The engine

This project's own rule library, scored on the same terms as the books it implements.

Home · The measurement · The engine · Provenance · Tests

The engine was measured on the same terms as the books, and it came in below its null.

Publishing that is the point. A project that scores a century of authors against matched controls and exempts its own rule library from the same test has not measured anything — it has built a scoreboard it referees.

What the engine is

A library of rules, each one implemented from a specific page of a specific book, each emitting falsifiable claims: a level, a direction or a date, with a tolerance, priced before the outcome is known. Claims are graded forward, never re-scored after the fact.

What it measured

A 63,990-claim walk-forward. The comparison that matters is pass_rate against null_all (like-for-like) — a like-for-like pair — and the engine comes in at -2.9pp.

And the figure this page used to lead with was one its own report forbids.

FLAGGED 2026-09-07, NOT YET RE-DERIVED. This row published -2.98pp, which is accuracy - null_rate -- and the report that owns it carries the field accuracy_is_not_comparable_to_null: true and a reading note saying in terms: do not read that difference as skill. The like-for-like pair is pass_rate vs null_all, about -2.91pp. The house was publishing the one quantity its own source forbids, twice, as its headline. The sign and rough size do not change; the label does. Awaiting reports/walkforward/, which is not in the restored state.

Fifteen of the eighteen effective series were never measured

A second defect, found on 2026-09-08 and worth more than the headline. The project prices pooled results by an effective sample size — how many genuinely independent series the universe is worth, given that SPY and QQQ are not two facts about the world.

That statistic was computed as measured + 1.0 × (declared − measured): it charged one effective series for every symbol file that was absent. With 6 of 21 files present, the headline 17.8 was 2.8 measured and 15.0 charged by fiat. With all 21 present it is 7.455, entirely measured.

The part that makes it a finding rather than a bug: the constant had two consumers with opposite signs. One made pooled results harder to clear; the other, added three days later and never declared, made them easier. So a placeholder introduced to be conservative was, in net, anti-conservative — pooled rules were priced on up to 2.39× more observations than they had.

No verdict flips under the correction. 50 of 85 scored rules move, 26 of the 31 with positive lift get harder, and the only survivors at either constant are planted controls. The claims that do not pool across this universe — the corn seasonal, the 1909 record, all corpus work — are untouched.

The general lesson

A constant’s conservatism is a property of its consumers, not of the constant.

Nothing about the number changed. A second reader arrived with the opposite sign, and its meaning inverted while the declaration that called it conservative sat unrevised.

The structural fix that follows: a constant declared conservative should enumerate its consumers and assert the count, so that adding one fails a test rather than silently reversing what the constant means. That is proposed here, not yet implemented, and saying so is the difference between a plan and a claim.

What the engine does not show

The forward record is short — a single market episode, graded once — and instruments observed over one overlapping fortnight share the same market-wide move, so the independent observations number far fewer than the rows. A count of rows is not a sample size. Nothing here is evidence in either direction, and it is published so that the next measurement has something to disagree with.