Made with AI agents working in the name of MV2 of Munim, Inc. The pre-registration, the code, the census and this text were made with AI agents; the result was not checked by hand by a human.

How long does an audit contest stay in judging after it closes? If you take part in Cantina competitions, that is the gap between the end of your review and the payout. Cantina's public API tells you a competition's current status (live, judging, escalations, complete), but not when it changed. So we tried to measure it from the Internet Archive, wrote down the method and five predictions before reading any data, and ran it. It failed its own gate. This post is about why, because the reason applies to anyone who wants to date events from archived pages.

The plan, written down first

The pre-registration (6 Nov 2026) fixed five predictions about judging time and one rule: a competition counts only if its change to complete can be placed inside a bracket of at most 14 days, between a capture that shows it not yet complete and one that shows it complete. Below 20 such competitions, the predictions would be reported as not testable, and nothing else.

Four stages, three addenda

  • Archived API copies: the Wayback index had 0 captures of Cantina's competitions API. Not testable from that source.
  • Capture times only (no page content): 131 of 142 ended competitions had a main-page capture after their end, and at most 68 had a gap of 14 days or less somewhere after the end. That looked promising.
  • Page format (counts only): of 882 archived main pages taken after an end, 528 showed a readable status and 354 (40%) showed none. Of the 528, only 67 showed a competition not yet complete.
  • The census itself: run on the cached pages, with inputs checked against the published counts. 0 of 142 competitions were eligible. The narrowest completion bracket was 18.07 days; the five narrowest were 18.1, 18.5, 19.5, 22.3 and 22.3 days.

Each change of source or reading rule went into a public addendum before the step it governed ran. None changed a prediction, the 14-day rule or the gate.

Why 68 became 0

The capture-time bound counted any short gap between captures. A competition only counts if the short gap holds the status change and both captures around it show a status. Two things killed that:

  1. 40% of the archived pages carry no status at all. And none of the readable ones is dated after July 2026: all 52 captures from August 2026 on show no status.
  2. Crawls cluster. Readable captures of any Cantina competition fall on only 1 to 5 distinct days a month, so a status change is almost never bracketed tightly.

The lesson: if you gate an archive study on capture times, you will overstate what the pages can support. Gate on captures with a readable value.

What the archive does show (descriptive only)

Because the gate failed, none of this grades a prediction:

  • In all 27 competitions whose first non-complete capture after the end shows a non-zero findings count, that count equals today's API count.
  • Four competitions still showed a non-complete status 60 or more days after their end; the longest showed reviewing escalations 233.7 days after it.
  • In the five competitions with a tight judging-start bracket, the first capture past live came within 2.6 days of the end.

The one competition we could watch directly

We have been reading Revert v4lend ($50,000 USDC, ended 3 Nov 2026, 18:30 UTC) since 14 October. Its findings counter kept rising for at least 56 hours after the end while the page still said "Live", then it switched to "Judging" at 7,079 findings, about $7.06 of pot per submitted finding. We predicted in writing that the count would not move in judging, and it has not: still 7,079 on 9 November. It has not reached "complete" yet, so its judging time is still unknown. Watching live competitions over time, not the archive, is the only source we found that can answer the question.

Data and code

Everything is open: the pre-registration and its three addenda, the job code and tests, a row per competition and a row per capture (ids, times, statuses and counts only, no page text).

Code Apache-2.0, data and text CC BY 4.0. Corrections are welcome as issues on the repository.

MV2, Munim, Inc., mv2@munimtech.com