[svozza]

What this records

Every Grep call failed in the 2026-11-10 large-PR study: 1,456 of 1,456, in all 325 sessions, in the reviewer arms and the plain Claude Code arm alike. The experiment's source-tool shim looked up ripgrep with shutil.which('rg') and no rg binary was on the session PATH. Production uses Claude Code's built-in Grep and was not affected; the October studies were not affected.

The coverage arms were re-run the same day with Grep working (202 sessions, all delivered) and re-judged.

Changes

• docs/findings/0005-large-pr-delivery-and-coverage.md: a Correction section with the cause, the scope, and the re-measured coverage table beside the original figures.
• docs/adr/0022-addendum-coverage-re-measured-with-search-working.md: ADR-0022 is left as written; the addendum gives the re-measured figures and what does and does not stand.

What stands and what does not

• Stands: coverage, not precision, is the large-PR problem (6.1 of 17.1 families per review; 98.6-100% of findings supported). A second pass adds 2.6 families against 1.5 for an independent rerun (difference 1.1, standard error 0.3, 7 of 8 PRs). The skip after an empty first pass holds.
• Does not: the second pass is not separated from two half-file reviewers (0.75 apart, standard error 1.2), and plain Claude Code is not separated from the reviewer (+0.8, standard error 0.65).
• The delivery results do not depend on Grep.

Docs only; no code or prompt change.

🤖 Generated with Claude Code