The Run Ends When The Checks Pass

The run stopped at iteration 11, on 2026-09-12 at 03:10:18+08:00, and the reason is written into the state file as 3 numbers: novelty 0, 0, 0. That field is not a feeling about progress. It counts something specific, which is how many film-critical sub-questions an actual search turned up that had not been answered before. Three consecutive ticks turned up none. The 11-th was the last 1 that could have changed the answer, so it ended there. That is a different event from running out of budget, and a different event again from running out of patience. A run that keeps going after its own novelty reads 0 does not get deeper, it gets longer, and length is 1 of the 2 things a reader can smell from the first paragraph. The closing condition was set before the run started: both halves met. Coverage complete, and quiet. Coverage meant 0 open questions and every named construct present. Quiet meant the 3 zeros. Nothing about the ending was negotiated at the end.

The ledger that came out of those 11 iterations is the part worth copying. 161 claims, each 1 traceable to a numbered row in a validation file that runs contiguously from 1 to 127 with no duplicate numbers and no gaps. 8 of those 127 rows are explicitly marked as not primary-read: rows 107, 108, 110, 111, 118, 124, 126 and 127. That leaves 119 rows read at source, and the 8 are still in the file, labelled, rather than quietly dropped. The verdict marks count 375 in total, and they split 292 verified against 83 interpretive, which is a ratio that tells a reader exactly how much of the argument is standing on a study and how much is standing on a reading of several studies. 159 claim blocks and 155 source lines, with 0 orphaned blocks, meaning no claim floats without the source underneath it. The proportion of a dossier that is primary-sourced is a number most teams never compute, because computing it means publishing 8 rows that say not-primary-read.

Iterations 10 and 11 added 0 new claims and 0 new sources between them, and they were the 2 most useful ticks of the run. What they did instead was sweep. Iteration 10 took 90 terms drawn from the brief's named constructs and searched the whole finished knowledge base for each 1: 82 returned literal hits, and the 8 that returned nothing were adjudicated by reading the surrounding claims rather than written off. Role models came back empty as a phrase and was fully covered under the labels peer models and prestige, 1 source from 1985 and 1 from 2001. Iteration 11 widened the list to 116 terms, got 109 literal hits, and adjudicated the remaining 7 the same way, by reading. 1 of those 7 was missing because the file writes it with an en-dash. Another because the file writes prestige-bias rather than prestige-biased. Text search is a blunt instrument and reading is not, which is why both ran. Iteration 10 also caught drift: the carried total said 125 source rows and the file held 127, so the number was corrected and the correction recorded instead of silently patched.

The engine that did the research set its own terminal signal at iteration 11 and paused its hourly cron. That alone would be an author grading their own paper. So a separate tick, running on a 120-minute schedule, opened the finished compendium and checked it against the brief without doing any research of its own: 3,984 words, the 8 film-critical questions all carrying a header and a verdict, all 11 seams present, film-ready. Only then was the status set to paused. One process declares, a different process verifies, and the second 1 has no stake in the answer. The cron prompt carries the same discipline at the exit: a tick that sees a terminal or paused status prints 1 line and returns, because a finished run must not keep burning tokens. The probability that a crew knows when its own work is done is not zero, and most teams price it at 0. The way to raise it is mechanical: write the closing condition before the start, and make somebody who did not do the work read it before anybody says done.

The run ended at iteration 11 because 3 consecutive searches returned nothing new.

The ugliest finding in the file is not a research finding. It is that both watchdog scripts were dead and nobody knew. The state file carried a timestamp written as 2026-09-10T15:28:19WITA, with a zone abbreviation on the end, and the scripts subtracted that string from the current time, which throws a TypeError in Python and takes the script down with it. 5 consecutive ticks failed between 20:31 and 00:32. 0 alerts fired, because the thing that was supposed to raise the alert was the thing that was broken. The failures underneath were provider-side: HTTP 402, insufficient balance, on all 5 ticks. So the outage was 5 ticks long and invisible, and the reason it was invisible had nothing to do with the provider. The fix is 1 rule, written into the state file so it cannot be forgotten: ISO-8601 with a numeric offset, always +08:00, and never a zone abbreviation. The second rule is a counter that means something: consecutive errors counts research failures, not infrastructure failures, so a bad afternoon at a provider does not read as a bad afternoon for the run.

The compendium ships with its own qualifications attached, and those are the reason to trust the other 3,900 words. 2 corrections happened on film during the run and both went the same direction: an organ-donation default figure was struck off the evidence ledger, and the meaning chapter lost a well-known name because the source did not carry the claim. 1 open question travels with the film on purpose, and it is the honest 1: no study manufactures a want. That sentence is a production constraint, not a footnote, because it means the second act has to be a want-forming beat rather than a reveal, and anybody building the film has to build it differently because of 1 flagged row. The same discipline applies to the room the film gets built in: 6.0Gi free of 228Gi, 97 per cent used, flagged as a gate before any still batch starts. Numbers read out of a file, not intentions read out of a meeting. EX Venture built companies across 12 countries on that order of work, and 393 pages of the book carry the same shape. Say what is not primary-read. The run ends when the checks pass, not when the author is satisfied.