IQM Garnet Depth-10 Sign-Replication Preregistration

Date: 2026-07-22

Committed BEFORE any of its own hardware data exist. Execution requires the readiness gates below plus an explicit owner per-submit GO (granted in session for this campaign, recorded here for the audit trail: owner directive "priprav a zadokumentuj všetky dokonale a následne spusti FU-1").

Motivating (already-collected) evidence

Across the executed powered block and depth-profile campaigns the pooled relative parity-leakage asymmetry on layout [2, 7, 12, 13] is positive at depths 4, 6, 8, and 12 but negative at depth 10 (−0.049; absolute even-minus-odd difference −0.0237 at 4,096 shots per arm). The sign pattern (+, +, +, −, +) places the negative point between positive neighbours, which suggests an execution-window fluctuation — but the depth-profile campaign's committed per-depth power (±0.076 relative MDE) cannot decide this: |−0.049| is below that MDE.

Does the depth-10 negative sign replicate in a fresh execution window at a power matched to the observed effect size?

Either answer is publishable: replication revives the zero-crossing reading of the depth profile; failure to replicate (bounded null) strengthens — without proving — the window-fluctuation reading.

Circuit Matrix

Field	Value
Device	IQM Garnet (20 qubits, square lattice) via Resonance
n	4
Circuit family	committed iqm_dla_pinned_n4_d10_{even,odd} builders (bit-identical to the powered block)
States	0011 even, 0001 odd
Depth	10 only
Repetitions	8 per state at 1024 shots → 8,192 shots per arm
Readout states	0011, 0001, 0000, 1111 at 2048 shots (fresh window)
Layout	[2, 7, 12, 13] pinned; fallback [9, 4, 3, 8] recorded before submission if unavailable

Circuit count: 16 main + 4 readout = 20 circuit executions. Shot count: 16,384 main + 8,192 readout = 24,576 shots.

Batching disclosure (frozen): all 16 main executions are submitted as ONE batched job and the 4 readout circuits as a second job (cost discipline: ~2 jobs). The 8 repetitions are therefore execution-order replicates within one queue slot, NOT independent windows; the cross-window comparison is provided by the 2026-07-21 d10 data themselves. The per-repetition drift table is still reported.

Endpoints and Decision Rule (frozen)

Power (committed before execution): at 8,192 shots per arm the one-sided minimum detectable absolute even-minus-odd difference at α = 0.05 and 90 % power is 0.0228 — matched to the observed −0.0237 (the depth-profile campaign's 4,096-shot arms could not resolve this).

• Primary (sign replication): pooled across the 8 repetitions, one-sided two-proportion z-test of leak_even < leak_odd at depth 10 (i.e. the NEGATIVE asymmetry replicates), α = 0.05.
• Rejects → the depth-10 negative sign is real in this window too; the zero-crossing reading of the depth profile revives.
• Fails to reject → bounded null at 90 % power for the observed effect: no evidence of a negative asymmetry at depth 10 in a fresh window. Combined with the positive signs at depths 4/6/8/12 this STRENGTHENS the window-fluctuation reading, but does not prove it; the report says exactly that.
• Secondary S1: two-sided test at depth 10 (direction-agnostic) with Wilson 95 % intervals per sector.
• Secondary S2 (cross-window): comparison of this window's depth-10 even-minus-odd difference with the 2026-07-21 value (two-proportion z-test on the difference of differences), cross-window caveat attached.
• Secondary S3: mean total leakage versus the 2026-07-21 value 0.469 (equilibration consistency).
• Per-repetition drift table (execution-order replicates, as disclosed).
• Raw-count treatment identical to the executed powered-block primary; readout calibration circuits committed alongside.
• No coherent-dynamics claim in any branch (exact statevector baseline pins noiseless leakage at zero; AUD-6).

Live Readiness Gates (block submission until all pass)

• IQMFakeGarnet dry run of all 20 circuit executions from committed code;
• transpiled depth envelope: the May reference (d10 → 159) + 25 % = 198;
• pinned-layout calibration check on the day's live snapshot; fallback substitution recorded before submission if required;
• garnet:mock end-to-end submit/retrieve integration check (zero spend).

Budget and Stop Rules

• Expected cost by the observed batching pattern: ~2 jobs ≈ 2 credits of the ~13 visible; single-block campaign — no further submissions under this preregistration.
• Abort before retrieval only if submission errors leave the matrix incomplete; a partial matrix is reported as such, never silently topped up.

Claim Boundary

Blocked regardless of outcome: quantum advantage; coherence protection; coherent-dynamics parity claims; extrapolation beyond the sampled device, calibration window, layout, and depth; any modification of frozen submissions under paper/submissions/. Results extend the submission_007 draft (or its successor) only. IQM and IQM Resonance are credited in every resulting output.