passenger experiment's controls or calibration.

The next development stage should use a small set selected by recorded geometry and data completeness, before inspecting Earth-model residuals. Preserve raw angle and velocity increments, logged installation epochs and timing uncertainty. Compare accumulated increments over supported intervals as well as rate estimates; this tests clock assumptions without selecting the normalization separately for whichever model fits better. Determine which onboard corrections are documented and which remain hypotheses. Looking for an expected Earth-rate component is a diagnostic, not by itself proof that no Earth-dependent processing occurred.

Then build an archival observation model that separates aircraft rotation, sensor bias, mounting uncertainty, transport and the background field. Embedded GPS supplies trajectory context; ground track does not supply body heading, and acceleration during flight does not give gravity direction alone. Fused Group-1 attitude and gyro outputs can check the decoder and describe motion, but cannot supply the independent rotation being tested. Any unavoidable use of fused orientation must be explicit. Test how its model dependence affects the comparison. GPS-derived coordinates and their geometric assumptions also need an explicit treatment under each candidate model.

Before fitting measured gyro outcomes, calculate whether the recorded changes in heading, speed and latitude separate each model pair after those nuisance freedoms are included. A fixed heading with an unknown constant bias may leave an Earth signal unidentifiable, regardless of IMU quality. Actual route changes and any independently identified stationary or taxi intervals may provide controls; do not invent reversals or smooth away short motion. Preserve useful pairwise comparisons when a complete three-model conclusion is unavailable.

Only informative, independently supported measurements advance to the bounded exploratory comparison above. If motion or processing remains inseparable from the predicted signal, report that specific limitation and the additional observation needed. A later scientific decision requires its own frozen assumptions, calibration and independent validation.

The first archival observation check has now completed on the six existing IMU6 representatives. It prepared 2,930 complete diagnostic seconds and tested geometry using primary-receiver coordinates, without fused attitude or gyro rates. The 26 October 2010 route retains predicted pairwise separation after constant/linear bias removal, including 2.50 degrees/hour RMS for still globe versus disc, under an idealized independent aircraft-motion correction. Freely varying gain/mount tangents remove nearly all separation, but that unbounded local projection is not a fit with realistic calibration bounds. No Earth fit had been attempted at that stage. The joint raw-IMU/GPS model and later fit studies followed; the passages below describe that development history.

That joint motion model is now implemented and its deterministic controls pass. It integrates full-rate finite rotations and rotated specific force, predicts navigation under explicit candidate geometries, and requires calibration bounds/processing assumptions as inputs. Six strict source scans prepared all 2,930 previous complete seconds; 2,928 have real GPS observations near both endpoints, with their timing offsets preserved. No fused attitude/gyro entered the observations and no Earth fit ran. The subsequent recorded-data modeling used the six real IMU/GPS recordings with stated correction, calibration and timing allowances, preserving unfinished calculations. The completed pilot and 53-stretch study followed. The separately frozen matched solver trial is complete and did not improve numerical completion. The saved-fit audit examines calibration limits and residual errors without new fits. The completed epoch residual diagnostic reproduces saved predictions under both clocks, preserving individual errors and their temporal structure. No refits ran. The earlier records below remain a history of completed stages, not the current work queue.

The bounded estimator is now implemented. Five analytic software controls recover bias, effective orientation, gain and timing under known conditions, and expose orientation/bias ambiguity when both are free. Fifteen focused tests pass. Pair-specific bounded tangent profiles and the empirical prerequisite/start ledger are implemented; no observed Earth fit ran. Local tangents are not a guarantee of separation between two independently fitted nonlinear models. Archive calibration, processing, clock and validated receiver covariance remain prerequisites before the bounded empirical fit. The subsequent receiver/processing audit recovers 2,968 dated primary-receiver positions with reported north/east/vertical uncertainty. Their units and epochs reproduce the receiver's GPS sentences. It also reassembles 29,681 binary RT17/RT27 survey envelopes, whose measurement payloads remain unconverted. Four receiver streams are complete; two have incomplete final packet/record tails, retained without repair. All six outer source/hash audits pass. Sixteen focused tests pass. Reported sigmas are not independently validated full covariance, and no Earth fit or scientific promotion occurred.

The satellite-measurement audit then decodes 19,699 epochs and 603,603 signal measurements at 10Hz in the two RT17 streams and 5Hz in the four RT27 streams, plus 326 GPS ephemeris packets. All six source/frame/record sequence checks pass; 31 focused tests pass. Of 29,567 GPS sky-angle comparisons, 29 flagged cases are one satellite reporting zero angles for 29 seconds. All measurements and prior tail failures remain preserved. RT17 filtering flags do not establish code smoothing; RT27's explicit smoothing flags are clear. No positioning or Earth fit ran at that decoder stage.

The separately frozen GPS-only reconstruction now solves all 2,968 previously selected epochs with two fixed code-range methods: 5,936 primary solutions and 24 displaced-start checks all converge. Recorded broadcast orbits, satellite clocks, signal travel time, ionospheric coefficients and an altitude-supported atmospheric approximation supply the corrections. No fused navigation enters the reconstruction. The first receiver position only initializes the solver; reported positions are comparison products, not constraints. Twenty-eight focused tests pass (13 new, 15 preceding).

L1 RMS differences from the same receiver reports range from 0.17–0.85m per horizontal axis and 0.46–1.79m vertically. Those are consistency differences, not measured errors. Dual-frequency differences are larger; unknown signal biases and shared processing prevent an accuracy ranking. Lag-one difference correlations span 0.20–0.89, and mean vertical method offsets span about -1.61 to +2.97m. Formal covariance remains uncalibrated. All232originals remain, with zero Earth fits or scientific eligibility changes.

The subsequent correlated-error sensitivity analysis evaluates 13 fixed assumptions over both methods and the same 2,968 epochs. Formal cross-axis terms, exponential temporal correlation, common offsets and linear drift propagate through mean positions and fixed 2/6/12/30/60-second kinematic windows. There is no residual-based selection or automatic smoothing choice. The assumed scales are not calibrated bounds. Twenty-three focused tests pass (ten new, thirteen preceding). Persistent errors can prevent mean-position uncertainty from averaging away; a short-motion control loses over 100 times its acceleration response in a 60-second window. No observed Earth fit or gate change ran. The joint IMU/GPS controls now propagate that full assumed covariance through five analytic calibration/timing fixtures. Numerical recovery is reported separately from local precision; unresolved directions have no finite marginal precision assigned. GPS offset/drift controls reproduce initial position/velocity shifts. Thirty-three distinct focused tests pass (eight new, ten preceding sensitivity, fifteen estimator). This uses two-second toy translation data, not archived IMU fits or measured IMU6 performance. Supported processing, calibration, mounting and physical integration/latency bounds remain necessary before observed Earth fitting.

The duration/rotation controls now compare2/20/60s analytic trajectories with/without a one-second0.1deg pitch correction under three fixed GPS-error assumptions. All18trajectory checks and4single-parameter gyro-bias/timing fits pass;11focused tests pass. Joint eight-parameter precision improves with duration but gyro-gain information remains weak. Covariance-dependent rank changes are preserved and are not interpreted as improved data. No archived IMU was fitted or downsampled. The IMU6 evidence matrix distinguishes empirical units/mappings from unsupported processing, noise/drift bounds and physical clock/latency conventions. The subsequent information stability diagnostics compare four central derivative steps on those same18fixtures and retain every result. Smooth reference information separates numerical rank from useful local sensitivity relative to the declared bounds. The reference penalty and GPS covariances are explicitly assumed, not measured calibration or confidence intervals. Independent instrument-specific support remains necessary; the empirical Earth-fit evidence gate is unchanged.

The instrument documentation review now finds a published POS AV510/IMU-6 association with tactical-grade Litton-200 hardware. This supports an LN-200-family inference for the six files; it does not identify their exact installed variants. The review associates existing source hashes/dates with logged POS system identifiers and firmware/ICD revisions, and prepares an unsent technical question packet. Public descriptions establish a sensor-correction stage before navigation integration but not the legacy Group4 logging tap point. Generic NovAtel LN-200 output scales differ from our empirical Applanix encoding; they are not substituted into the decoder. Recorded correction, calibration, integration-clock/latency and mounting support remain unresolved.

The legacy-record search obtained no manual/calibration record matching those V5 revisions. NSIDC's current guide supports raw/unprocessed Level-0 archive provenance but does not define onboard Group4 corrections. A verified support route and concise unsent request are ready. No external message or empirical fit ran during that document search. An inquiry has since been sent and NSIDC is asking the producers for documentation. Analysis continues while that reply is pending; the absence of a matching proprietary manual does not prevent using the independently decoded measurements with stated processing cases.

The same evidence characterizes the six prepared stretches from embedded GPS: median heights 6.35–11.93 km above mean sea level and median ground speeds 482–937 km/h. These are not whole-flight or above-terrain summaries. Earlier candidate windows cover both low and high altitudes. Slower flight reduces the predicted curvature rate, while a geographically fixed background field is independent of aircraft speed. Instrument streams and uncertainty must remain distinct.