Pre-registration — does the GBPUSD pass follow the GBP leg?

Written and committed BEFORE running. Criteria in this file are not edited after the commit that adds it. Any change requires a new file and a new run.

2026-08-27.

The hypothesis under test

On 5 years of Tickstory 5m data, ZMS defaults cleared all eight falsifiers on GBPUSD (PF 1.945, ex-outlier 1.896, median month +131.9, ex-top-3 +5,745) and failed on EURUSD (PF 1.129, ex-top-3 −595), at every H4 grid offset tested. The two pairs are ~0.9 correlated and share the USD leg. The only thing GBPUSD has that EURUSD does not is the GBP leg.

H1 (treatment): the pass follows GBP. Then a GBP cross with no USD leg should pass. H0 (control): the pass is the favourable tail of a family of runs. Then GBP crosses should behave like anything else, and non-GBP non-USD crosses should sometimes pass too.

Instruments — fixed now, none added later

role	symbol	span	status
treatment, primary	GBPCAD	2021-03-25 → 2026-03-24	full 5y
treatment, secondary	GBPJPY	2024-03-21 → 2025-05-09	~14 months, file truncated mid-line — declared UNDERPOWERED in advance
control	EURCAD	2021-03-26 → 2026-03-25	full 5y
control	AUDJPY	2021-03-26 → 2026-03-25	full 5y
reference	GBPUSD, EURUSD	5y	re-run under the new cost model

GBPJPY carries ~14 months against the others' 60. It is reported for information and cannot satisfy any criterion in either direction. The treatment arm is therefore a single instrument, which is a real weakness of this test and is stated here rather than discovered afterwards. EURGBP — still a broken 14 KB stub — remains the control that would settle this best and is still unavailable.

Cost model — changed for this test, and applied to everything

Every previous run in this project used slippage_ticks = 2. The Tickstory files carry the broker's own spread column, and the measured medians are:

	EURUSD	GBPUSD	GBPCAD	GBPJPY	EURCAD	AUDJPY
median spread (ticks)	5	11	37	26	24	10
p90	8	17	53	37	35	16

So GBPUSD's pass was computed at roughly a fifth of its real spread, and crosses are 2–4× wider again. Charging the crosses realistically while leaving GBPUSD optimistic would rig the comparison, so:

slippage_ticks is set to each instrument's own measured median spread, applied adversely on market and stop fills only (limit fills unaffected, per execution.py:_slip). This is deliberately conservative — it charges the full spread on entry and again on a stop exit — and, more importantly, it is constructed identically for all six instruments, which is what a treatment-vs-control comparison requires.

Configuration — fixed now

ZMS defaults, unchanged: H1+H4 bias on, T1 split 50%, R:R gate 1.0, max 3 entries, no session flat, divergence off, Rolling-24 anchor, percentile thresholds. 5m chart, H4 grid offset 3,600s. pyakao.zms.run_zms, conservative fill model. No parameter is tuned in this test.

Sample adequacy

An instrument is UNTESTABLE — and can satisfy no criterion in either direction — if it produces fewer than 200 closed trades, or if run_zms emits any warning (HTF leg warm-up, chart percentile warm-up, H4 grid advisory). GBPJPY is expected to land here.

Criteria

All five must hold for H1 to survive.

C1 — GBPCAD passes all eight falsifiers (Report.verdict(halves=...)["passes"] is True): PF > 1, ex-outlier PF > 1, best month ≤ 60% of net, median month > 0,

50% of months profitable, net ex-top-3-months > 0, both halves PF > 1.
C2 — GBPCAD's largest-net cell has n ≥ 30 and PF > 1, and is the same cell key that carries GBPUSD under this same cost model. The 2026-08-26 finding was that cell roles invert per instrument; if they invert here too, there is no shared mechanism to follow.
C3 — neither EURCAD nor AUDJPY passes all eight.
C4 — GBPCAD passes all eight at all four H4 grid offsets: 0, 3,600, 7,200, 10,800s. This is the sweep GBPUSD survived; the treatment must survive it too.
C5 — GBPUSD still passes all eight at its real spread (11 ticks). If it does not, the phenomenon this test was designed around never existed at realistic cost.

Decision rule — stated before seeing any result

outcome	conclusion
C5 fails	Test VOID. The GBPUSD pass was a cost artifact. Closure stands, strengthened.
C1–C5 all hold	H1 survives its first real test. ZMS is not reopened and takes no capital; it becomes one named forward-observation candidate (GBP crosses, defaults, real spreads) with a stated review date.
C1 fails	The pass does not follow the GBP leg. Closure stands, strengthened.
C3 fails	The pass is not GBP-specific. Closure stands.
C2 or C4 fails	No consistent mechanism carries it. Closure stands.

A mixed result is a fail. "GBPCAD passed but a control also passed", "GBPCAD passed at three offsets of four", "GBPJPY looked good" — each of these is reported as a failure of this test, not as encouragement.

Anti-rescue commitments

No instrument is added, removed, or substituted after results are seen.
No parameter, cost, timeframe, or offset is changed after results are seen.
All six runs are reported regardless of outcome, including the ones that hurt.
The verdict is read off Report.verdict(), not argued.
If this test fails, the GBPUSD pass is closed as a tail and is not re-examined without new data (EURGBP or a full-length GBP cross).

Prior probability I would state now, before running: low. Six modifications, nine instruments and two timeframes have failed this project's tests, every apparent positive so far has been instrument-specific, outlier-carried, single-month, or directional beta, and this hypothesis was generated post-hoc from the very run it seeks to explain.