# PRI corrections log

## 2026-09-09 — method v0.2.1 interval reproducibility

Reporter: internal reproduction check.

The September 5 reproduction of method v0.2.0 found changed time_to_award
median interval endpoints in 404 office-years, with a maximum endpoint change
of 22.5 days. Seeded bootstrap sampling depended on unordered input rows.
Counts, point estimates and amendment values were unchanged.

Method v0.2.1 sorts each office's actions and elapsed-time vector by
(action_date, contract_transaction_unique_key) before the 1,000 seed-0
bootstrap resamples. Permutation tests and a complete CSV-subset-to-release
reproduction test now exercise this behavior, including in weekly CI.

Comparing the new canonical ordering with the original published v0.2.0
release changes intervals in 477 office-year/scope rows; the maximum endpoint
change is 64.5 days. These replacement-release figures are distinct from the
404/22.5 diagnostic above. All non-metadata differences are interval endpoints:
point estimates, populations, amendment values and other measures are identical.
The larger change reinforces that these are estimated intervals, not precision
guarantees.

The original v0.2.0 files remain immutable. They are superseded for interval
computation by v0.2.1; deployment awaits release verification and review.
The correction has a new derived-knowledge as_of. Original source receipt dates
and meta.source_as_of remain separate: this is not a fresher source download.
Full raw-archive reproduction passed: all seven output byte hashes match after
reloading all 13 receipt-listed source archives/CSVs, rebuilding linkage and
recomputing the release. [Full reproduction receipt](/pri/reproduction-v0.2.1.json). Computation commit: 248d4586af593555ec9c042cfd73638ea156fa7c.
