Every data-integrity issue we've found in the published record, and what we did about it. Nothing is deleted — corrected entries stay in the archive, marked.
Both were NYSE holidays — no market session occurred. The site published regime reads on these dates anyway. We removed both from session counts, regime tenure, and confidence statistics. The original records stay in the archive, marked invalid.
Four weekend dates with no market session also carried published reads. Two used a stale Friday close duplicated across both weekend days; two used values that don't match any adjacent trading day. Same treatment as above — removed from counts, kept in the archive, marked invalid.
Both invalidation conditions were already true at the moment they were published — not valid falsifiable calls. We excluded both from hit/hold statistics. The original reads stay visible, marked.
June 22 stated a condition that, if crossed, would flip the regime. It crossed on June 23. The regime didn't flip. We're marking this an open exception pending methodology review, not resolving it quietly.
These carried a regime label outside our approved set. We haven't decided whether the approved list needs to expand or the classifier needs fixing, so we're not relabeling them yet. They're marked and under review.
Beyond being an invalid session date (#1), this record's internal regime field disagreed with its own generated summary text. Flagged as a same-record inconsistency.
Full data: correction_ledger_v2.json raw machine ledger
Last updated: September 21, 2026