Arithmetic can be legislated; capital allocation cannot

The workflow

The project ran as a loop: frozen specification → agent implementation → human adjudication. Decisions that could be reduced to arithmetic — roll rules, instrument-substitution gates, sizing rules — were written down as thresholds before the data arrived. That pre-registration did real work: it prevented in-flight rationalization, the quiet process of inventing a reason to like whatever the data happened to show.

Where thresholds shine

For execution questions, frozen thresholds were exactly right. “Substitute instrument B for A if correlation ≥ x” is a rule a machine can hold faithfully and a human cannot fudge after seeing the result. The arithmetic is legislatable: write it once, enforce it mechanically, record the outcome.

Where they overreached

The rules failed once — by overreaching. A threshold written for execution feasibility briefly entangled the account size itself, and a human had to step in and decouple capital allocation from the rule. A second, smaller case: an instrument rule ended in a one-basis-point tie — statistical noise — and was settled by human judgment on liquidity depth, with both the rule’s output and the override recorded in the decision grid.

The division of labor

The endpoint: machines hold the frozen arithmetic; humans hold the capital and the spec-change pen. A pre-registered threshold is a commitment device for the things that should be mechanical. The moment a rule starts deciding how much money is at risk, it has left the domain where pre-registration protects you and entered the one where it merely removes a human who should be paying attention. Knowing which side of that line a decision falls on is itself the methodology.