METHODOLOGY, PRIMARY BENCHMARK SOURCES & MEASURED VS. PROXY / ASSUMED DISTINCTIONS
1. Pinned Repository Evidence & Collection Commands:
Evaluated against main at 6a3f67671a974703590efde5ae2a323f3f7471fa (window 2025-11-23 → 2026-11-01, 343.4 days):
tokei --output json (static line census); sqlite3 .codegraph/codegraph.db (56,257 AST nodes / 268,598 edges);
git rev-list --count 6a3f67671a9 (4,190 commits, full history).
PR merges are first-parent commits whose subject ends in (#N), read from
git log --first-parent 6a3f67671a9 --format='@%cs %s' --shortstat; PR size = insertions + deletions from that shortstat, with no files excluded.
Windows filter on committer date (%cs ≥ start date); median is the midpoint average and p90 the inclusive decile.
All history: 2,421 PRs. Since 2026-05-10 (175 d): 2,418 PRs, median 225, p90 1,008, mean 664.8, 31.6% >400, 10.1% >1,000.
Since 2026-09-02 (60 d): 861 PRs, median 290, p90 1,485, 40.1% >400. Since 2026-10-02 (30 d): 387 PRs, median 325, p90 1,782, 44.7% >400.
June 2026: 833 PRs, median 188, p90 625, 24.1% >400.
Commit counts per window use the same committer-date filter on git log 6a3f67671a9 --format=%cs.
Reverts = subjects matching ^revert (case-insensitive): 4 (7 subjects mention “revert” anywhere).
Fix share = subjects containing BUG- (1,163) plus subjects starting with fix and not containing BUG- (163); subjects such as [AI:x] fix(...) are not matched, so this undercounts fixes.
The 30-day churn and code-evolution figures come from git log --numstat --since="30 days ago" run on 2026-11-01 (795 commits).
The .aida/review-verdicts/**/*.json snapshot (3,824 files across 2,959 targets) is local, untracked state (no files tracked at the pinned commit): verdict counts, review rounds, and findings per verdict cannot be reproduced from the commit and are not evidence of signatures.
2. Primary External Reference Sources & Caveats:
•
DORA Four Keys: dora.dev/guides/dora-metrics/ (2023–2024
Accelerate State of DevOps production deployment & incident definitions).
•
LinearB / GitClear-style benchmark figures: pickup time, cycle-time phases, <21d rework bands, unreviewed-merge rates, PR-size bucket distributions, and oversized-PR shares are
not traced to a specific published report in this audit. Each is marked
* and is an illustrative reference range only.
•
Review size: SmartBear/Cisco Systems peer-review case study (Jason Cohen,
Best Kept Secrets of Peer Code Review, 2006; 2,500 reviews) recommends reviewing no more than 200–400 LoC at a time.
•
Proxies, Unknowns & Hypotheses: Actual production deployment frequency, deployment CFR, and incident MTTR are
Unknown (no deployment/incident dataset is linked in the repo). The 27.7% <21d churn figure is a
file-level commit-window proxy (files touched multiple times within 21d in
git log --numstat), not line-by-line
git blame survival. Figures marked
* are illustrative: 8-developer squad scaling (~0.2–0.75 PRs/dev/day), architectural heuristics, or industry ranges without a cited report. PR pickup latency, CI settlement time, commit-to-merge time, and total cycle time are
unmeasured in this audit; any values shown for them are labelled design claims or estimates. Target KPIs in the action plan are assumptions, not forecasts. Co-occurrence between >400-LoC batches, hub files, and churn is an architectural hypothesis rather than proved causation.
