2026-09-07, MacBookPro. Authorship traced through tool-call timestamps. Fix built, wired and tested.
The chat that opened with go thru active chats and sum up last 4 paragraphs sent i bullet points published the board at 14:11:29 CDT and republished a corrected one at 14:23:24. One session, both versions. Its build script carried every number as a hardcoded Python literal against invented ids. It first read the transcript store at 14:19:47, eight minutes after publishing.
Twenty-three chats moved today. Two were building boards. I read every tool_use whose input named the route and stamped each write and deploy.
| Time CDT | Session | Event | What it produced |
|---|---|---|---|
| 14:10:42 | 5ed46212 | wrote build_board.py | 15 rows, counts as literals |
| 14:11:24 | 5ed46212 | rendered the board | the version with 9 impossible counts |
| 14:11:29 | 5ed46212 | DEPLOY | published to GitHub Pages |
| 14:12 | Brandon | sent the link | the audit begins |
| 14:19:47 | 5ed46212 | first real measurement | Scan every transcript touched today from the on-disk store |
| 14:23:14 | 5ed46212 | wrote build2.py | 22 rows, real ids, measured counts |
| 14:23:24 | 5ed46212 | DEPLOY | the clean board now live |
The other builder, 9acf1afe, is a thirty-minute loop that mentions the route 295 times and never wrote to it. It is not the author.
This is the head of the script that produced the bad board, quoted from the session transcript.
CHATS = [
{
"id":"local_2d4ad288-de99-4ca3-b9e6-c5cee04c2349","title":"Songs lickable text to Songsterr",
"last":"2026-09-07T18:49:09Z","msgs":623,"state":"done",
...Two separate failures sit in those four lines.
"msgs":623 was typed. Nothing in the script opens a transcript. The real assistant count for that chat is 193.2d4ad288 exists in the transcript store. The string appears in exactly three places on this Mac: the session that wrote it, the loop that read it, and my audit.{"fid":"bf53c889","t":"13:49","n":193}, which is real. The numbers are still hardcoded. The difference is that this time the session measured first and pasted the result. Nothing structural stops the next build from regressing.| Gate that ran | What it read | Result | Why it could not catch this |
|---|---|---|---|
| ai_tells_gate.py | prose style | PASS | A fabricated number is well written. |
| session_fraud_check.py | language patterns | passed the counts | It reads phrasing, never arithmetic. |
| curl --retry poll | http status | 200 | A page of wrong numbers serves perfectly. |
All three answer is this well formed. None answers is this true. That is the hole.
A transcript only ever grows. That single fact turns a fuzzy judgment into arithmetic, because the direction of an error carries its meaning.
| Direction | What it means | Severity |
|---|---|---|
| claimed below measured | the chat kept working after the snapshot | warn |
| claimed above measured | growth cannot explain it, so it was not read | hard |
| idle staler than mtime | the chat wrote after the snapshot | warn |
| idle fresher than mtime | claiming a file moved when it did not | hard |
No configuration, no thresholds to tune, no judgment call. It separated the bad board from the good one on the first run.
~/.claude/skills/claim-provenance/. Run it on any page that reports per-item numbers.
python3 ~/.claude/skills/claim-provenance/verify_claims.py <file.md>| Check | Catches |
|---|---|
| C1 orphan-id | An id-shaped token resolving to no file. Catches local_2d4ad288. |
| C2 ceiling | A count above the physical maximum of its source. Caught all nine. |
| C3 repeated-value | One value shared by three or more rows in a field that must vary. Caught 1h 20m. |
| C4 remeasure | Every id to count and id to idle pair, re-derived and diffed by direction. |
| C6 uniform-offset | Every row wrong by the same amount, which means the build stamp is wrong rather than the data. |
It also enforces a format. A row carrying a count must name the source that proves it, and a row with a number and no id is a hard fail because nothing on that page can be re-derived.
- **Status:** X | **Idle:** 1h 13m | **Replies:** 242 | **Transcript:** `30e6208e`Memory says there is one gate system and never to build a second, so this went into ~/.claude/skills/fifty-gates/ as gate 52 of 52, wired to the existing Stop hook.
G52 number-without-measurement, hard, session scope. It fires when a session deploys a page of per-entity numbers whose cited sources it had not read at deploy time. The skill checks the artifact. The gate checks the order of operations, because a correct number can still be a lucky guess.
overflow.sh inside an AskUserQuestion preview I had written, and counted my own hypothetical as a real deploy. Deploy detection is now restricted to Bash commands. Second, a loose source pattern let any incidental file read count as measurement. It now requires the specific ids the page cites.$ python3 run_gates.py --self-test
G52 [no source id on any row] fires OK
G52 [cites an id it never read] fires OK
G52 [cites an id it did read] stays quiet OK
G03 fires OK
G45 fires OK
G11 fires OK
self-test: 6 passed, 0 failed
$ echo '{}' | python3 run_gates.py ; echo $?
0 # runner healthy, gates not silently dead| Fixture | Expected | Got |
|---|---|---|
| the real 14:11 rows | hard fail | exit 2, 13 findings |
invented local_* ids | hard fail | exit 2, 3 findings |
| the live board | no hard fail | exit 1, warnings only |
| the guilty session, replayed | no false fire | none fired |
| this session | no false fire | none fired |
5ed46212 today fires nothing, because G52 reads the route's file as it stands and the bad version was overwritten twelve minutes later. At Stop time, when the gate actually runs, the file on disk is the one just deployed. The self-test fixtures are what prove the firing behaviour, not the replay.overflow.sh.public/.