Search, trials, and deflation

The evidence for a selected result depends on how many alternatives were tried before it was selected.

The selection problem

Suppose 500 independent noise strategies are tested and only the best Sharpe ratio is shown. The winner is not a one-in-one discovery. It is a one-in-500 selection, and its evidence must be evaluated against that larger search.

Penrose records:

• search_cohort_id — which registered search produced the claim;
• search_denominator — the size of that search;
• strategy family — which attempts share a comparable hypothesis space;
• parameter-grid breadth and additional robustness trials.

Why metadata must be preregistered

The denominator must be known before outcomes are observed. Reconstructing it from surviving claims after the fact is biased because discarded candidates are precisely the records most likely to be missing.

If historical metadata cannot be recovered from an authoritative record, Penrose refuses to invent it. That can leave a fresh corpus sparse, but the alternative is scientifically worse.

Family scope

Unrelated hypothesis families should not indiscriminately penalize one another. Penrose scopes trial accounting to declared and normalized strategy families while preserving a conservative floor for external claims.

Generated searches

A generator registers its full budget before emitting candidates. Every slot counts, including failed, discarded, and unreconstructable candidates. Emission is not admission, and admission is not a positive verdict.