Four gates, in order. No idea skips one.

The chain

Four legs, and the one we added late

The fourth leg is younger than the other three. Attribution became a gate on 2026-07-30, after an audit and a placebo failure described below. Every verdict this house issued before that date cleared three legs, not four, and we date the change rather than absorb it.

Gate S

Survivability — guilty until proven robust

The gate runs after grading and treats the winner as a curve fit until it proves otherwise. That inversion is the whole design. An optimiser's job is to find the parameters that scored best on the data it was given; the harder it searches, the better that score and the more of it is luck. The grade-A backtest is therefore not the safest candidate. It is the most suspect one, because it is the one the search worked hardest to inflate.

Every candidate runs unlevered at fixed size: sizing is an L-question, and compounding artefacts have manufactured large mirages out of losing books. The gate is floors-first: seven pass/fail checks, and only a candidate that clears every one earns a continuous score. A large edge cannot buy its way past a fatal flaw. The floors, without their thresholds: the edge is distinguishable from chance; it is unlikely to be a net loser under block bootstrap; it survives execution-cost jitter; it clears a multiple-testing noise floor that rises with the number of trials tried; the tuned winner does not collapse below its own authored defaults out of sample; the winner sits on a plateau rather than an isolated parameter spike; and no walk-forward fold liquidates.

Ranked by how often they cut a graded winner, the top two killers are knife-edge isolation and tuning overfit, which are exactly the two failure modes a shop that deploys its best backtest ships without noticing. The ordering is published; every threshold value and the per-floor kill counts are withheld as calibration.

Which statistics actually predict out-of-sample survival is a separate question, and we measured it on our own corpus rather than assuming it. Taking an archived sweep cohort whose selection timestamps give a clean forward split, 43 of 690 archived candidates reproduced exactly under the current registry; each in-sample gate leg was scored by its discrimination against survival over the first 30 out-of-sample days. Every magnitude statistic we tested (headline grade, raw Sharpe, deflated Sharpe, the permutation p, the cost-jitter check) ranked inversely. Everything that certifies "this edge is large and real" landed below a coin flip. Only two legs ranked correctly, and neither of them measures the size of the edge. The inversion is published, because it is a negative and it generalises. Which gate legs the two are, and the discrimination values of our own instruments, are withheld as calibration.

Then the same measurement humbled itself. Re-run at a corrected, dimensionality-scaled trial budget across roughly 216 graded cells, the leg that had ranked best under the earlier flat budget inverted. Same metric, same computation, opposite sign. A metric's calibration is a property of the pipeline that produced it, not of the metric alone, so any leg validated on a mis-specified budget is suspect until re-validated on the real one. We now re-check a leg every time the budget or the search space moves.

So the gate ranks on consistency and emergence rather than on magnitude. A front-loaded book that made most of its P&L early in the window and then reversed is floored regardless of its headline number: that shape is a rejection signature in our own forward record, and no headline figure overrides it. The floor's parameters, and the size of the forward penalty, are withheld.

Gate P

Specificity — the load-bearing filter

Each candidate faces a control ensemble built to null exactly the mechanism it claims. A basket strategy is re-run on random same-size baskets; if random selections do as well, the chosen basket carried no information. Every control is a full re-run through the same engine, and the ensemble size is pre-registered before any of it is read.

Timing strategies face a different null (a regression of forward returns on the market factor, testing the residual), because a shuffle null destroys beta and mints false positives. Per-window p-values combine by Fisher's method (−2Σln p ~ χ²) and the family-wise verdict applies Benjamini–Hochberg at a pre-registered false-discovery rate, with consistency required across independent windows.

The gate proves itself on planted controls. Every market we enter gets a deliberate factor strategy seeded among the candidates, expected to clear S and die at P. The G10 carry basket did exactly that; the full exhibit is in the 2026-06 working paper Enforced structure ports. Statistical structure does not. If a market's planted control ever passes, every verdict from that market is void until the gate is fixed. That rule has been invoked.

How often P fires is the number the page owed and never printed. Of 265 candidates that accumulated enough independent windows to face it, 19 survived. Roughly nine in ten of the things that reach the specificity gate die there, and each of those candidates was already the surviving parameterisation of a several-thousand-trial search.

A matched control answers the wrong question

A matched control answers whether a book beats other books like it; it cannot answer whether the book beats nothing at all. A calibrated edge-free generator therefore runs beside the control ensemble, and a book must clear both. The incidents that forced that correction, and the doctrine reversal we published with it, are set out in Two nulls, because they answer different questions (2026-08).

Gate L

Leverage — what size can it actually carry?

Most survivors cannot carry even 1x. That is a distribution, not a failure of the books: it is what honest sizing returns when the worst window binds. The distribution itself is withheld.

Books whose relationship is enforced (a rebalance mandate, a defended band, share-class identity, contract convergence) take scenario-sized leverage instead. A bootstrap cannot contain a break that never occurred in its sample, so a discontinuity the size of the largest peg abandonment on record is imposed by hand on any band book rather than inferred from history. Every enforced book also registers a same-day kill trigger on its enforcement mechanism, because the risk that ends an enforced relationship is the enforcer stopping, and that arrives as news rather than as a drawdown.

Capacity is measured with shared legs netted. Two books that trade the same instrument do not add capacity; they divide it. Netting runs across the whole book set before any sleeve is sized, and it always nets down.

The fourth leg

Attribution — did the mechanism earn it?

Beating your controls is not the same as earning from what you claimed. A book must attribute its P&L to the mechanism it named. This became a gate leg on 2026-07-30 for two reasons, both of them ours.

The first was an audit. Of a twenty-book set under review, fifteen were pattern-only: no named enforcer at all, just a relationship that had held. "These series co-move" is a screening observation, not a mechanism, and a candidate that cannot name an obligated party and the trigger that binds them now pre-classifies as a control at authoring time rather than being discovered to be one later.

The second was the placebo that passed. A placebo which shuffles the data but leaves the claimed leader intact does not null the mechanism; it nulls nothing, and it will pass alongside its treatment. So the placebo is now specified against the mechanism, named at authoring, and a mechanism field is machine-readable on every seed. Enforcement alone is also not sufficient: a real, dated, contractual obligation can be fully priced, discharged exactly as written, and pay nobody. We measure the dislocation against round-trip cost before authoring, not after.