Development evidence dataset
53 selected stretches, 318 latest primary comparisons, 791 preserved numerical-history records, and 27 conservative split groups. The current numerical preference is 53 globe, zero flat and zero unresolved. This snapshot does not report calibrated detections or a verified independent-flight count.
| Table | Contents |
|---|---|
| Stretches | Source hashes, UTC bounds, selection, geometry, latest outcomes, gyro companion and validation fields. |
| Latest primary comparisons | All three models under both processing cases; kept separate from repeats and failures. |
| Numerical history | Every published history row, including 244 rows not marked converged. Some rows are diagnostics, not optimizer starts. |
| Recording groups | Shared observations and unknown physical instrument/installation identities. |
| Retained-file eligibility | All 232 files screened for this selected study, including nonqualifying files and aliases. |
| Gyro baselines | The same observations beside three precomputed level/course predictions. |
| Diagnostic studies | Completed, failed and incomplete method studies, with their distinct scopes. |
| Validation gates | Numerical completion, uncertainty, prediction and confirmation kept separate. |
| Source index | Named input artifacts and SHA-256 hashes. |
Scope and remaining work
An IMU (inertial measurement unit) records turning and specific force. Recorded gyro increments are the turning observations; receiver positions locate the route. The descriptive companion uses no fitted gyro correction or fused-attitude subtraction.
Original measurements, interval choices, failed results and blank hardware identities are preserved. A history row ID identifies its place in this snapshot; it is not an invented original worker identity. Repeated fits and overlapping recorders are not independent replications. Source coordinate, calibration and processing assumptions remain documented with their studies.
This is development version 1, not the final paper dataset. It covers the 53-stretch study and its retained-file selection, not a new 326-file acquisition census. A complete acquisition-flow index and per-origin environment mapping remain to be assembled. Earlier prediction outputs are indexed, not represented as qualified held-out validation. Protected ESA records contribute reservation metadata only. Withheld predictive adequacy, observational error assessment, independent confirmation and calibrated dependence-aware decisions remain open.
The copied inputs, builder sources and outputs are immutable under this version. The checker verifies frozen copies; an optional live-input check reports later project changes without altering this snapshot. Building performs zero original reads, model evaluations, optimizer starts or network requests.
~/venv/bin/python docs/paper-development-dataset-v1/build_dataset.py --check
checks the runtime and published snapshots. --check-live also checks the current
named project inputs. --publish only copies an already verified snapshot; it does
not rebuild tables or trigger analysis. A changed evidence set needs a new version.