A Source Line Or It Does Not Ship

The ledger closed on 2026-09-12 at 127 rows. Not 127 ideas. 127 sources, numbered 1 to 127, contiguous, no gaps, no duplicates, every row dated. Each row holds 5 things: the author and the year, the venue, the claim in the knowledge file it stands under, the way it was read, and the day the read happened. The knowledge file is downstream of that table and nothing else gets in. When a claim and its row disagree, the row wins and the claim is rewritten. 3 conventions hold the whole thing to the floor. The date is written in 1 fixed format, so the file sorts itself. The confidence is printed rather than hidden. And the number travels with the claim instead of being summarised into an adjective. 127 rows is not a lot of material. It is enough. A file with 4 rows and 1 flag is worth more than a file with 40 rows and none, because the second one tells you nothing about how any of the numbers were read.

The format is fixed at YYYY-MM-DD in the first row and never negotiated again. That matters more than it sounds. The first check reads 2026-09-09. The run on 2026-09-11 checked sources 109 to 127 in a single pass, with 6 of those flagged as not read at primary level and 2 more carrying a version caveat. The terminal sweep reads 2026-09-12. 11 iterations sit inside those 4 dates, and the order of the reads is the order of the file, so a reader with no context can see what was known on which day. Sort the column and the whole provenance story falls out in 1 keystroke. The date is the cheapest form of honesty in the file and it costs 10 characters. A dated row is also the only way to answer the 1 question every reader eventually asks, which is not whether the claim is true but when anyone last looked.

Every claim in the file carries a verdict mark. 159 claim blocks, 155 source lines, 375 verdicts in total: 292 recorded as supported, 83 recorded as qualified. Qualification is a vocabulary, not a mood. Abstract-level only. Record-level only. Full text read directly. Read at second hand inside another paper's introduction. The flag count was re-confirmed at 8 rows against the row-level marks: 107, 108, 110, 111, 118, 124, 126 and 127. A looser pattern had counted 11, catching the words not and primary sitting near each other on unrelated rows, and the correction came from reading the rows rather than from trusting the count. A claim read only at abstract level is flagged once. A claim resting on somebody else's description of a study is flagged twice and never smoothed over. 8 flags out of 127 is not a defect in the file. It is the reason the other 119 can be quoted.

A number without its sample, its interval and its venue is an adjective wearing digits. The self-control material is the clearest case in the file. The original meta-analysis put the effect at d = 0.62 across 198 tests, interval 0.57 to 0.67. The 23-lab replication then put it at d = 0.04, interval minus 0.07 to 0.15, across 2,141 people. A second replication, 12 labs and 1,775 people, landed at 0.10. All 4 numbers sit next to each other in the table. None of them is optional, because 0.62 on its own is a claim that something works, and the other 3 turn it into a claim about what happened when 3 separate teams tried to reproduce it. Elsewhere the file holds a 150-record meta-analysis over 58,597 people, correlation of .29 between what was expected and what was felt. It holds a competition effect of 0.20 overall, 0.43 measured in the field against 0.08 in the lab, and minus 0.04 for women. It holds 64 men in a rigged one-on-one, where the men whose testosterone rose after losing are the men who chose the rematch. The number travels with the claim, or the claim goes in the bin.

A row is 5 fields or it is a rumour.

Blog posts and AI summaries are refused as claim sources outright. That rule costs the file sources and the file accepts the cost. 2 pages are carried as exceptions, both flagged at the row: a narrative review and a widely-read health commentary, and the pair is used only to evidence the absence of strong primary support for 1 folk claim, never to support it. A third row is the author's own working paper, kept because it reports the design and the findings of a peer-reviewed study that could not be read in full. Then there is what the flags actually caught. An 81 percentage-point figure is attributed with a caveat that the published journal article was not read and the authors' 2005 discussion paper was. A d of 0.68 turned out to be a number quoted inside another team's discussion, where the direct figure in the meta-analysis itself is 0.62. 3 effects, minus 6, minus 4 and plus 8 points, were read in a government evidence summary rather than in the paper's own tables, so the directions ship and the magnitudes wait for a primary read. An 84.7 against 80.1 contrast came from a simulation inside the paper and not from observed people, so it is not described as an experiment. Each of those is 1 line in a table. Each of those lines is the difference between a file you can build on and a file that sounds the same until somebody checks.

At the close of the run the file read 127 rows, 8 flags, 0 orphaned claims and a novelty count of 0 for 3 consecutive ticks. The last 2 ticks did no research at all. They swept 116 named terms against the finished file, got 109 literal hits and 7 zero-hit terms, and adjudicated each of the 7 by reading the surrounding claims rather than by searching again. 1 prohibition came out of that sweep: a reciprocity claim repeated all over the field has no row in the file and may not be written on screen until it gets one. The final distillation is 3,984 words and every claim in it traces back to a numbered row. The probability that an unchecked number survives 3 handovers unchanged is not zero, and most files price that probability at zero, which is exactly how a citation becomes a rumour. The same rule runs outside the file. 210 energy systems across Africa and Asia is a real number and it is worth nothing until the row underneath it names who counted them, on what date, and where the record sits. Build the row first. The claim can wait 1 afternoon.