Next dataset revision: complete the provenance indices
Completed in development version 2. The design below records the metadata checks; it is not a pending execution queue.
Development version 1 is sealed. Its checker must continue to verify; additions belong in a separately identified version. This work uses saved metadata only, with no downloads, original-file decoding, new fits or prediction evaluations.
Link every numerical-history row to its origin
The 791 history records have sixteen published origin tags. Eleven embed their
source-study directory after completed_; their manifests are already indexed
by the current findings receipt. The five earlier aliases need explicit mappings:
| Origin label | Source stage |
|---|---|
original_first_pass |
All qualifying stretches: 318 original model/case attempts. |
qualified_subset |
Position-stability qualification: 24 attempts. |
completed_wider_recovery |
Full position recovery: 159 attempts. |
qualified_paired_precision |
Paired precision qualification: 16 attempts. |
completed_cost_precision |
Cost-precision recovery: 86 attempts. |
Preserve each original tag and worker identity. Join by source stage, role, stretch, processing case, model and identity where available; do not invent a worker identity for legacy blanks. Source manifests, completion hashes and published results must substantiate the mapping. Conditional profiles and fixed-endpoint checks retain their operation labels and cannot become full fits.
Record exact numerical environments, thread settings and source/kernel hashes from each origin. Different freezes may legitimately use different allocations; never assign the current environment retroactively. Missing evidence stays explicit rather than being filled from a neighboring study.
Reconcile acquisition and retention
Start from the 326-file catalog in the physical-validation catalog snapshot and
the named acquisition inventory in data/ilvis0-ready. Combine the documented
screening outcomes and duplicate-cleanup journal with the 232 retained-file
eligibility rows. Follow existing task IDs, filenames and stored hashes.
Every catalog entry needs one sourced final storage disposition and a separate analysis-eligibility disposition. Retention, physical conversion, current four-minute/high-speed eligibility and independent-flight identity are different questions. Earlier counts of 233 originals remain historical; verified duplicate removal explains the current 232. Do not delete or reacquire anything to build this index. A missing hash or unresolved ledger relationship remains unknown.
Verify and publish
Tests must reject ambiguous joins, changed hashes and mistaken operation classes, preserve failed/interrupted rows, and keep all 53 selections unchanged. Reconcile catalog, retained-source, selected-stretch and origin counts from the actual ledgers before publishing a version 2 snapshot.
This metadata revision does not establish predictive adequacy, calibrated uncertainty or independent replication. Those remain separate gates in the roadmap.