Development dataset revision 2: acquisition and execution lineage
Extend the sealed first snapshot without changing it or rerunning analysis. All work uses named catalog, inventory, cleanup, selection and numerical-result metadata. No downloads, deletions, original-file reads, native predictions or optimizer starts occur.
Acquisition census
Join every catalog item by task ID, filename and URL with the original 326-entry inventory. Reconcile the 232 retained-file selection rows with the later no-level cleanup journal and verified duplicate-removal journal. Prepared and removed events must agree on source identity. A duplicate's canonical source must be retained and have the same original-byte SHA-256.
Keep initial screening, recorded pipeline storage and current scientific eligibility separate. Absence from the retained ledger is not deletion evidence. Missing disposition evidence or hashes remain unknown. Preserve rounded catalog sizes separately from observed original byte lengths. Supplied historical reference copies outside the pipeline are not part of its storage census.
Numerical lineage
Link all 791 published history rows to their exact publication record, retaining blank legacy worker IDs, failed/interrupted outcomes and operation types. Validate case/model/stretch/role identity, saved cost, convergence label and parameters. Conditional profiles cannot become full fits. Interrupted evaluation counts remain unasserted; their reserved allowance stays visible.
Trace execution manifests separately from publication-review manifests. A review that reads saved results performs no fit and must not donate its environment to the original operation. Use sealed result membership, declared tasks, charged starts or preserved result envelopes to resolve ownership. Reject conflicting owners rather than choose one by order or matching software versions.
Record exact Python/NumPy/SciPy versions, thread settings, source/kernel metadata and manifest hashes. A diagnostic operation can reuse a previously fitted point; preserve that source identity separately instead of treating the check as a new fit. Unknown point ancestry or clock/calibration details stay unknown.
Snapshot and gates
Build a separate immutable version with copies of all named input metadata and builder sources. Preserve version 1 and verify both snapshots. Test identity conflicts, ambiguous owners, tampered hashes, missing deletion evidence, unchanged failure records and publication-versus-execution environment separation.
The original 53 selections and numerical preferences remain unchanged. A complete metadata index does not establish withheld predictive adequacy, observational covariance, independent replication or calibrated significance. Those remain separate roadmap gates.