Beating always-long at the same geometry and gate is the fifth check

DESIGN §3 makes GOOD SETUP? four refusals, all evaluated, all recorded: the plateau, beats the null, clears the declared floor, leave-one-out (occams/guards/forward.py). On 2026-09-17 the second programme's first question, Q2-001 — after four consecutive lower closes, an S&P 100 constituent carries positive EV over the next bars, inside the up regime of the classifier frozen on SPY — passed all four on the measurement partition: the winner (hold 20, stop 2 %) at EV +0.213 net R over 6,946 trades, 412.7 a year, bounded costs (#19). The Strategy is at FORWARD.

The record appended beside it says something the four checks cannot. M12.6's Shrinkage (#24) runs always-long at the winner's geometry and gate — a market order on every box, long, the same hold, stop, horizon and regime gate, no entry signal at all — on the same partition, and it makes +0.214. The margin is −0.001. On the survey's definition cell the same comparison had given +0.239 (EV +0.256 against always-long at +0.017, #9). Read together: on the sixteen years the verdict was measured on, waiting for four down days before buying did exactly as well as not waiting. What the verdict measured is being long S&P 100 names inside the up regime for twenty bars with a 2 % stop. What the mechanism sentence claims is the streak. The two are not the same hypothesis, and the four checks passed the second on the strength of the first.

Why the four could not see it. Beats-null draws its null as random entry with a coin-flip side — long and short, a seeded coin per box (position_boxed.null_distribution, day_boxed alike) — block- bootstrapped over positions in time order. That null asks better than chance?; a long-only strategy inside a rising, gated partition answers yes without its entry signal doing anything. The floor is absolute (0.15 R, 50 a year) and asks big enough?. The plateau and leave-one-out ask a region, not a spike? and not one name?. None asks better than the passive alternative at the same geometry? — which is the question the survey's trust tiers ask of every cell (held, positive, margin; M12.3, M12.4) and the reason M12.6 records the margin per verdict. M12.6 called the recorded diagnostics "a gate only if the author declares it". This is that declaration, drafted for the author.

Decision (adopted 2026-09-17)

1. **A fifth check, beats_always_long.* The winner cell's EV must exceed the distribution of always-long at the same geometry and gate at the Bonferroni-corrected alpha: random entry *days, long side only, the same exits, stop, horizon and regime gate, a market order on every box, positions in time order, block bootstrap over them — the construction beats-null already uses, with the side fixed to long instead of a coin. The same thinness rule applies: a distribution with too few draws to resolve the corrected alpha is refused, not passed. Refusal text, in the Register's voice: beats-always-long: being long at the same geometry and gate does as well; evidence: p_baseline, alpha_corrected, winner_ev, baseline_mean, draws.
2. The contract carries it. Measurement gains baseline_ev: tuple[float, ...], the Monte Carlo EVs of always-long, beside null_ev. The guards read occams/measurement.py and nothing else (CLAUDE.md); an engine that cannot supply the field is not an engine. Both real engines compute it from the ALWAYS/LONG probe exactly as null_distribution computes the coin-flip probe. The synthetic engine draws it as the means of −cost + σ·z: under no drift its always-long is its null, so make signal (a planted effect at the floor) passes the fifth check and make null (a coin flip) is refused as before. S3 and S10 hold after the change or the change does not land.
3. Same gate, deliberately. For a regime-gated question the baseline is long inside the regime, so the check attributes the effect to the entry rather than to the gate. The gate's own value — whether the classifier times the market — is a different question, one the regime axis has never registered, and this check does not pretend to answer it.
4. Shown passable before alpha moves. question prepare shows the always-long baseline on the definition partition beside the cell's EV — from a survey it already prints the survey's margin and tier; from a Draft it computes the same probe — so the fifth gate meets M12's rule that a gate is shown passable before anything is spent. The survey's gate readiness is unchanged: held and positive already require a positive margin on the definition partition.
5. From the next registration, never backwards. A verdict is a one-time resolution under the checks the question was registered with (ADR-0036; "changing the checks after a verdict is a recorded decision, not a re-run", M6.2). Q2-001 stands as recorded: supported by four checks, with the margin the fifth would have refused written beside it (#24). Re-asking it is a new registration under five checks, costs alpha, and is the author's act; the R4.9 distinction is that the fifth check is now declared.
6. The record names five. forward.CHECKS gains the fifth; HypothesisResolved.refusals carries its sentence when it fires; the passes list, the console card and the programme page show five; a refusal by this check is a null verdict and counts toward the falsifier like any other.

Consequences

• occams/measurement.py (baseline_ev), both real engines and the synthetic engine (the probe and the draws), a new occams/guards/beats_always_long.py, forward.py's tuple, the controls' assertions ("all five checks passed"), DESIGN §3 ("five refusals"), the console's and programme page's check lists. Tests: a fixture whose signal equals always-long at its geometry is refused by name and one whose signal exceeds it passes; the existing four-check fixtures gain a baseline; the synthetic control names five.
• The engine code hash changes; the survey's cells are not recomputed (the survey does not run the guards) and a later grid carries the new hash.
• Alpha is untouched: a check is not a cost. The Shrinkage record stays as it is — the definition margin beside the measured margin — and the check's p-value becomes part of the verdict's evidence rather than a diagnostic after it.
• A supported verdict now attests that the entry beat the passive alternative at its own geometry and gate on a partition nobody read first. That is the claim the mechanism sentence makes, and the claim the forward window (ADR-0014) would then falsify or not.

Considered options

• Keep it a diagnostic (the status quo). The margin stays recorded beside every survey verdict and is not a gate. Rejected in this draft: a verdict would then attest the gate and the side while the sentence claims the entry, and the lab's stated purpose is that the sentence and the test agree. Kept as the author's alternative: nothing is wrong in the Register, and the record is legible either way.
• Change beats-null's null to long-only. Rejected: it changes the meaning of a check four verdicts resolved under, and it loses the two-sided null, which still answers a question worth asking — better than chance in either direction. Two nulls test two things; the Register can say which one refused.
• A point-estimate margin floor (margin ≥ X R). Rejected: a threshold with no significance behind it, and a new author's number. The Monte Carlo at the corrected alpha needs no new number and refuses when it cannot resolve.
• Margin at or above the declared floor. Rejected: it doubles the floor's meaning. The floor is absolute EV, declared for the money's sake (M0.15); this check is attribution, and the two should be refused separately so the Register says which.
• Apply it to Q2-001 now. Rejected: a verdict is one-time; the checks are stamped at registration; and the record already carries what the fifth check would have said. A lab that re-judges a verdict under a rule written after seeing it is the thing every gate here exists to prevent.