Zero of eighty-one methods
survive their own measurement.
Eighty-one rules mined verbatim from the primary sources were coded, run forward across ninety-eight years and twenty-two instruments, and graded against a chance model built for each claim type. They emitted 1,478,199 unique claims. None beat its own null after family-wise correction. The system as a whole reads 0.36 percentage points below chance.
Source of truth: reports/walkforward/deep_latest.json and reports/confluence/century.json, restated by every dataset published here. 1,468,974 of the emitted claims were graded; 1,464,422 of those were testable (a claim whose chance of passing by luck exceeds one half is discarded as non-evidence). 83 nodes were registered in the pipeline; 2 were structurally silent and emitted nothing, leaving 81 rules with graded claims. Span 1927-12-30 to 2025-08-06.
Everything the library teaches — not just the parts that worked.
240 method records were mined from the corpus, each one a falsifiable sentence with its verbatim quote and page locator. 73 of them became the code behind the 81 measured rule ids. The other 167 are published with the reason they never reached the run — the missing input, the absent phase, the recommendation against wiring. A method inventory of 306 numbered entries records everything the books teach, including what was struck and why.
The whole statement, as written.
The findings document below is published verbatim, exactly as it was written on 2026-08-17. It is the honest state of the project.
Rendered from FINDINGS.md, unedited. The project plan it grew out of is published at GAME_PLAN.md.
The methods were never wrong.
They were never tested.
Now they are. A pure-function grader ruled on every claim with its tolerance declared before the outcome was known. Duplicates collapse. Passes that were likelier than a coin flip are discarded. No number is quoted without its sample size. Under those rules the corpus reads slightly below its own null — and the three rules that briefly survived family-wise correction were each audited adversarially and refuted, the last of them by the fifth error in the ruler.
This is a negative result, published in full. Nothing on these pages is a signal, a recommendation, or a claim that any method predicts markets. A fold with a positive lift is a fold, not an edge.
Headline numbers, read from the files.
| invariant | published | source file |
|---|
Each row is read at page load from the JSON named in the third column. If a value here ever disagrees with the file, the row is marked and the file wins.