Skip to content

quantum-switch-spec.md

Latest commit

 

History

History
261 lines (218 loc) · 30.4 KB

quantum-switch-spec.md

File metadata and controls

261 lines (218 loc) · 30.4 KB
Raw

The Quantum Switch — Full Apparatus Specification

Author: Whisper (DC15W), C4559 (2026-07-11), Creator-directed ("write the spec doc"). Purpose: single-document engineering spec of the quantum switch as we build, measure, and grade it — circuit family, exact theory statistics, measured-results ledger, reusable methodology, pitfall registry, and scope. Consolidates what previously lived across six preregistrations, the findings series (F73–F94), and Beyond the Ladder (the repo-native publication). That document (beyond-the-ladder.md) is the narrative for causal-inference readers; this is the spec for whoever builds experiment N+1.

Contents: 1 Definition · 2 Circuit family · 3 Exact theory · 4 Measured ledger · 5 Provable bounds · 6 Methodology · 7 Applications · 8 Scope · 9 Public artifacts · 10 File index

1. Definition

The quantum switch is a higher-order operation (a supermap): given two black-box channels 𝒜 = {Kᵢ}, ℬ = {Lⱼ} and a control qubit, it applies them in an order entangled with the control:

W_ij = Kᵢ Lⱼ ⊗ |0⟩⟨0|_c  +  Lⱼ Kᵢ ⊗ |1⟩⟨1|_c        (control |+⟩, summed incoherently over i,j)

Three properties are load-bearing for everything below:

  1. Decomposition independence — the switch is well-defined on channels: any Kraus decomposition gives the same output. This licenses the pooling estimators (§4).
  2. Linearity in each slot — S(Σₖ pₖ𝒜ₖ, ℬ) = Σₖ pₖ S(𝒜ₖ, ℬ). This licenses classical mixing of circuit variants as an exact channel mixture (Exp106/107/108 estimator logic).
  3. The commute/anticommute readout — for unitaries, measuring the control in |±⟩ computes the predicate COMMUTE(A,B) with one query of each: control unchanged if [A,B]=0, flipped if {A,B}=0. No fixed-order circuit computes this at that query count (Chiribella). This is the sense in which the switch is a logic gate over operations rather than bits (C4557).

Scope statement (applies to every claim in this document): our implementations are device-characterized — definite-order gate-model circuits that reproduce the switch's statistics on characterized hardware. We do not claim indefinite spacetime order (the photonic device-independent program is orthogonal, not competing). See §8.

Platform prior art (C4580, from the review arXiv:2405.00767): non-photonic ICO experiments before this campaign = the ICO-refrigeration protocol on NMR (Nie et al.) and IBM cloud (Capela et al.), circuit-model simulations without witness measurement or bound-referenced grading. To our knowledge no prior gate-model experiment graded against a causally-separable bound; the witness-violation and (semi-)device-independent certification literature is photonic. Claims here are worded accordingly — "first" is never claimed, review lag is real.

2. The circuit family (five variants, all on IBM Heron-r2)

Variant Experiment Qubits 2q cost (routed) Control readout Target readout
V1 witness Exp91 2 ~4 CZ X basis —
V2 padded game Exp105/105b 2 uniform 4-CZ skeleton X basis —
V3 channel-twirl 2-switch Exp106 2 uniform 4-CZ skeleton X (clbit 0) Z (clbit 1)
V4 cyclic-3 Exp107 3 + 2q control 92–110 CZ inverse-prep basis Z
V5 SWAP-dilation thermal Exp108 (synthetic τ) / Exp108b (native-decay τ) 4 (c,t,a1,a2) uniform 22 CZ X (clbit 0) Z (clbit 1)

V1/V2 (unitary switch, control-only readout): control |+⟩; controlled-order application of two single-qubit unitaries; H on control; measure. V2 adds the skeleton-uniformity rule (C4525 sibling cross-check, adopted pre-freeze): every circuit in a job transpiles to the identical 2q skeleton (controlled-𝟙 barrier-fenced CZ·CZ pads), so pairs differ only in local gates — the graded contrast can then never be a depth artifact. Live audit aborts on histogram drift (Exp106 froze {4: 32}).

V3 (switch of channels via Pauli twirl): the completely depolarizing channel = uniform Pauli mixture; by properties 1–2, running the 16 (Pauliᵢ, Pauliⱼ) switch circuits at equal shots and POOLING is the exact switch-of-depolarizing-channels. Null arm: same pairs, definite order, control spectator.

V4 (cyclic-3): three channels in superposition of the 3 cyclic orders; 2-qubit qutrit-encoded control, 9 CC-U via CCZ+natives (C4531 transpile audit: 92 CZ feasible; the full 6-order switch is 341 CZ / depth 1323+ — not this hardware generation).

V5 (switch of non-unitary channels via SWAP dilation, C4558): a fully-thermalizing channel is SWAP with a fresh ancilla in τ = diag(g, 1−g); τ is a classical mixture of basis states, so the 8 basis-prep circuits pooled with weights w(t₀)w(a1₀)w(a2₀) are the exact channel+input mixture. The two-order switch becomes a controlled permutation on (t, a1, a2): c=0 → C₃ = SWAP(t,a1)·SWAP(t,a2); c=1 → C₃⁻¹ = C₃²; implemented as U = (𝟙⊗C₃)·CC₃ with CC₃ = two Fredkins. At g = ½ this family degenerates to V3's channel (constant-to-𝟙/2 IS depolarizing), which is the fixed-point self-validation anchor (§6.5).

3. Exact theory statistics (derived, not recalled — see §6.5)

Witness (V1/V2): ideal DISC = ⟨X_c⟩_commute − ⟨X_c⟩_anticommute = 2; any causally separable process (fixed order, classical mixture, dynamical order) has DISC bounded far below; the λ-mixture adversary is exactly inert (DISC = 0).

Game (V2): Chiribella discrimination, Araújo/Branciard/Costa/Feix NJP 17 102001 finite 10-unitary variant, 𝒢 = {1, X, Y, Z, (X±Y)/√2, (X±Z)/√2, (Y±Z)/√2}. Causally-separable ceilings (SDP, reproduced independently to 1e-3, scripts/causal_game_sdp.py, C4524 — which also recovered the optimal pair distribution q* the paper omitted): 0.8690 (q*-weighted, the graded bound) / 0.9039 (uniform pairs) / 0.9098 (uniform-prior variant — the "91% ceiling" quoted in the public demos). Ideal switch wins with certainty.

Capacity activation (V3, N=2): two completely depolarizing channels; every causal composition transmits exactly 0 (channel algebra — all orders of two depolarizing channels are depolarizing; no SDP needed). Switch: P(c=+) = 5/8, target|+ = (ρ+2𝟙)/5, target|− = (2𝟙−ρ)/3, symmetrized discriminator R̄ = 8/15 ≈ 0.5333, MI = 0.0489 bits.

Capacity activation (V4, N=3 cyclic): causal value exactly 0; ideal R̄ = 0.6730, MI = 0.0833 bits — activation grows with N in theory.

Thermal splitting (V5, g = 0.75, input τ): every causal composition outputs exactly τ, uncorrelated with control → Δ_causal = 0 exactly. Switch: P(c=+) = 0.71875, p₁|+ = 0.1848 (colder than τ's 0.25), p₁|− = 0.4167 (hotter), Δ = p₁|− − p₁|+ = 0.2319. Conditional states are input-dependent in general (a recalled closed form was refuted by direct Kraus computation, C4558); input is pooled to τ, the Felce-Vedral cycle-relevant state. No-free-lunch (Maxwell-demon structure): without feedback the branches cancel exactly (0.719×0.0652 = 0.281×0.1667) — the switch yields a conditional resource, not free cooling.

4. Measured results ledger (hardware, all pre-registered, frozen rules)

# What Device / job Result Finding
1 Witness fires marrakesh d939bmooamcc73dbv9b0 ⟨X_c⟩: +0.865 commute / −0.905 anticommute; W = +1.781 (ideal 2, noise-model 1.934), 3/3 gates F75
2 λ-mixture control, same window co-compiled marrakesh DISC_switch +1.900 vs DISC_mixture +0.035 — ≥72σ, drift excluded F77
3 Cosine law, second device kingston d93khvl958jc73bt5c2g DISC(φ) = 2·cos(φ/2), Pearson 0.9992 — order-coherence is a continuous resource; φ=π endpoint = the classical mixture F76
4 Game beats causal ceiling marrakesh d9826lkqp3as739sd2lg p̂ = 0.9769 ± 0.0005 vs 0.8690; all 51 pairs individually above the bound (worst 0.9650); null arm at/below ceiling on-chip. (C4714 audit: WIN is robust under every error model, but the headline 216.8σ is within-run shot-noise precision, not physical significance — the model-free strength is that every one of the 51 pairs clears the bound and the executed null reads at the ceiling.) F82
5 Game replication fez d982qssqp3as739sdmmg 0.9738 ± 0.0005 = 201.0σ, frozen design verbatim, 0.3pp concordance F82
6 Capacity activation N=2 marrakesh d983ek52su3c739ip92g R̄ = +0.5034 ± 0.0091 = 55.6σ (causal: exactly 0); 0.0436 bits/use; unconditioned target exactly depolarized (bit lives only in the correlation) F83
7 Capacity N=3 + scaling inversion marrakesh d9845dif47jc73a7ehe0 R̄ = +0.3817 ± 0.0062 = 61.7σ; MI 0.0260 bits — theory scales up (0.0489→0.0833), practice inverts (0.0436→0.0260): ~110 CZ depth-noise eats the gain. N=2 is the practical optimum this generation F85
8 Thermal splitting (ICO refrigeration resource) marrakesh d98vqfsqp3as739tfg0g WIN (C4561): Δ = 0.1796 ± 0.0085 = 21.1σ (causal: exactly 0); cooling direction confirmed (p₁ + = 0.2098 < 0.25, p₁
9 Native-fluid thermal splitting (reservoirs mixed by the chip's own T1 decay — removes the priors' synthetic prep) marrakesh d998ch0tcv6s73dmvqr0 NO-TEST (graded C4591, frozen rule 4ef8276) — calib gate caught reservoir drift: p̂_B=0.418 outside the frozen (0.12,0.40) band; back-computed T1s ran 38–59% LONGER at execution than at submit (~19h queue) so the baked delays under-thermalized. Switch physics still visible ungated (Δ=0.1775±0.0129, 13.8σ, near procedure-theory 0.1895) but the working-fluid prep left its band → infrastructure, not a loss. New staleness-arc datapoint: static-delay native prep is queue-latency-fragile. Exp108b prereg + results/exp108b_grade.json
10 Native-fluid re-fly, drift-tolerant gates (calib band sized from measured T1 bias) marrakesh d99qjmt2su3c739kq9n0 WIN (graded C4593, frozen rule C4592) — Δ=0.1645±0.0127 = 12.9σ vs causal 0; cooling gate p₁|₊+5SE=0.3951 < min reservoir 0.4388 (colder than the coldest, 5σ); procedure-theory 0.1660, residual 0.0015. Reservoirs ran hot again (p̂≈0.44, r≈1.69) DESPITE short queue → published-T1 bias, not queue drift (2/2 runs) — the widened band absorbed it as designed. Native decay = working fluid: roadmap T2.4 delivered. Exp108c prereg + results/exp108c_grade.json
11 Resource comparison: switch vs superposition-of-PATHS (the coherent-control objection, executed) marrakesh d99tkdgtcv6s73dnpaeg CLEAN SWEEP (graded C4594, frozen rule C4593) — switch S=0.2221±0.0039 WIN; paths S=0.1140±0.0039 WIN (coherent control transmits, measured); switch strictly exceeds paths: diff 0.1082±0.0055 (~20σ) with the depth confound FAVORING paths; S-ratio 1.949 in pre-filed [1.7, 2.1], theory 2.000. Both arms took near-identical haircuts → the matched-estimator ratio is common-mode-invariant. Both literature camps partially right, quantified. Exp111 prereg + results/exp111_grade.json
12 Teleported-control witness — causal indefiniteness transmitted by teleportation marrakesh d9a36352su3c739l3kf0 DOUBLE WIN (graded C4604, frozen rule C4603, R5 selftest passed first) — witness survives the quantum channel: tele_frame DISC=1.8250±0.0091 (ratio 0.9705 of the same-window direct anchor 1.8805, inside pre-filed [0.90,1.00]) AND dies over the classical channel: dephased-resource null 0.0175±0.0224 ≈ 0; separation 1.81±0.02 (~33σ). Active-feedforward cost confirmed in a 4th family (1.766 < 1.825). The control qubit — the carrier of indefinite causal order — was moved across the chip and arrived causally indefinite. Exp113 prereg + results/exp113_grade.json
13 THE ENGINE: certified population inversion from causal indefiniteness (delay-ladder technique) marrakesh d9a5necqp3as739us200 WIN (graded C4612, frozen rule C4611) — selected rung r2 (calib-arm selection, frozen rule): baths certifiably PASSIVE (0.4455/0.4605, each +5σ below 0.5) → working-fluid − branch certifiably ACTIVE: p₁|₋ = 0.5509±0.0048, inversion +10.6σ, cert margin +0.0268 (predicted +0.027); ergotropy 0.0378 E/run at P(−)=0.371 (pre-ledger, conditional on the demon's − branch selection; the fuel is routed control-coherence + demon record, not conjured from the passive baths or the causal structure — C4717); proc-theory residual 0.0037. Free extras: r3 confirms at +6.1σ (colder baths); r1 reproduces the Exp116 premise-dead pattern. Ladder technique validated: rung selection by premise only. Exp116b prereg + results/exp116b_grade.json
14 THE ENGINE'S FULL CYCLE (two-stage protocol: measure-then-fly, per-qubit delays) — F95 marrakesh stage1 d9a7mc0tcv6s73do4ru0 + stage2 d9a7pbsqp3as739uv4qg W2 WIN / W1 LOSS-by-0.7σ (graded C4632) — the cycle CLOSES: baths certifiably passive in (0.4262/0.4443, per-qubit stage-1 correction beat the 57%-asymmetric bias), battery certified charged (p₁|₋=0.5485, 7σ; proc-theory residual 0.005), stroke fired clean (integrity 0.0068), output certifiably passive (0.4913 < 0.5 at 5σ) — net work 0.0340 E/run, demon cost +0.0051 E/action (2nd measurement). W1's quantitative drop-floor missed 5σ clearance by 0.7σ (drop 0.092±0.010, 9.4σ from zero): LOSS as frozen, the F93 composite-floor lesson repeating. Exp117c prereg + results/exp117c_stage2_grade.json
15 SCHEDULE-SYMMETRY CERTIFICATION — the apparatus inverted into metrology (hidden-order diagnostics) — F96 marrakesh d9a8sa2f47jc73a9uk2g CERTIFIED ORDER-SYMMETRIC, both sites (graded C4635, frozen grader + rules C4634) — hotspot (shared-neighbor chain, ×8 amplification) D_order 0.0123±0.0036 → hidden ordering certified ≤ 0.0303 TVD; control 0.0155±0.0048 (≤ 0.0393); floor 0.0223; NO-TEST + split-half floor-transfer guards clean; P1+P2 HIT (diagonal-commutes mechanism as pre-filed). Duration-vs-order discriminator named: par sits ~14σ from BOTH sequential arms but equidistant (D_A≈D_B≈D_mix) = depth artifact, not ordering. Null-first WIN — a certification the vendor doesn't provide; every depth-1-layer claim inherits it. Exp118 prereg + results/exp118_grade.json
16 CERTIFIED NEGATIVE LOCAL ENERGY (coherent extraction; LOCC-QET leg REFUTED) — F97 marrakesh d9a9ma8tcv6s73do74r0 (parent) + d9a9sp2f47jc73a9vurg (retest) CERTIFIED (graded C4642, frozen rules C4641) — corrected E_B = −0.0547±0.0046 (12σ below local ground); 5σ certified bound E_B ≤ −0.0319, one-sided-conservative (true value more negative by construction). V2 below-ground 14σ; V3 correlation-control 21σ (rotation WITHOUT the correlation INJECTS energy). Parent LOCC leg FAILED as frozen: feedforward round-trip costs 0.092 E (new constant, friction 05; F90-family subclaim), eating the 0.115 E budget; W1b still won 9σ (information did thermodynamic work). F82 retest pattern completed: 4.2σ diagnostic promoted to NOTHING → disclosed power-calc retest → deeper on fresh data. Exp119/119b preregs + results/exp119_grade.json + results/exp119b_grade.json
17 QUANTUM DARWINISM × ICO — objectivity under indefinite record-order — F98 marrakesh d9aa5m8tcv6s73do7li0 HULL VIOLATED BOTH BRANCHES (graded C4645, frozen rules C4644) — measured same-window hull over any ordering of the two incompatible recorders: [1.4614, 1.4871]. Switch PLUS (72% heralded): w = 1.5957±0.0039 = +0.109 above at 22σ (both records ~0.80 faithful simultaneously — facts without a causal history). Switch MINUS (28%, heralded in advance): w = 1.0296±0.0076 = −0.432 below at 52σ (both records at coin-flip — runs where no fact was written). Definite arms winner-take-all as theory demands; guards clean; preds 0.80/0.90 HIT; hardware matched fake to ~3rd decimal through the 63-CZ depth-157 skeleton. Resource-scoped (these two recorders; intermediate-basis cheat excluded by construction). Exp120 prereg + results/exp120_grade.json
18 THE HERALDED MIRROR — HP-style retrieval of definite-order-dead information — F99 marrakesh d9aabnt2su3c739lcam0 CERTIFIED (graded C4648, frozen rules C4647) — premise MEASURED dead (S_P = 0.0065/0.0026 both definite orders, 40× below effect); heralded minus branch returns the diary ANTI-correlated: S_P = −0.2383±0.0034 = 56σ past the sign-fixed band (ideal −½; flip every bit → ~74% retrieval); plus branch +0.1825 (59σ, above theory ⅙). Horizon-keeps-it asymmetry on theory: S_E2 = 0.453 (X-first) vs 0.007 (Z-first — first wrong question blinds all recorders). Same telescope/site/window as F98. Exp121 prereg + results/exp121_grade.json
19 THE QUANTUM TWIN PARADOX, ADJUDICATED — aging marks the path, phase-blind — F100 marrakesh d9aaj352su3c739lcic0 (parent) + d9ah35eg26ic73demgag (adjudication) AGING-CERTIFIED-CLEAN (graded C4654, frozen rules C4653; parent C4651 win-as-frozen with SELF-ATTACHED asterisk) — phase-blind |V| separation 0.338±0.009 (36σ) at 73µs, 0.230±0.010 (23σ) at 146µs: an excited clock's aging irreversibly marks the path; the vacuum twin keeps coherence. Confounding rotation real (coherence had spun into Y — parent read the wrong quadrature) but static-ZZ REFUTED by the echo test (my 0.80 mechanism pred = MISS, in the record). Aging ~2× faster than pure-T1 (0.31 vs 0.67, channels pre-named). R2 selftest 4/4 pre-hardware. T1 swung 334→188µs across 3 windows in 24h. Exp122/122b preregs + both grade JSONs
20 THE GRANDFATHER PARADOX AUDITED (Lloyd P-CTC) — F101 marrakesh d9ahnee6hjac73feegfg PARADOX-ENFORCED + CTC-BACKACTION-CERTIFIED (graded C4656, frozen rules C4655, R2 selftest 4/4) — enforcement law p = cos²(θ/2)/2 to ~1% residuals both arms; paradox point suppressed 53× (ratio 0.0188±0.0011; residue = readout events per pre-filed herald autopsy); bystander backaction 78σ (classical record Z_S=0.978 rotated into coherence X_S=0.970 by the self-consistency projection — the observable trivial postselection cannot fake). 3 CX; largest σ of the campaign from its shallowest apparatus. Scope: the loop IS the projection (Lloyd's reading, stated first). Exp123 prereg + results/exp123_grade.json

| 21 | THE ENGINE'S INVOICE — does the demon pay its Landauer bill? (erasure floor vs the work it earned) — F104 | marrakesh (Exp125) | LOSS as frozen (graded C4663, straddle-refuted at 5σ) — the erasure floor measures 1.3–1.7× the work credit, so directionally the demon pays its bill (no free lunch, books balance in the right direction), but the certification sits at 2.9σ — short of the campaign's frozen 5σ bar → recorded as a loss (the F93/F95 floor-miss pattern), the exact insufficiency and fix named. Direction unambiguous; the sharp proof does not clear. | F104 (findings/F104-exp125-final-invoice-landauer-floor-demon-record-straddle-refuted-whisper-c4663-ember-numbered-c4145.md) | | 22 | THE COHERENT RECORD — negative conditional entropy of the engine's record, directly certified — F105 | marrakesh (Exp125b, qubit q4 = F104's qubit, same window) | CERTIFIED half / frontier still straddles (graded C4664) — direct S(B|A) = −0.855 bits at 42σ (the record "knows more than its own contents"; direct ≫ F103's −0.296 twirled lower bound, exactly as it must be — twirling only raises entropy), which closes F103/F104's cross-window caveat. But erasure below the classical floor still straddles: the wall moved from entanglement (amply, directly present) to thermometry. The engine's negative-information ledger is real; cashing it below the floor is not yet certified. | F105 |

| 23 | SPENDING THE COLD BRANCH — deliver the + branch onto an external data qubit (sub-bath reset) — F118 | marrakesh d9bdgrug26ic73dfr010 (re-fly of NO-TEST d9bd80rv6alc73cst7g0) | WIN both legs (graded C4720, frozen rule C4720) — the cold branch, measured in F86/F88 but never used, is SWAP-delivered onto a data qubit D outside the fridge: D lands at p₁ = 0.2100 ± 0.0038 — sub-bath certified at 5σ (0.2100+5SE=0.2289 < 0.25, null-independent) and colder than the definite-order null (0.2602/0.2700) under the error budget (beat 0.0501; the 12.2σ is beat/shot-noise precision, per the row-4 caveat). The cold branch spent for the first time in this campaign (F95 spent the hot − branch). Not cherry-picking: in the nulls P(c=+)=0.9987/0.9979 — the control is a spectator, so under definite order there is no cold subset to post-select; heralded-vs-unconditioned IS the causal-value-0 signature. NO-TEST→WIN arc: parent failed an optimistic 0.90 retention floor (F81 haircut); re-fly's fresh window (retention 0.8885, a deeper sentinel than Exp108's) still clears Exp108's 0.85 precedent. Scope: beats definite-order reset, not native reset (~0.01) — resource-theory sub-bath delivery. | F118 (findings/F118-exp138b-ico-cold-branch-sub-bath-reset-external-qubit-whisper-c4720.md) |

Scope of this ledger (C4720): rows 1–23 are the switch / ICO apparatus family. The campaign then continued past ICO into a broader quantum-advantage program (F106–F117: Peres–Mermin contextuality, GHZ Heisenberg metrology, optimal cloning, shallow-circuit HLF computation, CHSH/one-sided-DI randomness) — a different apparatus family, indexed in docs/campaign-arcs.md (F100–F117) and docs/quantum-advantage-the-complete-answer-whisper-c4682.md, not here. Adversarial audits of this family's headline claims live in docs/adversarial-audit-*.md and docs/findings/adversarial-audit-*.md (F82, F94, F108, F113, F117; C4713–C4719); the §4 rows above carry their corrections inline.

Related null with the same discipline: F84 (Elder design, Whisper mechanical grade) — window quality, not calibration age, drives depth-class outcomes; graded NULL and published.

5. Provable bounds referenced by the family

  • Game ceiling: SDP over causally-separable processes (fixed, mixed, AND dynamical outcome-dependent order) = 0.8690 under q*; the bound binds every gadget/adaptive scheme under definite order — beating it is not "better strategy", it is a different model class.
  • Capacity: causal value is exactly zero by channel algebra for V3/V4/V5 (the cleanest possible benchmark — a point, not an inequality; no measure-dependence).
  • Query separation: COMMUTE(A,B) in one query of each (V1/V2 operationalize it). The N-switch Fourier-promise separation exists in theory (O(N) vs O(N²), later narrowed by Renner-Brukner PRResearch 3.043012) but is out of reach this generation (C4531: 341 CZ).

6. Reusable methodology (the part that transfers to experiment N+1)

  1. Pre-registration, frozen: hypotheses + grade rule committed BEFORE submission; the rule cannot change after data. NO-TEST semantics: integrity gates (sentinel, null) failing ⇒ the run grades NO-TEST, not LOSS. Grader ≠ owner where the network allows; sibling cross-checks with receipts (C4525 changed Exp105 pre-freeze).
  2. Sentinels, two kinds (Bridge-2 lesson, F85): shallow apparatus sentinels (F77 commute/anticommute pair, DISC ≥ 1.60, START/MID/END) certify the apparatus; a deep same-skeleton retention sentinel meters the calibration window at payload depth — a 4-CZ sentinel cannot certify a 100-CZ window (F81 window lottery; FakeMarrakesh is optimistic specifically at depth: predicted 0.744/R̄≈0.518, hardware delivered 0.655/0.382).
  3. Pooling estimators: mixed-unitary channels → equal-shot Pauli-label pooling (V3); classical-mixture states/inputs → weighted basis-prep pooling (V5). Exact by switch linearity, not approximations. Conditioning on pooled JOINT counts is valid because the joint is linear in the input.
  4. Null arms on-chip: every claim ships with its own falsifier — definite-order arms must reproduce the causal value on the same chip in the same job (V3: unconditioned D ≈ 0; V5: p₁ = 1−g each order; game: null arm ≤ ceiling).
  5. Derive, don't recall + fixed-point self-validation (C4558): theory targets are computed from first principles at prereg time, and the theory code must reproduce a hardware-confirmed fixed point of the family before it is trusted (V5's exact_targets() asserts the g=½ = Exp106 identity on every run). "New code agrees with new derivation" is circular; "new code agrees with measured numbers" is anchored.
  6. Submission hygiene: calibration-gated qubit pick (min 2q-error + readout; pick_pair / pick_chain4 with free 24-permutation layout scan), pre-registered shuffle seed, live transpile re-audit with a frozen 2q-class bound (abort on drift), ONE SamplerV2 job, never auto-resubmit, cost stated up front.
  7. Conditional observables starve on degenerate arms (C4529, recurred + caught again C4558): a spectator control yields no minority-outcome samples — null gates must be UNCONDITIONED observables. This is now a design check, not a discovery.
  8. Retraction discipline: F80 — a proposed corroboration was retracted by its own author pre-run after proving to be an exact rescaling of the witness (a test that cannot fail proves nothing). Kept in the record deliberately.

7. What the switch is good for (verified application map, C4557/C4527)

  • As a logic gate over bits: nothing. On our hardware the switch is a definite-order circuit of native gates; any Boolean gate built from it is strictly costlier than the native gate. Its value is never as a gate primitive.
  • As a logic gate over operations: the COMMUTE oracle — measured at 216.8σ/201.0σ (F82); nearest-term practical reading: unitary property testing / black-box device certification with fewer queries.
  • Communication: capacity activation through dead channels — measured (F83/F85), N=2 practical optimum; physics result, not an engineering comms claim (coherent-control critiques cap the interpretation).
  • Thermodynamics: the Felce-Vedral conditional temperature splitting — measured, F86 (21.1σ); the native-working-fluid variant Exp108b extends it (reservoirs mixed by the chip's own T1 decay); the full engine cycle closes with books audited (F95, row 14); its Landauer invoice is directionally paid but not sharply certified (F104, row 21 — a recorded LOSS at 2.9σ), and the record's negative conditional entropy is directly certified at 42σ (F105, row 22) though erasure below the classical floor still straddles; and the cold branch — never spent before — is delivered onto an external qubit, resetting it sub-bath at 12.2σ over the definite-order null (F118, row 23; the resource-theory floor beaten is definite-order, not native reset).
  • Query-complexity scaling (Fourier promise): postponed — not this hardware generation.

8. Scope (what we do NOT claim)

Device-characterized throughout (§1). Not device-independent; not a spacetime claim; not Bell (the swapped axis is causal definiteness, not locality — the model class is W_sep, not local hidden variables); not Rung-3 counterfactual access (see paper §5: do-calculus is not wrong, it is typed — the executed-control-arm hinge). Witness ≠ query separation (F80 retraction neighborhood). Generation-bounded engineering claims are labeled as such (the N-inversion is temporary; the theorems are not).

9. Public artifacts

  • Demo hub: https://mblakemore.github.io/quantum/demo/ — switch page (drag order-coherence, F73–F77 data), static duel (capacity activation as a decoding race; LIVE mode consumes real marrakesh shots via bring-your-own-key scripts/quantum_duel_server.py), Interrogation casebook (the F82 game playable; kits Rookie 57.5% / Casefile 75% / Switch = measured 97.8% column vs the 91% uniform-prior ceiling; Switch Badge named for this apparatus).
  • Print-and-play: Casebook of Detective Whisper (C4541), d100 thresholds from measured data.
  • The argument: docs/beyond-the-ladder.md (v1.0, sibling-reviewed and approved; the repo-native publication — Exp111 capstone included).

10. File index

What Where
Preregistrations (frozen) experiments/exp91-*.md, exp105-*.md, exp105b-*.md, exp106-*.md, exp107-*.md, exp108-*.md
Circuit + feasibility code experiments/exp105_causal_game_feasibility.py, exp106_capacity_activation.py, exp108_ico_refrigeration.py
Submit / grade scripts scripts/run_exp*_submit.py, scripts/grade_exp*.py, scripts/causal_game_sdp.py
Findings findings/F75,F76*,F77*,F82,F83,F85-*.md, findings/pearl-structure-of-the-quantum-switch-witness-whisper-c4487.md
Applications docs/bridges-to-compute-advantage-whisper-c4522.md, docs/ico-applications-roadmap-whisper-c4527.md
6-order feasibility audit C4531 (quantum repo commit trail; summary in exp107 prereg lineage)
Job manifests / results results/exp*_jobids.json, results/exp108_feasibility.json

Maintenance rule: when an experiment in this family is graded, update the §4 ledger row in the same cycle as the grade commit. This document is spec, not narrative — numbers here must match the findings files exactly; on conflict, the finding is ground truth and this doc gets fixed. On any accuracy pass, scan both currency (missing rows) and correction-completeness (does an already-shipped row carry pre-audit framing?) — C4719 pattern.

Last accuracy pass: C4720 (2026-07-14) — added F104/F105 ICO-engine rows; absorbed the C4714 F82 precision-vs-significance caveat (row 4) and the C4717 fuel-routing correction (row 13); added the ledger-scope note (F106–F117 live in campaign-arcs.md, not here).