Method
How an idea earns weight.
Every decision is measured. Each piece earns its place through evidence, and the work learns from every outcome, including the ones it refused. This is how the practice is designed to run. Some of it is built and some of it is being built, and the last chapter marks what is still ahead.
01
Propose
A person proposes what to measure. Candidates that fail are kept, not discarded.
02
Measure
Each candidate is tested on data it has not seen, using only what was knowable at the time.
03
Run aside
A new candidate is designed to run on paper first, where it can be watched and cannot move a decision.
04
Earn weight
Weight is designed to follow evidence that holds, and to leave on its own when the evidence fades.
05
Write back
What was learned, including what was refused, is designed to be written back into the machinery that proposed the idea.
01
Proposing and measuring
An idea begins as a hypothesis and stays one until it is measured.
Candidates come from many directions.
A person proposes what to measure. A measurement is taken on every bar, with no opinion about whether to buy. Nothing is measured only where it was expected to work.
Measurement uses prices that could have been traded.
Returns are taken from a price a trade could really have been placed at, not from the print that closed a bar. A result that assumes an impossible fill is discarded.
Most ideas never earn weight.
Each candidate is measured on its own. A person does not choose which combinations matter, so the only way in is through the evidence.
Refusals are part of the record.
Ideas that were turned down are kept with what happened afterward. A practice that studies only what it did cannot learn the line it chose not to cross.
02
From pieces to a book
Small measurements that survive become families. Families become a book.
A family is several variants of one mechanism.
Each family is named before any result is seen, with a stated reason it might exist. Every variant stays in. Weak variants are not removed after the fact.
The book is what gets tested.
A book holds the families, a combiner that sets weights, a rule that turns a forecast into positions, and a horizon rule. A change to any part makes a new version of the book.
The combiner is walk-forward.
Weights follow how reliably each member has predicted, estimated only from data that came before. No single member is the book.
Families on different clocks meet only in the result.
The way a multi-manager firm keeps each team in its own pod, families that run on different clocks combine at the level of profit and loss, and risk sits above all of them.
03
Earning, holding and losing weight
Proof comes in two stages, and weight is never permanent.
The search is a screen. Confirmation comes once.
A result on data an idea was fit to decides what deserves a real test. The confirmation is one test of a frozen book on data nobody has searched.
Every look at the history is counted.
Many candidates are tested at once, so chance alone will make a few look real. Looks are recorded, and selection is corrected for the number of them.
Weight leaves on its own.
The design is that when a measured edge fades, its weight is reduced at once with no one needing to approve it. Recovery is meant to need a new block of independent evidence, not one good reading.
Capital follows evidence, from the low end.
A positive statistic alone is not capital. Size is to be set from the low end of the evidence and to grow only as forward results arrive.
04
Watching itself
The practice is built to be observed while it runs.
Inputs are watched, not only results.
The practice checks whether the measurements have changed shape, not only whether the latest score is good. A quiet shift in the data is the same problem as a loud one.
The record can be replayed.
The same inputs produce the same result. A conclusion is kept as a record of how it was reached: what was seen, what was refused, and when weight changed.
A failure is loud.
A visible failure gets repaired. A wrong answer that stays quiet corrupts everything built on top of it.
Every parameter has a history.
No threshold or weight is a fixed constant. Each is a versioned record that says who changed it and why.
05 · Being built
Where it is going
This chapter describes work that is being built. It is the direction of the research, not a description of what runs today.
Self-improvement.
What the work learns is being written back into the machinery that proposes ideas, including how often and over what windows a measurement is taken, so each cycle starts better than the last.
Combiner refinement.
How measured candidates are weighed into a book forecast is being refined as evidence arrives, always from past data only and always inside the same gates.
Risk monitoring.
Limits and monitors are being built to sit above the book, with the authority to halt it, and with nothing in the research layer able to override them.
Agentic and evolving research.
The aim is for agentic workflows to drive the research. A population of candidate ideas is generated, tested on data it has not seen, and selected with no person in the loop, by variation, recombination and targeted mutation, inside the same gates as everything else.