Method

How a result has to be earned before we call it one. Counts below are read from the registry at build time.

Pre-registration

A study is a frozen folder: a written specification (hypothesis, data, the full grid of variants, costs, the decision rule) and the strategy code, both hashed. The family is registered before any return is computed; the grid fixes how many variants count against it, and every variant is logged, winners and losers alike. A study is never edited after its result: it is rerun from the commit that recorded it.

Hash-chained registry

Every event (family, screen, trial, verdict, sealed opening, forward start) is one line of an append-only file. Each line carries the hash of the line before it, the code commit and whether the tree was clean, so a deleted or rewritten line breaks the chain. The head hash is committed to git; it is printed in the footer of every page.

Sealed holdouts

Data after (or before) a development span is cut into sealed slices that no study can read directly. A slice opens only through the registry, once, and the opening is itself an event: a second look is impossible to hide. Sealed tests so far: 2, verdicts KILL, KILL.

Costs, stressed

Every result is computed at modelled costs and again at double costs. The first check on any screen is the cheapest one that kills: gross edge against the round trip cost. Screen floors (fixed before the screen is run):

• net Sharpe >= 0.5 at 1x costs (book component): risk-adjusted return below the screen floor at normal costs
• gross >= 2x the round-trip cost: gross edge too small against trading costs
• net mean > 0 at 2x costs: loses money at double costs
• matched-null p <= 0.10: not distinguishable from a matched random null
• years to a claim <= 10: too weak to reach a claim within a decade of forward data
• net Sharpe > 0 in each of 1994-2002, 2003-2009 (M2): not positive in both pre-registered sub-periods
• min_cohort >= 5 (declared floor): too few names held
• rank rho <= 0.5 with B/M and S07 (M2): too close to known value and quality signals

Claim tiers

Evidence is an e-value: a sealed-slice p-value calibrated into an e-value, multiplied by the forward log's running e-process; the anytime p-value is one over the largest e-value seen. A forward log also bets against itself and kills its family when that bet grows twentyfold. Tiers (fintech.claim, read at build time):

Silver needs all of:

• survives Holm's correction inside its cohort of 8 slots, at a level that halves with each new cohort
• t-statistic of at least 3
• deflated Sharpe ratio significant after counting every variant tried in the family
• track record at least the minimum length its Sharpe needs
• survives the robustness checks (double costs, sub-periods)

Gold needs Silver plus:

• survives Holm's correction across the whole register
• deflated Sharpe significant against every trial ever logged
• replicated on commercial data when first found on academic (WRDS) data

No study has reached Silver.