Every Claim Carries A Source And A Verdict

The brief set a target of 70 to 100 validated sources. Underneath that target sat 1 rule, and the rule outranked it: every claim carries a source plus a verdict, and a source is never fabricated. 1 line decides more than it looks. A claim that cannot name where it came from is not a claim yet. It is a draft line waiting on 1 of 3 outcomes - a primary read, a correction, or deletion - and the file is not allowed to pretend otherwise while it waits. The ratio matters more than the number. 100 sources with 40 of them unchecked is a weaker base than 70 sources where every 1 has been read at the primary level, because the 40 do not carry weight, they carry the appearance of weight. The engine counts the 2 apart for that exact reason. So the target was never the achievement. 70 to 100 sources is a modest number. It becomes load-bearing only when 1 rule sits above it and refuses to let a claim in without its provenance attached. 1 probability belongs at the top. The probability that an unchecked source is later found to support the claim it was attached to is close to a coin toss. The probability that a source read at the primary level does is high, and the gap between those 2 numbers is the whole reason the engine refuses to average them together.

The knowledge base is 4 files rather than 1, and each 1 has a different job. KNOWLEDGE.md holds the claims, cross-linked to each other. VALIDATION.md holds every source as checked. COMPENDIUM.md holds the final distillation at no more than about 4,000 words. _STATE.json holds the machine's own state - the iteration, the novelty counter and the open questions. 1 file would have made the first 2 fight. A document cannot be append-only evidence and an editable summary at the same time, and a summary that quietly absorbs unverified material is indistinguishable from evidence once the 2 sit inside the same document. 4 files also make an accident visible. When a vault-wide reorganisation moved the brief and 3 of the files into a library folder, _STATE.json was deliberately left where the film lives, with a note written above the brief explaining why. A duplicate would have forked 8 iterations of validated research. 1 file in 2 places is 2 engines. 2 engines give 2 answers to the same question and 0 ways to tell which 1 ran. 1 probability closes it. The probability that a duplicated state file silently diverges from its twin is high once both are being written to. The probability that anyone notices before the 2 disagree in public is low, which is why the file count and the folder map are part of the method rather than housekeeping. 4 files, 1 engine, and no second copy of anything.

1 bounded iteration runs per run. That iteration picks up to 4 open questions, film-critical 1s first, verifies them against primary sources, appends the claims that survive, updates the source ledger, closes what has genuinely been answered, and opens new questions only when they are genuinely novel. The engine opened with 8 film-critical questions that had to be settled before a single line of script: the most defensible definition of motivation; why humans uniquely act on imagined futures with no immediate payoff; whether winning begets winning and inside which boundaries; what survives the replication crisis about discipline and willpower; what re-ignites a person after defeat; whether the frustration-equals-expectation-minus-reality line is supported by evidence; which channels of winning-begets-winning are actually causal; and whether getting inspired can be triggered deliberately. 8 questions is more than 4, and that is the point of the cap. The cap forces the ordering decision to happen in advance, while the material is still unread, rather than in the moment, when whatever was read last tends to feel most important. New questions are opened only when genuinely novel. A question already answered under another name gets closed rather than reopened, and the counter says so on the record. 1 probability sits under the cap. The probability that an unbounded iteration drifts toward whichever question is easiest to answer is high. The probability that a 4-question cap drifts is lower, because any drift would have to happen inside a budget that is already spent.

Every source is checked into VALIDATION.md, and every claim gets a verdict. 3 verdicts exist: supported, partially supported, or flagged as interpretation. No claim sits in the file without 1 of the 3 written beside it. 2 of the checks are the 1s most engines skip: the verdict field, and the honest flag on the interpretation. Skip the verdict and the checkmark becomes decoration. Skip the interpretation flag and a reading gets quoted to a viewer as though it were a citation. The brief names the 2 places where it expects the uncomfortable answer. The 10,000 hour claim and the willpower finding are both cases where the popular version and the primary literature disagree, and both are written into the file as they are rather than smoothed into a version that suits a film. That is what a verdict buys. It does not make a claim stronger. It makes the strength of the claim readable at a glance, so that a script writer 4 months later cannot mistake a partially supported claim for a settled 1. 1 probability closes the section. The probability that a claim with no verdict gets treated as verified by the next person who opens the file is high. The probability that a claim with partially supported written beside it does is low, and 1 word is the entire cost of that difference.

70 to 100 sources. 1 rule sitting above all of them: no claim without a source.

A body of practitioner material sat beside the science as background triangulation only. It was never named, never cited and never quoted on screen. The purpose was not to ignore it. The purpose was to test it. So a folk claim was read, checked against the primary literature, and then carried as its mechanism with no attribution attached. The line frustration lives in the gap between expectation and reality is the worked example: read it against expectancy-disconfirmation, against affective forecasting, against the shoulds that sit at the centre of REBT - and if it survives, the mechanism goes into the film and the name stays out. That rule does 2 jobs at once. It stops an unsupported claim from being carried into the script on the strength of the person who said it, and it stops a supported claim from being thrown out for the same reason. Both errors are 1 error made in opposite directions: judging the claim by its source rather than by its evidence. 11 seams were mapped as territory before the first source was read: what motivation is, learning and inspiration, execution, purpose and meaning, mastery, communication, the hormonal machinery, past experience, contagion and competitiveness, goal achievement and discipline, and the practitioner material as its own seam. 1 probability sits under all of it. The probability that background material leaks into the final text uncredited is high when no rule forbids it. The probability that it leaks under an explicit translation-only rule is lower, and the rule is written down for exactly that reason.

The engine does not stop when it feels finished. It stops when the counter reads 3 consecutive zeros. The novelty counter ran [1,1,1], then [1,1,0], then [1,0,0], then [0,0,0]. 3 consecutive iterations, each 1 searching for a gap and finding nothing new. Only then did the engine write the compendium and set its own status to reached. The signal prints. The run stops. Never keep burning effort on a question that has already been answered. The base it closed on: 127 sources, 119 of them read at the primary level, 161 claims, 11 construct-complete seams and 0 open questions. 8 of 8 film-critical verdicts passed. COMPENDIUM.md at 30,222 B carried the whole distillation inside about 4,000 words. The same order shows up in the work the method was pointed at. Companies built across 12 countries. A EUR 75 million industrial group restructured. 210 energy systems deployed across Africa and Asia. 200 wood-gasification machines built across the UK and Europe. 16 years of field work from rural Nigeria to post-conflict Serbia, Kazan and Fukushima. A book of 393 pages published on 8 April 2026. 25,000 applications arrive for 25 places in a cohort, and 210 alumni across 19 countries now run the same order in their own work. In every 1 of those, the ledger came before the story. 70 to 100 sources is not a large number. It is a small number made load-bearing by 1 rule - every claim carries a source plus a verdict, and a source is never fabricated. 1 last probability. The probability that a finished research base is really finished is only as good as the counter that says so, which is why the counter is 3 zeros rather than 1. So the next step is small and it is yours. Take the 1 claim you repeat most often about your own work, and write its source beside it. Which 1 of your claims cannot name where it came from?