Development evidence dataset, version 1
Assemble an immutable metadata snapshot of the 53-stretch study. This is an incremental paper dataset, not the final validated dataset or an independent replication. No measurements, selections, numerical results or thresholds change.
Contents
- All 53 selected stretches, original file hashes, UTC bounds, geometry, configuration, selection version, current numerical preference and explicit validation gates.
- All 318 latest model/processing-case primary comparisons, separately from the published numerical history, including failed and interrupted records.
- Conservative split groups and shared-leg identifiers, preserving unknown physical instrument and installation identities as unknown.
- The 232-file eligibility ledger for the retained study pool, including nonqualifying files and aliases. Earlier archive-acquisition exclusions remain in their separate ledgers; this is not a new 326-file acquisition census.
- All 159 raw-gyro baseline comparisons and their declared assumptions; the full fixed-scenario archive is indexed separately.
- A source/evidence index with SHA-256 hashes, including completed, failed and incomplete prediction-method studies. Protected ESA membership is metadata only; no protected measurements or model outcomes are read.
Integrity and interpretation
Verify source receipts before joining. Require matching stretch sets, source hashes, configuration and selection identities. Every latest comparison must retain all three models and both processing cases. Preserve every history row, even blank legacy attempt identities; assign a separate snapshot row ID without claiming it is the original worker identity. Never substitute a repeat for a primary.
Keep numerical preference, descriptive agreement, predictive adequacy, calibrated uncertainty and independent replication in separate fields. The dataset reports no calibrated detections, confidence levels or verified independent-flight count. Do not infer hardware identity from type/configuration bytes. Missing prediction qualification remains missing; earlier development predictions stay indexed.
Build once from named documented inputs. Freeze copied inputs, builder/module sources and generated outputs under this version. A read-only check verifies the frozen snapshot; later changes to live project pages cannot rewrite it. An optional live-input check may identify drift without repairing or regenerating the snapshot. Changed evidence needs a separately identified version.
This metadata-only build performs zero original reads, model evaluations, optimizer starts or network calls. It does not need a new observed-data campaign allowance. The final paper dataset remains gated on withheld prediction, observational error assessment, independent confirmation and dependence-aware calibration.