How a verdict is produced
Nothing here is a black box. This page describes exactly what happens between an idea and a score, what each part can and cannot establish, and where the whole thing is weakest.
Last updated 6 August 2026.
The short version
Two things happen on this site and they are different. The first is a DETERMINISTIC READ: every figure about an industry is computed from a local corpus of United States federal public records. It calls no model, costs nothing, takes under a second, and returns the identical answer every time. That is what you get from the free pages.
The second is an UNDERWRITING: a scored judgement about a specific idea, produced by a chain of specialised model stages including an adversarial layer that argues against the idea, answers itself, and rules. That produces a verdict, a score, a confidence, and a falsifiable test recorded before the outcome is known. It costs real money and takes roughly half an hour. Every one we have published is at /analysis.
The deterministic read is the foundation. The underwriting is judgement layered on top of it, and it is always shown with the argument against it.
What the corpus holds
| Source | Rows | As of |
|---|---|---|
| Establishment and employment density | 1,100,961 | 2023 |
| Occupational employment and wages | 136,541 | 2024 |
| Occupational licence register | 27,067 | 2024 |
| Business dynamics, 1978 onward | 13,248 | 2023 |
| Federal contract awards | 957 | 2025 |
956 six-digit industries carry establishment data; 388 also carry occupation data, so the licence gate is answerable for roughly 41% of them. Where it is not, the page says so. Full detail at /data.
How a score is built
Each path has a rubric of seven weighted dimensions. A dimension is scored only where evidence exists for it; where there is no producer, it returns nothing and is EXCLUDED from the weighted total, and the covered weight is published beside the score. A score built on 55 points of evidence and presented as if it were 100 is the failure this system was built to argue against.
The model judges the dimensions. Code computes the total. A model doing arithmetic can be wrong about arithmetic, and the same judgements must always produce the same total.
What the interval means
| Rubric | Interval | How it was measured |
|---|---|---|
| w3.0-20260804 | ±3.1 at 95% | the arbitration stage re-run 10 times on identical frozen input, sd 1.6, measured 2026-08-04 |
| every other rubric | NOT MEASURED | no interval has been established for these, and the one above does not transfer to them |
An interval measures REPRODUCIBILITY — what re-running the same input actually produces. It is not a claim about accuracy. Accuracy is a separate question and it is answered at /ledger, where every prediction is recorded before its outcome and graded afterwards.
Where this is weakest
We have registered 63 predictions and 0 have resolved. Until enough resolve, no honest claim can be made about how often this system is right, and none is made. Every competitor we surveyed publishes an accuracy percentage; not one publishes the record that would let you check it.
Signals are backtested against ESTABLISHMENT GROWTH, which is not the same thing as returns to an operator. An industry growing is not the same as a business in it succeeding, and this is the best proxy the public record offers rather than the right one.
We publish 70 of our own defects at /library/failures, generated from the regression suite rather than written by hand, so a defect reaches that page only by first acquiring a check that holds it shut.