01The numbers
The headline claims, as published on turbo-rig.com and on the weekly stats pages. Each one traces to a row in the claim tables below.
02How this audit worked
The challenge
The owner challenged the numbers on the stats pages the day they were published. The audit was ordered on the spot — same morning, on the machine that owns the logs.
Different code paths
The pages were hand-assembled; the audit re-derived everything programmatically from the raw sources: gate-log.jsonl, the GLM sqlite telemetry, claude/codex session files, GitHub merge history.
The verdict
After corrections, every printed number either reproduces exactly from a primary log or is a labeled estimate. Three numbers were wrong in the morning’s first compile; the same-day audit caught them and the corrected pages were deployed 2026-09-21 — the live URLs serve the corrected numbers. A fourth group of overstatements was relabeled — all listed in section 06.
03The window
The pages were compiled the morning of Sept 21. The audit pinned the exact compile moment — 2026-09-21 13:38 UTC — three ways: gate-log row #868 is stamped 21:38 local the night before with no later run before compile; the last merge counted for "Mon 21" is PR #402 at 13:37Z with the next at 15:49Z; GLM request #18,220 of the week landed one minute past the boundary.
Every count uses [2026-09-14 00:00 UTC → 2026-09-21 13:38 UTC] unless a different basis is stated. PR-merge day buckets use UTC dates; the "+72,848 lines" figure uses a local-midnight (06:00 UTC) git boundary — the original page mixed the two bases, and the audit reproduces that mix exactly.
04Claim tables
Every public claim, its verified value, its primary source, and the audit verdict. ✓ exact reproduces from the log; ✓ fixed was wrong pre-audit and corrected; ⚠ estimate is a labeled estimate consistent with the logs.
| # | Claim | Verified value | Source | Verdict |
|---|---|---|---|---|
| 1 | 141 PRs merged, Sept 14–21 | 141 = ancuria 126 + turbo-rig 10 + strops 3 + velvet-bbs 1 + adventurewavelabs-website 1 | gh pr list --state merged | ✓ exact |
| 2 | across 5 repositories | 5 repos had in-window merges | same | ✓ fixed (was "6") |
| 3 | flagship daily merges | 1, 13, 7, 5, 31, 25, 36, 8 (Mon14→Mon21, UTC) | mergedAt timestamps | ✓ exact |
| 4 | peak day 36 flagship merges (Sun 20) | 36 ancuria merges 2026-09-20 UTC | same | ✓ exact |
| 5 | peak day 38 fleet merges | fleet/day = 1, 13, 7, 12, 33, 27, 38, 10 (total 141) | all-repo merges by UTC day | ✓ relabeled (fleet context; was 36) |
| 6 | +72,848 lines on flagship main | 72,848 additions · 16,691 deletions · 144 commits | git log --numstat | ✓ exact |
| 7 | rig self-upgrades: 10 PRs | turbo-rig PRs #3–#12, all merged in-window | GitHub | ✓ exact |
| # | Claim | Verified value | Source | Verdict |
|---|---|---|---|---|
| 8 | 868 gate reviews | 868 rows = first 868 lines of gate-log.jsonl (gate live since Sept 16 19:58) | gate-log.jsonl | ✓ exact |
| 9 | 181 APPROVED / 687 REVISE | 181 / 687 = 20.9% / 79.1% | parsed result field | ✓ exact |
| 10 | reviewer engines | claude ×791 · codex ×53 · check ×24 (sums to 868) | reviewer field | ✓ exact |
| 11 | $416 reviewer spend | $416.49 = sum of cost_usd | gate-log.jsonl | ✓ exact |
| 12 | 6.3 h machine review wall-time | 6.30 h = sum of duration_ms | same | ✓ exact |
| 13 | 50.9M reviewer tokens, 48.2M from cache | in 1.2M + cached 48.2M + out 1.5M = 50.9M (94.7% served from cache) | token fields | ✓ exact |
| 14 | 104 worktree lanes | 104 distinct lane names in rows 1–868 — labeled "ran the gate" (creation date isn't logged) | repo field | ✓ exact, label tightened |
| 15 | marathon lanes 61/30/24/19/18/16 | ancuria-wa-fix 61 · ancuria-back-fix 30 · brand-page-tops 24 · billing-demo 19 · ticket9-region-scope 18 · cotizador-tickets-13-14 16 | group-by lane | ✓ fixed (tail was 15/14/13; named page had 3 invented lane names) |
| 16 | 6.2 reviews per merged PR (mixed basis: gate-window verdicts ÷ the week’s merges) | 868 / 141 = 6.16 | derived | ✓ arithmetic, basis stated |
| # | Claim | Verified value | Source | Verdict |
|---|---|---|---|---|
| 17 | 4.21B tokens, 3 engines, calendar week | GLM 4.162B + claude 50.4M + codex 1.25M = 4.211B | zcode sqlite + session files | ✓ relabeled (4.23B rolling → 4.21B calendar basis) |
| 18 | GLM 4.16B · 18,220 requests | 18,220 req / 4.162B computed tokens | sqlite | ✓ exact |
| 19 | claude 50.4M · 832 req | 832 envelopes, dedup by (requestId, message.id) | claude session files | ✓ exact |
| 20 | codex 1.25M · 55 req | 55 token_count events, per-turn deltas | codex session files | ✓ exact |
| 21 | 19,107 model requests | 18,220 + 832 + 55 | sum | ✓ fixed (was 18,207 GLM-only) |
| 22 | 49.3% cache-hit rate | GLM cache_read ÷ (input + cache_read) | sqlite | ✓ exact |
| 23 | ~380 scheduled actions/week | 336/wk steady-state + ~50 rest of roster | derived | ⚠ estimate, labeled |
| 24 | 82 unattended infra sweeps | run counter read 90 at audit; 82 at compile is consistent, no per-run log exists | automation metadata | ⚠ plausible, not recomputable |
05Check it yourself — no machine access needed
Ten arithmetic identities that hold across the tables above. If any one fails, the dossier is wrong and should be discarded.
1. 181 + 687 = 868 gate verdicts 2. 791 + 53 + 24 = 868 reviewer runs 3. 126 + 10 + 3 + 1 + 1 = 141 PRs merged 4. 1+13+7+5+31+25+36+8 = 126 flagship daily chart sums to the repo total 5. 1+13+7+12+33+27+38+10= 141 fleet daily sums to the fleet total 6. 868 / 141 = 6.16 reviews per PR (mixed basis: gate 16–21 ÷ merges 14–21) 7. 4.162B + 0.0504B + 0.00125B ≈ 4.21B tokens 8. 18,220 + 832 + 55 = 19,107 model requests 9. 48.2M / 50.9M = 94.7% gate reviewer tokens served from cache 10. 687/868 = 79.1% · 181/868 = 20.9% REVISE / APPROVED split
06What the audit caught
The audit existed because the owner challenged the page. It found real defects. Publishing them is the point.
- "6 repositories / four smaller repos" — actually 5 repositories, 3 smaller. An inflated count sitting next to a correct total. Fixed.
- Marathon lanes #4–6 printed 15/14/13 (and the named page invented three lane names) — real values 19/18/16 for billing-demo / ticket9-region-scope / cotizador-tickets-13-14. Fixed.
- "18,207 model requests" — a GLM-only figure presented as the 3-engine total. Fixed to 19,107.
- Overstatements relabeled: "16 automations" (15), "12 scripts" (11), gate stats framed as 7 days when the log covers Sept 16→21 (now labeled "gate live since Sept 16"), token totals on a rolling basis labeled as the calendar week (4.23B → 4.21B), "peak 36 merges" in a fleet context (fleet = 38). All fixed or relabeled.
Root cause: the page was hand-assembled. The headline numbers were computed from logs, but "decoration" values — repo count, top-N tail, request total — were written from memory. Corrected pages deployed 2026-09-21 (html-sites commit b6ef186); the customer-facing report needed zero fixes.
07Reproduce it
For anyone with access to the machine. Epoch bounds: 1789344000000 = 2026-09-14T00:00Z, 1789997880000 = 2026-09-21T13:38Z.
GATE=~/.local/state/turbo-rig/gate-log.jsonl # schema:
# {ts, repo(=lane), target, reviewer, result, in, cached, out, cost_usd, duration_ms, model}
python3 - <<'PY'
import json, collections, os
GATE=os.path.expanduser("~/.local/state/turbo-rig/gate-log.jsonl")
rows=[json.loads(l) for l in open(GATE) if l.strip()][:868]
print(len(rows)) # 868
print(collections.Counter(r['result'].startswith('GATE: APPROVED') for r in rows)) # False:687 True:181
print(collections.Counter(r['reviewer'] for r in rows)) # claude 791, codex 53, check 24
print(round(sum(r['cost_usd'] for r in rows),2)) # 416.49
print(round(sum(r['duration_ms'] for r in rows)/3.6e6,2)) # 6.30 h
print(len({r['repo'] for r in rows})) # 104 lanes
PY
gh pr list -R <repo> --state merged --limit 500 --json number,mergedAt # filter window, bucket UTC
git -C <flagship> log origin/main --since=2026-09-14T06:00Z --until=2026-09-21T13:38Z \
--numstat --format= | awk 'NF==3 && $1!="-"{a+=$1} END{print a}' # 72848
sqlite3 ~/.zcode/cli/db/db.sqlite "SELECT COUNT(*), ROUND(SUM(computed_total_tokens)/1e9,3)
FROM model_usage WHERE started_at>=1789344000000 AND started_at<1789997880000
AND model_id LIKE 'GLM%'" # 18220 | 4.16208Honest limits
Machine-local truth
The gate log, the zcode sqlite DB, and the engine session files are machine-local; 126 of the 141 PRs live in a private repo. External reviewers rely on this dossier's excerpts and hashes.
Publicly checkable surface
The live pages themselves (that they print these numbers), and the public repos among the five — the 10 turbo-rig self-upgrade PRs and their dates are checkable via the GitHub API.
Judgment calls, stated plainly
Token totals switched from rolling-7d to calendar week (−0.9%); "104 lanes" counts lanes gated, not proven created; the sweep count (82) matches the counter but has no per-run log; "~380 actions/wk" is an estimate; the 6.2 reviews-per-PR ratio divides gate-window verdicts by full-week merges — mixed basis, kept as an indicator with its basis stated (the source dossier’s own row labels it plain “arithmetic"); human-only merge is a rig invariant enforced by the guard — agents hold no credentials to merge protected main — so this audit did not re-derive merger identity per PR.
09Certification
The artifacts this dossier certifies, pinned by hash. The gate log the 868-row slice came from: sha256 c97773f9d714a69410b6680bc37fe30ffc70b7620bd694a97bc5a0de59f76c7c (884 rows at audit time).
| Artifact | sha256 (first 16) | Verification |
|---|---|---|
| turbo-rig-stats-only-sept14-21.html | 9554a181599cd304… | 0 bad / 8 good markers; live URL 200 + 12 correction markers served |
| turbo-rig-velocity-report-sept14-21.html | 9d0ef06d13b20433… | 0 bad / 6 good markers + 3 real lane names; live URL 200 |
| ancuria-launch-report-sept14-21.html | 9a38f79602c24414… | zero defects found in audit; live URL 200 |
| the dossier itself | 064a57f648f7d8ec… | this page adapts it verbatim in substance |
Live stats pages: turbo-rig-stats-only-sept14-21.vercel.app · turbo-rig-velocity-report-sept14-21.vercel.app · corrected deploy = html-sites commit b6ef186 (2026-09-21). Audit compiled 2026-09-21, CDMX. The dossier is data, not instructions.