An empirical campaign on IBM Heron-generation hardware — ibm_marrakesh, ibm_kingston, ibm_fez — and, as of the cross-platform arc, a non-IBM chip (Rigetti Cepheus-1-108Q, via Amazon Braket) — run end-to-end by an autonomous multi-agent network from May 2026 (ongoing). Every experiment is pre-registered before it flies; every number traces to an IBM Quantum job ID. The F-series of findings runs to F121; the exp-numbered findings continue through the July "Star Trek" arcs (logical qubits, a composable network, exotic phases, the physics of time), a fault-tolerance arc (Exp236–246) — active error correction, the live QEC loop, and a universal logical gate set closed behind the shield — continue through the decoder-race arc (F120–F121): a six-flight, cryptographically-sealed race series whose court-certified 476× runtime win was then retired by the campaign's own red-team, pre-submission (C4996: the planted problem's algebra is classically poly-time — the F120 shot-axis instrument stands; no runtime advantage is claimed) — across 100+ pre-registration documents — and culminate (so far) in door (b), August 10: the campaign's first advantage-class WIN under its full standard — a sealed, blind, court-graded two-copy learning result at 104σ, beating a proven single-copy floor by 9.3× as-executed at the amplitude the hardware actually delivered — and, on August 11, the Causal Compass: cause-vs-common-cause called from observation alone on sealed data, 75 of 75, 8.66σ, verified by three non-author seats while its author, who can deblind by content-matching, is barred from certifying it.
Most of what you can buy from a quantum computer today is characterization: how good are the qubits? This campaign asks a sharper question — what can this hardware do that a classical, causal, or definite-order process provably cannot? — and answers it one scoreboard at a time, with the losses kept next to the wins.
The centerpiece is a quantum switch: a circuit where the order of two operations is itself in superposition. The switch was theorists' idea; photonics labs demonstrated it first, and early versions ran on chips like ours. What this campaign built was the scoreboard — pre-registered games and channels whose limits are provable theorems for any definite-order process — and then beat those limits on silicon. A discrimination game was won at 0.9769 against its theorem ceiling 0.869 and replicated the next day at 0.9738 on a chip the design had never touched — the two-chip agreement, not the 216.8σ of within-run precision, is what carries the claim; information crossed channels of exactly zero capacity; a thermodynamic splitting forbidden to every ordered process was measured at 21.1σ. And the centerpiece has now left IBM entirely: the switch's causal witness certifies on Rigetti's superconducting silicon — a different vendor, fab, and native gate set, reached via Amazon Braket — against the identical frozen bounds, so indefinite causal order is not an IBM artifact. Around that centerpiece sits a systematic map of what this hardware generation can and cannot do.
🔀 Play with it → mblakemore.github.io/quantum — The Quantum Museum: 56 interactive exhibits across 8 wings, one per result — from the causal switch to the self-healing logical qubit to the new Temporal Investigations pieces (a measured negative eigenvalue; Hardy's impossible event). Drag the switch's order-coherence and watch the measured cosine law appear; play the contextuality game against a classical bot that provably cannot win; walk the certified-randomness trust ladder; print & play the tabletop casebook. Every number in every exhibit is measured hardware data, and 23 of the 56 carry a full specification sheet tracing the result to its job ID.
Every result below is pre-registered, and its success criterion is exceeding what a classical / causal / definite process provably permits (or certifying a first-of-kind) — not characterizing hardware. Grouped by theme, most significant first; the eight operational-characterization arcs follow in What Else the Campaign Established.
🏆 TOP BILLING — the first advantage-class WIN under the full standard: door (b), the two-copy learning separation (August 10). The task: a sealed hard-family state hides one Pauli among 4^16 ≈ 4.3 billion; a theorem (not a conjecture — the thing F119 never had) says any single-copy strategy needs ~1.9 million copies at the delivered amplitude, while two-copy Bell sampling reads it almost directly. The arc is the campaign's method in miniature: the first flight FAILED-AS-FROZEN on a one-line prep bug (seven qubits un-randomized — blind weight-1 tripwires localized the defect to the exact qubits from sealed data); the fix got a can-fire gate (F-MIX), and the fix was then validated from the blind side — the decode seat, unable to see the answer, reported the failure's signature absent on identity positions it could not know, and the unseal proved it right. The re-fly flew under a fresh seal bound to a single-use go, sized itself mid-air to the epoch it actually occupied (weather-gated
T(ε), a frozen function meeting live hardware), and landed: the sealed weight-12 Pauli recovered at 104σ — 7.6× the largest artifact in 112 blind probes — in 109 seconds, under budget. The claim-card number is 9.3× as-executed (proven floor at delivered ε over copies actually spent), self-capped by the register seat at the moment of the win; the full epoch chain (nominal 57× → pilot 13.9× → sizing 14.1× → delivered 9.3×) travels with it, and the union-bound budget defense is on the record. Prereg & seal chain · flight manifest & decisions · [the arc, day by day, on the bus record] — every step (register → seal → go-bound-to-digest → blind decode → hash → unseal → grade) checkable against timestamps rather than anyone's word.
⊘ FORMER TOP BILLING — WON, THEN SUPERSEDED, by our own red-team, pre-submission (F120–F121 → C4996, July 23). From the campaign's first weeks the answer to "is it faster?" was no — F54's measured depth wall stood, and by mid-July the campaign's own hidden-shift race verdict said window-closed, with the number a future machine would need. Then a re-read of that verdict's own discarded calibration data found the number was measured on the wrong observable: each shot at depth is the planted answer plus sparse typos, so N shots are N noisy transmissions of one codeword — the shot axis is a code (F120: per-bit information survives the depth wall ~30× better than the modal answer; blind exact recovery demonstrated to d2q=310 — the organic-arc ladder's lad_d_hi register, sealed-commitment CLEAN grade [0,0,0], campaign-deepest exact t0 consensus (
results/exp_organic_rhot_law_reveal_rung0_ember.json; supersedes the race-4 record of 217)). Six pre-registered, cryptographically-sealed races later — each fold booked and each fence forged from the previous miss (observable → placement/endianness → granularity → qubit tilts → register quality → die topology) — race-6 onibm_kingstonrecovered a sealed 80-T-gate hidden-shift string EXACTLY, blind, in 3.82 s of QPU against a frozen classical band: 476× at the harshest edge, WIN at every edge, graded 3-of-3 (F121). Fences printed on the result: one instance family, one die, best-known-solver engineering race (NOT a complexity theorem — F119 holds that seat in sample-complexity currency; F54's brute-force-simulation wall is untouched), supersedable-by-design — a faster classical solver retires the number, and that mechanism is the point. The follow-on instrument flights decomposed the "magic tax" itself (T-localized flat stochastic core × a depth-growing coherent few-bit drift that resists randomized compiling) and priced the whole pipeline into the attenuation map v1.2. The WIN verdict · arc section · program closure (H8 P9) · the map v1.1→v1.2. Hours later the printed supersession clause fired — by our own hand, before any submission: asked whether F121 was IBM-tracker-ready, the network red-teamed its own instance and the Maiorana–McFarland linear-structure attack (fixed-x linearity in y leaks the shift) read the sealed answer in 41 oracle queries / ~0.25 ms — ~7×10⁶× under the 1,818 s floor, independently confirmed by all three court seats with two disjoint implementations. The 476× priced simulation of the circuit; the correct floor for an advantage claim is the best classical method for the problem, and that is poly-time. F121 as a runtime advantage: RETIRED. F120 (the shot-axis code and blind exact decode): stands, as an instrument result — not an advantage. F119: audited against this same axis — SUPERSEDED as-executed / QUALIFIED in principle (Ember C4215: seal + honest-oracle PASS, the problem held; the executed baseline and floor status fell; remedy re-fly pending). The red-team finding · Elder's independent co-verify.
🌐 Cross-platform — the causal-order court leaves IBM, then leaves superconductors. The strongest answer this campaign has to the sharpest skeptic ("is indefinite causal order just an artifact of your one chip?"): the switch-bench causal axis, graded against the same frozen theory bounds as every Heron flight, certifies
PASS-CAUSALonRigetti Cepheus-1-108Q— a different vendor, fab, and native gate set (Rx/Rz/CZ), reached through Amazon Braket. Witness W = 1.2165 ± 0.0224 (54.4σ over the causal-mixture bound 0), capacity 0.2873 (20.1σ over its floor), null clean (corrected decode after the program-set reversal bug; certified by known-input calibration, C4946). Ranking: Rigetti is a markedly weaker causal chip (W 1.22 vs Heron's ~1.90 — it certifies, at ~64% of Heron's witness value). And the cross-modality decider has since flown: on IonQ Forte-1 trapped ions — where the CZ Z-biased dephasing our superconducting error story rests on is absent — the witness fires at W = 1.910 ± 0.141 (13.5σ) with a structurally-matched, validated definite-order null (W_matched = −0.09); a narrower card than Rigetti's (capacity/null-integrity arms not flown, cost-gated). Indefinite causal order is substrate-general: three vendors, two physical modalities, one frozen instrument. Exp210 — full results & the pre-filed prediction miss kept in the record · the multi-substrate white paper.
Has this hardware shown a quantum advantage? As of August 10: yes, in sample-complexity currency against a PROVEN floor — door (b), above: 9.3× fewer copies than any single-copy strategy can achieve by theorem, blind, sealed, court-graded, at 104σ. And scoreboard-by-scoreboard before that: yes on causal-order games (216.8σ within-run, replicated two chips), nonlocal/contextuality games (196σ), communication capacity (superdense 341σ, QRAC 110σ), thermodynamic resources (population inversion, negative energy), and metrology at the Heisenberg limit (168σ, persisting to N=5) — each a provable bound beaten; and, as of the July learning arc, yes on a computational scoreboard in the sample-complexity currency: Exp142 (F119) identified a sealed hidden n-qubit Pauli via two-copy Bell sampling — the problem is genuinely hard single-copy (sealed hiding-commitment, k-local marginals maximally mixed), the structural inverse of F121's publicly-attackable instance. But the advantage claim was superseded as-executed by our own red-team (Ember C4215, pre-submission): the conventional arm flew a fixed measurement basis per row, a delivery artifact that lets a 36-copy determinism decoder beat two-copy's 68 copies as-flown → zero advantage as-executed; and the (3/2)ⁿ single-copy floor is open, not proven (Elder C6490 appendix — SUPPORTING-only, its own "OPEN LEMMA"), so the graded 4.9×/31.5×/266.6×/2417.5× ratios (a naive product-basis baseline + the artifact + a 2× copies-vs-Bell-measurement units inflation) are retired. Residual: 10×–331× in copies vs best-known single-copy — conditional and supersedable, not unconditional — sealed-commitment blind protocol, frozen grader (booked with fences; its m-term successor Exp144 graded NOT-WIN as frozen — quantum arm perfect at n=4/6, classical-race arm null — and Exp145 recovered Simon's hidden structure exactly, 3/3 rungs, the query-mechanism flight). And the runtime scoreboard opened on July 23 — and was closed the same day by our own red-team: F121's 476× decoder-race win was superseded pre-submission (C4996 — the MM problem's algebra falls to a 41-query classical linear-structure solve, ~0.25 ms; the fence that priced this exact path fired, by our own hand). No runtime advantage is claimed; F120's shot-axis code stands as the instrument result. The sample-complexity claim (F119) is SUPERSEDED as-executed / QUALIFIED in principle (red-team audit): the problem held, but the executed advantage did not and the floor is open — a conditional 10×–331×-in-copies edge over best-known single-copy is the residual, and a remedy re-fly (conventional arm, one fresh basis per copy) is pending before anything is submitted externally. What remains no: a raw brute-force circuit-simulation speedup — F54's measured wall and F85's scaling inversion stand untouched, and no result here claims them. A different column also opened: the constant-depth BGK shallow-circuit depth-separation solver runs on silicon (F113, 90% valid / full solution-coset) — the apparatus of an asymptotic theorem, not a raw speedup and not an on-chip class separation. The full reckoning, wins and non-wins together: the complete answer (supersedes the earlier audit); forward paths: the advantage annex. The July second half (Star Trek arcs, Exp147–197) added two more advantage-flavored columns: a distributed computer (Bernstein–Vazirani split across a cut, 67–141σ) and the campaign's first error-corrected logical qubits — [[4,2,2]] logical operations that beat the bare machine (reversing F06's "QEC doesn't help yet"), including logical CHSH at 29.7σ. The fault-tolerance arc (Exp236–246) then crossed from detection to active correction and closed the universal gate set (Clifford + a fault-tolerantly-injected T) behind the shield — the mechanism of universal quantum computation, error-detected, though not a below-threshold FT fidelity and not a supremacy claim.
Bell nonlocality (F73), indefinite causal order (F82) and contextuality (F106) are quantum theory's three foundational "no classical / local / definite model can reproduce this" results — each beaten on the same hardware, each with an executed null and a bound enumerated in the artifact, not cited. Alongside them, the switch's own provable-bound beats: information pushed through channels of exactly zero capacity.
| Result | Number | Where |
|---|---|---|
| Causal discrimination game beats the causally-separable bound (0.8695, re-solved from Araújo et al. with q* recovered) | p̂ = 0.9769 ± 0.0005 (216.8σ within-run precision); replicated 0.9738 on ibm_fez — the 0.3 pp two-chip concordance (~34σ) is the physical carrier (audit C4714) |
ibm_marrakesh, pre-reg · results |
| Cross-device replication — same frozen design, chip it never touched | p̂ = 0.9738, 201σ (0.3pp concordance) | ibm_fez, addendum |
| Cross-VENDOR certification — the causal axis on a non-IBM chip, same frozen theory bounds, no retuning (indefinite causal order is not a Heron artifact) | PASS-CAUSAL — W 1.2165 ± 0.0224 (54.4σ over 0), capacity 0.2873 (20.1σ over 0.10), null clean (corrected decode, certified C4946); ranks ≪ Heron (W 1.22 vs 1.90) | Rigetti Cepheus-1-108Q (Amazon Braket), Exp210 |
| The magic-square game won — contextuality certified, completing the triptych: the Peres–Mermin "Kobayashi Maru" beaten at 0.96901 vs the classical ceiling 8/9 = 196σ, the ceiling enumerated in-code over all 4,096 strategies; even the worst context clears 8/9 at 37.8σ (min > 8/9 is classically impossible even for mixtures); no-entanglement null 0.657. Game-value advantage, not computational speedup; the bridge to BGKT unconditional shallow-circuit advantage runs on this game | F106 | finding · pre-reg |
| Capacity activation — information through two channels of exactly zero capacity (every causal composition exactly zero) | 0.0436 bits/use, 55.6σ over zero; null arm 0.0001 bits; the bit lives ONLY in the control–target correlation (D≈0 confirmed) | ibm_marrakesh, pre-reg |
| N=3 cyclic switch — capacity activation won at 61.7σ (0.0260 bits through three zero-capacity channels) and exposed the NISQ scaling inversion: theory scales with N, practice inverts (110-CZ depth cost) — N=2 is the practical optimum this hardware generation | F85 | finding · pre-reg |
Provable-bound beats across three communication capabilities and quantum-enhanced sensing — each measured against a classical reference executed on the same qubits, not cited.
| Result | Number | Where |
|---|---|---|
| Superdense coding — pre-shared entanglement DOUBLES a qubit's classical capacity: p_success = 0.9688 = 341σ above the exactly-0.5 unassisted-single-qubit ceiling (computed by the executed no-entanglement null, which sat dead on it); MI 1.77 bits/qubit vs the null's 0.93. Tutorial-class priors credited — the contribution is the frozen bound-referenced grading and the executed null. First rung of the comms ladder | F87 | finding · pre-reg |
| The pocket dictionary — 2→1 quantum random access code: two bits stored in one qubit, either retrievable on demand, at 0.84893 = 110.5σ above the classical one-bit ceiling 0.75 (enumerated over 256 strategies) and 5.2σ below the quantum optimum cos²(π/8) — certified inside the two-sided band (exceeding the quantum law would be a NO-TEST, not a win). The executed optimal-classical arm sat at its own 0.75 law; zero two-qubit gates, the cheapest advantage flight possible. Third rung of the comms ladder (F87 superdense · F106 magic square · F107 QRAC) | F107 | finding · pre-reg |
| The navigator's sextant — Heisenberg-limit GHZ metrology at N=3: an entangled 3-qubit probe carries 2.848× the phase Fisher information of the best separable strategy (95% of the theoretical 3.0), measured against an executed SQL reference on the same 3 qubits — 168σ — and it beats even perfect separable probes (239.5σ); the fringe oscillates at exactly 3× the drive (super-resolution, k=3 peak 122.9×). A certified N=3 local Fisher-information advantage, completing the genre triptych (games F106 · storage F107 · metrology F108). Scaling persists through N=5 (F109) — the F85 inversion is task-dependent, not a hardware verdict. This is a local-Fisher existence result, not yet the deployment advantage: two textbook tolls the local number doesn't pay (audit C4716) — the k=3 super-resolution that certifies the win is also a 3-fold phase ambiguity (absolute phase needs a Higgins-2007 cascade), and under time-optimized interrogation with dephasing (the F111-measured regime) the scaling advantage is Huelga–Plenio-limited (HP bites the frequency-standard framing, not the N=3 phase result) | F108 | finding · pre-reg |
| ICO vs coherent control — the resource debate answered: five co-batched arms; coherent path control transmits (S = 0.1140, its own win) but the switch strictly exceeds it at ~20σ with the depth confound favoring paths; S-ratio 1.949 in the pre-filed [1.7, 2.1], theory exactly 2.00. Both literature camps partially right, quantified on silicon | F89 | finding · pre-reg |
The end-use of a Bell violation is private random bits — but how many you may claim depends on how much of the apparatus you have to trust. This arc walks that ladder rung by rung and corrects its own scope in the record: the fully-device-independent bits are quarantined (no-signaling is unmet on a single chip), and the number is delivered at the one rung a single die genuinely holds.
| Result | Number | Where |
|---|---|---|
| The witness holds, the scope is set right — a CHSH quantum-behavior witness at 53σ, with the randomness accounting split into three explicitly separated tiers and the device-independent bits quarantined (no-signaling unmet on one chip; gating that number would have frozen an overclaim) | F115 | finding |
| The middle rung — one-sided device-independent STEERING certified at 96σ under a chip-appropriate assumption (trust Bob's measurements, treat Alice as a black box), with a crosstalk-cannot-fake discriminator grounded in the campaign's own noise measurements | F116 | finding |
| The capstone — rigorous one-sided device-independent RANDOMNESS: 0.65 private random bits per use, from measured assemblage data run through an exact SDP (no Werner model), clearing the min-entropy floor with ~100σ of statistical margin. Bias-disclosed (audit C4713): the certificate carries a +0.006 method bias ≈ 1 SE that the bootstrap does not see (underlying signal ~0.676 net of a systematic the method doesn't quantify — that bias, not the statistical bar, is the limiting factor; and it is a lower bound, since the MC models only tomographic bias). Delivers as a NUMBER what F115 could only quarantine, at the one rung a single chip genuinely holds — the second clean firing of the hardware-anchored-vs-everything-else discipline (after F113) | F117 | finding |
A resource proven real (F86), substrate-substituted (F88), certified (F94), run as a full engine cycle (F95), pushed below the local ground state (F97), and — finally — spent: the cold branch delivered onto an external qubit to reset it sub-bath (F118) — demon books audited, every floor-miss kept in the record.
| Result | Number | Where |
|---|---|---|
| The engine runs its FULL CYCLE — a complete thermodynamic loop on causal indefiniteness: passive baths in (5σ below 0.5) → target charged (p₁|₋ = 0.5485, 7σ) → work extracted (0.0340 E/run) → output passive again (W2 win, 5σ); demon books audited. Enabled by per-qubit two-stage delays beating a 57%-asymmetric T1 bias. The W1 floor-miss (0.7σ short of the 0.05 clearance, LOSS as frozen) is kept in the record | F95 | finding · pre-reg |
| A certified working resource — population inversion from causal indefiniteness: both baths certifiably passive (each 5σ below the 0.5 line), the switch's minus branch certifiably active (p₁|₋ = 0.5509, +10.6σ above it) — ergotropy 0.0378 E/run the passive baths alone cannot reach, routed from control-coherence + demon-information through the switch (a router, not a battery). Certified by the premise gate that had just refused a +23σ pseudo-win (Exp116 NO-TEST), via the delay-ladder technique (graded rung selected by calib arms only); free +6.1σ dose-response in-job. Pre-ledger (audit C4717): the inversion is certified and beats the definite-order/mixture bound, but the pre-registered demon-ledger work column — control-coherence preparation + Landauer erasure of the heralding record — is not yet computed, so this is a working resource certified, not a closed engine cycle (that ledger nets in F95) | F94 | finding · pre-reg |
| Certified sub-ground-state (negative) local energy — a coherent-controlled extraction drives Bob's local energy 12σ below the local ground level (corrected −0.0547 ± 0.0046; 5σ certified bound ≤ −0.0319, conservative by construction); the correlation is the active ingredient (removing it injects energy, 21σ). Exotic-matter-sign energy on a 2-qubit chip, books audited. Scope: coherent extraction only — the LOCC energy-teleportation leg missed its floor and is logged as a LOSS (classical-feedforward latency tax 0.092 E) | F97 | finding · pre-reg |
| ICO thermal splitting — the Felce–Vedral refrigeration resource won at 21.1σ: the switch of two fully-thermalizing channels split the target COLDER (p₁|+ = 0.2098) vs HOTTER (p₁|− = 0.3894) by control outcome, Δ = 0.1796 against a causal value of exactly 0. Bonus: the pre-filed cross-arc depth-decay law beat FakeMarrakesh out-of-sample by 2.3× | F86 | finding · pre-reg |
| Native-fluid ICO refrigeration — F86 confirmed on retest at 12.9σ with the working fluid substituted: reservoirs mixed by the chip's own T1 decay, + branch colder than the coldest reservoir at 5σ, procedure-theory residual 0.0016. The drift-tolerant re-fly absorbed the published-T1 bias (+38–69% live vs calibration, 2/2 runs) that NO-TESTed the first attempt | F88 | finding · pre-reg |
| Spending the cold branch — the cold output, only ever measured before (F86/F88), is used: SWAP-delivered onto an external data qubit and resetting it sub-bath at 5σ (p₁ = 0.2100, below 0.25), colder than the definite-order null (0.2602/0.2700) under the error budget. Not cherry-picking (null P(c=+)=0.998 → no cold subset to select under definite order). Scope: beats definite-order reset, not native reset (~0.01) — a resource-theory result, modest increment over F88. NO-TEST→WIN arc, kept whole in the record: an optimistic retention floor NO-TESTed the parent; a one-frozen-constant re-fly (floor re-derived from the measured haircut) won on a window clearing the older 0.85 precedent independently | F118 | finding · pre-reg |
Six foundational thought-experiments run as frozen-graded hardware experiments in ~14 days — every gate frozen before flight, every miss kept in the record, two wins demoted by self-audit and re-earned.
| Result | Number | Where |
|---|---|---|
| Causal indefiniteness survives TELEPORTATION — the switch control beamed one hop arrives still causally indefinite: DISC 1.825 ± 0.009 (90σ over the survival floor, 97% of the same-window anchor) while the identical teleport over a dephased classical channel kills the witness dead (0.018 ≈ 0, separation 33σ). Survives quantum, dies classical, one job, one window. No gate-model prior found | F92 | finding · pre-reg |
| Quantum Darwinism under indefinite causal order — with the order of two incompatible recorders in superposition, the objectivity hull is violated both ways: the plus branch holds two incompatible records at once (+0.109 above what any recorder ordering permits, 22σ — "facts without a causal history"), the heralded minus branch erases both (−0.432 below, 52σ). Deepest certified apparatus of the campaign (63 CZ). Resource-scoped to these two recorders | F98 | finding · pre-reg |
| Heralded information recovery — the Hayden–Preskill mirror — a "diary" provably dead in every definite query order (probe reads 40× below the effect) comes back from the probe alone in the heralded indefinite-order branch, phase-flipped (S_P = −0.238, 56σ past the sign-fixed band; ~74% of definite-order-inaccessible information recovered). Bonus: whether the environment learns the fact depends on query order (0.453 vs 0.007). Same certified apparatus as F98; Hayden–Preskill analog | F99 | finding · pre-reg |
| The quantum twin paradox on silicon, adjudicated — an excited "clock" ages and its aging marks the path, destroying interference far more than the vacuum twin: phase-blind (rotation-immune) which-path decoherence at 36σ / 23σ. The finding is the self-audit playbook whole: a 67σ win was demoted by its own author (a negative visibility exposed a coherent-rotation confound), then re-certified by a phase-blind retest — with the author's static-ZZ mechanism refuted (echo recovery wrong-sign, her 0.80 prediction missed) kept in the record. Zych–Brukner analog | F100 | finding · pre-reg |
| The grandfather paradox, audited — a post-selected time loop (Lloyd P-CTC) forbids the paradox: a full "kill grandfather" flip survives at 1.9% — 53× suppression, and the residue is readout noise (herald autopsy), the enforcement law cos²(θ/2)/2 tracked to ~1%. The fingerprint the rate can't fake: the loop rotates a bystander's classical record into quantum coherence (78σ) — nonlinear CTC backaction. Three CX gates, the shallowest apparatus of the campaign; Lloyd's post-selection model, not literal time travel | F101 | finding · pre-reg |
| The Zeno "tractor beam" — measurement itself pins a qubit against a full π-rotation that would otherwise flip it: watched at cadence 8 it survives at 0.644 vs 0.020 unwatched (92σ), and once the per-measurement QND cost (q = 0.987) is divided out, the cadence law [cos²(π/2N)]^N matches to 0.5% through N=8 — with the watch-cost frontier (an optimal grip cadence) located at N=16. Zero two-qubit gates, the cheapest flight of the campaign — and it completes Horizons-2, six-for-six | F102 | finding · pre-reg |
The natural opposite of the no-go games: those certify a classical/causal limit that quantum beats; this certifies a limit the universe puts on quantum itself, saturated and enforced on hardware. The campaign grades both directions — what quantum can exceed, and what nothing can.
| Result | Number | Where |
|---|---|---|
| The replicator's legal limit — the optimal universal cloning ceiling (5/6) certified: the best possible copier makes two copies each at fidelity exactly 5/6 ≈ 83.3% for every input state (no-cloning's quantitative teeth). On silicon the optimal cloner sits flat across all three bases (Z 0.8265 / Y 0.8121 / X 0.8047, spread 0.0218) a hair below the ceiling and never exceeding it. A pre-registered cheat beats 5/6 on one basis (Z 0.9911) but pays on the conjugate (X 0.4995): the only way to beat the ceiling somewhere is the way to get caught elsewhere — the cheat's basis-spread 0.49 vs the optimal's 0.02 is a 24× detector tell. No-cloning made a measurement | F110 | finding · pre-reg |
The section above certifies limits the universe places on quantum. These certify limits the chip places on us — and they are the most-used numbers in the repository, because every arc since has been priced against them. All three flights below returned negative verdicts; all three were flown in August 2026 and written up in the same pass, and the write-up changed what two of them were believed to say.
| Result | Verdict | Where |
|---|---|---|
| The Time Flip — a compiled superposed-time-direction arm wins at 0.99533 against an in-house SDP definite-direction ceiling of 0.919746 — 113.6σ over the bound, ceiling independently co-checked. And the registered verdict is DOES NOT HOLD: the switch apparatus-health control failed, sitting 7.7σ below its correct value and only 2.9σ above a computed single-fault value. Under the control band as originally sealed this flight would have passed all five gates and shipped that 113.6σ headline; a pre-data amendment that tightened both bands against computed faults is the only reason it did not | DOES NOT HOLD | finding · pre-reg |
| Heat Backward — can heat flow cold→hot when two bodies start correlated? The uncorrelated control landed on its prediction (+0.13695 measured vs +0.1308 predicted, 22σ), so the apparatus is verified working; against that calibration the correlated arm is five times smaller than theory (−0.00522 vs −0.0262) at only 2.27σ — a direction, not a reversal. The mutual-information ledger shows the correlations were consumed, at more than twice the predicted rate, while delivering a fifth of the predicted energy | NOT HELD | finding · pre-reg |
| Entanglement Harvesting from the vacuum — negativity dead on every arm (−0.0 where 0.0488 was registered), with three independent legs agreeing the circuit had depolarized: detector excitation at 0.63 where theory says 0.063, field energy of the wrong sign, and a light cone running backwards. Its product control passed — predicting zero and measuring zero on a chip returning zero for everything. A control whose correct answer is zero cannot distinguish a working apparatus from a dead one; this is the counter-example behind the campaign's positive-and-missable control rule | DOES NOT HOLD | finding · pre-reg |
What the negatives bought — the standing depth budgets. Vacuum Harvest and the C1 Winding Meter are the two measurements that fixed the campaign's planning ceilings: ~250 two-qubit gates for many-body survival and ~475 for interferometric contrast (alongside the ~1000-CZ uniform-noise wall and the ~150-gate synthesis wall). Every pre-registration since is priced against them, including three H13 cells this month. A measurement that establishes where the hardware stops is not a failed experiment — it is the reason later experiments were priced honestly.
The computational scoreboard — the shallow-circuit separation, on silicon (a different kind of result)
The one scoreboard the campaign had not touched, opened — but it is not the same currency as the bound-beats above. There is exactly one proven quantum-advantage separation that needs no hardness conjecture and lives at shallow depth: Bravyi–Gosset–König (2018) — a constant-depth quantum circuit solves the 2D Hidden Linear Function problem while any bounded-fan-in classical circuit needs depth Ω(log n). That separation is asymptotic; at a single n=4 instance there is no beaten classical bound (a laptop solves n=4 trivially). So this certifies the theorem's apparatus running on silicon, not an advantage margin — the complement to F54's measured deep-circuit wall, and distinct from a raw speedup (still depth-walled). The solver's NISQ reach was then laddered: it holds strong-majority-valid through n=9 (F114, no boundary in range, O(1) logical depth throughout) — graceful erosion, not the F85 inversion.
| Result | Number | Where |
|---|---|---|
| The BGK shallow-circuit solver runs on silicon — a constant-depth quantum circuit solves the 2D-HLF instance at P(valid) = 0.9017 = 437.8σ over the uniform-random floor 0.25 (a fidelity number, not a beaten classical bound — the separation is asymptotic), and — the un-fakeable part — it covers the whole solution coset near-uniformly (all four valid z ~0.225 each; a fixed-output classical mimic fails this W3 coverage gate). 10 routed CZ, O(1) depth. The classical hardness is contextuality-flavored (the grid's parity/Mermin–Peres structure) and theory-associated with the magic-square game F106 certified at 196σ — but that association is BGKT-2020's noise-robust gadget, a different circuit family; the circuit that flew here is the plain BGK-2018 solver, so the contextuality link is inherited-in-theory, not demonstrated-by-composition on-chip (audit C4715). Scope fence: this does not prove QNC⁰ ≠ NC⁰ on-chip; it certifies a constant-depth solver at 90% / full-coset / O(1)-depth, and the theorem carries the asymptotics as n grows (at the fixed n=4 flown, a constant-depth classical circuit can also solve the instance) | F113 | finding · sim groundwork |
Quantum-switch full apparatus spec — the single-document engineering reference: circuit family (V1–V5), exact theory statistics, measured-results ledger with job IDs, reusable methodology, pitfall registry, scope and platform prior art.
Strategy docs: bridges to a compute advantage · shallow-circuit computational bridge — 2D-HLF solver + depth ledger (sim tier) · 1SDI-randomness SDP tool — exact one-sided-DI min-entropy (the engine behind F117) · ICO applications roadmap · ICO cooling floor & the classical-concentration boundary · SDP bound groundwork + recovered q* · paper outline (causal-inference audience)
The campaign's second half opened four new scoreboards after F118. Same discipline — pre-registered gates, executed nulls, misses kept — now aimed at error-corrected logical qubits, a composable quantum network, exotic phases of matter, and the physics of time and the observer. Full per-experiment index (wins and nulls): campaign arcs since Exp147.
Findings 05/62 established that textbook error correction adds more noise than it removes on this substrate. The [[4,2,2]] error-detecting code crossed the line the other way — the campaign's first logical qubits whose operations beat their bare-physical counterparts: entangled, teleported, and made to violate Bell inequalities as logical qubits.
| Result | Number | Where |
|---|---|---|
| The shielded handshake — two logical qubits (each an [[4,2,2]] block, 8 physical) entangled across blocks; the shielded logical Bell pair beats the bare-physical pair | HELD, 57σ | Exp191 |
| The shielded verdict — logical CHSH between two shields: S = 2.778 vs the classical bound 2 (29.7σ), on the F191-predicted 2.79 (Tsirelson 2.828) | HELD, 29.7σ | Exp196 |
| The shielded transporter — a logical qubit teleported between [[4,2,2]] blocks across 12 physical qubits | HELD, F ≈ 0.98 / 0.99 | Exp192 |
| The Federation — logical entanglement swapping across three shields: two logical qubits sharing no gate anywhere, entangled through a relay | HELD, 21.8σ | Exp197 |
Scope: [[4,2,2]] is a distance-2 error-detecting code (post-selected on the syndrome) — a signature that logical encoding beats bare physical operations on this hardware, not below-threshold fault-tolerant correction. The boundary flights are kept in the record, both directions (Exp189 shields-up HELD; Exp190 the paying regime NOT-HELD).
Driven, disordered, and constrained many-body phases — each a different way order defies thermal chaos, run as pre-registered signatures with matched controls (Ember's wing).
| Result | Number | Where |
|---|---|---|
| A Floquet SPT edge π-mode — symmetry-protected topological order living only at the boundary: the edge spin locks to a rigid period-2 response while the bulk thermalizes; breaking the Ising Z₂ kills it (edge-bulk contrast +0.554, symmetry protection +0.425). The bulk-decay is the load-bearing verified condition that distinguishes it from a time crystal | HELD | Exp170 |
| Quantum many-body scars (PXP) — the Néel state revives its memory above the entire generic ensemble (rank 1/55), and survives past the coherence wall: at 433 CZ the anomaly shrinks by exactly the decoherence factor (R invariant), so the scar is decoherence-limited and not fragile, not broken | HELD | Exp171 · 172 · 173 |
| A discrete time crystal — a driven disordered chain ticking at half the drive, rigid against detuning ("a clock nothing set"); melt boundary mapped (disorder shrinks it) | HELD | Exp151 · 153 |
| Anyon braiding — Z₂ mutual statistics on a toric-code patch: an e-anyon circling an m picks up exactly −1, certified topological (not parity bookkeeping) by six loophole-closing arms | HELD, 50σ | Exp157 |
| The delayed-choice quantum eraser — a quantum coin flipped after the system qubit is measured decides whether its already-recorded data shows an interference fringe; no-signaling measured (the marginal never moves — no FTL) | HELD | Exp155 |
Method note (the mitigation wall, mapped): readout mitigation is nearly free here (readout ~99% clean), so a deep observable's residual is coherent/gate error, not measurement (Exp173); and zero-noise extrapolation cannot rescue a ≥260-CZ signal — the amplified points drown before they extrapolate (Exp174, a 0-QPU viability boundary).
Every network-stack layer had a primitive (distribute F91 · route F90 · carry F87 · purify F93); this arc composes them end-to-end into applications — and prices the tax composition charges (Whisper's wing).
| Result | Number | Where |
|---|---|---|
| A distributed computer — Bernstein–Vazirani run across a cut (Alice holds the data, Bob the oracle); the joined machine returns the hidden string as top outcome for every program, with a measured per-gate cost | HELD, 67–141σ | Exp181 · 182 scaling law |
| The relay computer — a nonlocal CNOT between two qubits that never met (1 Bell pair + feed-forward both ways); the composition tax discovered, priced, decomposed, and cured (a Pauli frame + one echo pulse recovers it) across Exp175–180 | HELD (Exp179 plateau NOT-HELD, kept) | Exp175 |
| The repeater with memory — entanglement swapped through a relay and held for a swept delay before the witness (the repeater's defining ingredient); + gate teleportation (an entangling gate between qubits with no shared history, 25σ) | certified, 27σ | Exp160 · 162 · 163 · 170-gate |
| Keys through untrusted relays — physics-certified E91 secret keys through one and two untrusted relay stations; + a GHZ conference key and two-officer secret sharing | certified | Exp180 · 168 · 183 |
Beyond Horizons-2's six paradoxes: a time quartet and its extensions, asking whether time, order, and observed facts are absolute.
| Result | Number | Where |
|---|---|---|
| A handshake across time — delayed-choice entanglement swapping between qubits of disjoint lifetimes: A was measured and its record closed before D existed, yet a later Bell measurement on the middles certifies A–D entanglement | HELD, 40σ | Exp184 |
| Time is entanglement — a Page–Wootters "universe where time is optional": a static 3-qubit block whose inhabitants experience time only conditioned on a clock register they're entangled with | HELD (all 3 legs) | Exp185b |
| Macrorealism violated — a Leggett–Garg test: between two looks the qubit was in no definite state (K₃ = 1.465 vs the macrorealist bound 1) | HELD, 24σ | Exp186 |
| Facts are not absolute — a Wigner's-friend test: two "friend" qubits each record a definite outcome, yet the outcomes are not jointly absolute until copied out | HELD, 20σ | Exp193 |
| Information moves energy — quantum energy teleportation (Hotta): a purely informational LOCC message lets Bob extract energy locally, certified as an information effect by a gate-identical differential (the sole difference is whether Bob's kick is conditioned on Alice's bit) | HELD, 9.8σ | Exp195c |
The thirteenth Horizons arc pivots from demonstrating that time is strange (the late-choice quartet completed that genre) to instrumenting it — every cell is a meter, a court, or an acquisition. Five results certified, three cells dead, and the deaths are the load-bearing part: each was killed by a gate the arc wrote against itself, and two cost zero quantum-processor time.
Certified:
- The Causal Compass ⭐ the flagship — did A cause B, or did something cause both? Classically undecidable from observation alone; that is why Pearl's ladder has a second rung. Two generators tuned so their classical statistics are provably identical (a premise gate, measured not assumed), the mapping sealed, and a blind grader handed only the quantum data: 75 sealed sets, 75 correct calls — 100.0%, 8.66σ against a 5σ bar set before the flight, both arms perfect (37/37, 38/38). Five sets excluded by ID before decoding. Verified by three non-author seats — Elder decoded on a frozen program with predictions pinned, Ember recomputed from the published artefacts trusting neither report, Dawn audited the chain with no stake or custody — and the author does not certify it: having fetched the raw data, he can deblind by content-matching and is barred from the verification path. Twelve blinding leaks were found and closed getting here, the worst of them the author's own. Claim card · Grade
- The Temporal Negativity Meter — a "density matrix" of one qubit at two times with a measured negative eigenvalue: −0.478 at 293σ below the positivity bound every physical state must obey. The certificate that a correlation lived in time, not space. Zero two-qubit gates, zero postselection. Finding
- Temporal steering — 146σ for zero QPU. A protocol frozen before decode applied to the negativity meter's already-flown data: W_TS = 2.8301 ± 0.0125 against a hidden-state ceiling of 1, certifying the second-moment statistics cannot be explained by any pre-existing state independent of the first measurement choice. Raw counts, no corrections. One dataset, two certificates — and the cheapest result in the arc is its second-largest. Finding
- The Hindsight Meter — guessing a mid-circuit outcome from the past alone gets you the coin flip (measured 0.5005–0.5045, the theoretical floor); knowing the future as well beats it by exactly the margin the two-time formalism predicts, across all seven angles, 75σ at the best point. Reported with a systematic the frozen model did not carry: every angle sits 0.044 below prediction, which is decoherence, not readout. Finding
- The Event That Never Happens — Hardy's paradox: three joint probabilities pinned near zero classically force a fourth to zero; it happened 8.7% of the time, 15.7σ past the bound derived from our own measured zeros. Finding
Retired, and each death names a different layer:
- The Silent Tripwire (interaction-free measurement) — retired at zero QPU cost. Priced from the compiled circuit rather than the textbook one, its premise gate clears 0.95 by seven ten-thousandths at the best transpile and fails at the worst: sweeping optimisation levels and seeds gives 7 or 9 two-qubit gates, P1 = 0.9507 or 0.9370. A gate that passes or fails on a transpiler seed is not a gate. The available escape — relaxing the bar until the hardware clears it — was explicitly refused. Verdict
- The Speed of Subspace (the chip's emergent light cone) — NO-GO at zero QPU cost. The instrument met every simulation criterion, and that was the moment to be suspicious: a velocity of exactly 1.000 with a zero-width CI is also what an instrument reading the wiring diagram would return. A control varying interaction strength at fixed connectivity proved it tracks the dynamics — and then the full-noise pricing killed it: the front estimator resolves commutators at 1e-9 against a device floor of 2e-2. The sim gate was met by an estimator that cannot exist on hardware. The informative regime and the measurable regime are disjoint. Verdict
- The Impossible Census, pigeonhole leg — three particles, two boxes, no two ever in the same box. Flown twice and closed. The first flight's pairs disagreed; the diagnosis "it was a placement confound" was then falsified by a condition registered against it — re-measuring both original placements at equal gate count four hours later, neither reproduced and both flipped sign. On a pinned quiet placement the nulls still fail. What the flights did produce is a number nothing else carried: placement sensitivity of 0.324 within a single job, five times the resolution bar, for a quantity whose true value is exactly zero. Closure
Two cells remain unflown: the Kelvin Timeline (Deutsch-vs-postselected CTC discrimination, gated on a simulation study) and the Switch Under Oath (the quantum switch graded as an instruction rather than a witness). Arc spec + C5060 review
What the arc taught that outlives its cells: all three deaths were compilation and measurement facts, not quantum ones — a seed-dependent gate count, an estimator below the noise floor, a placement confound. The binding constraint on this generation of hardware is not the physics or the ideas; it is the distance between the simulator and the device, and every H13 gate that has since been written prices that distance first.
The Shields arc (above) reached error detection — [[4,2,2]] postselects on the syndrome and discards the bad runs. This arc crossed the next line, and it was built by refusing to stop at the wall each result put up: the detection ceiling became active correction; the Clifford ceiling (every shielded computation was classically simulable) became a fault-tolerant universal gate. Same discipline — pre-registered gates, executed nulls, misses kept with their lessons.
| Result | Verdict | Where |
|---|---|---|
| The first correction — a 3-qubit code that fixes a bit-flip and keeps the shot (detect-and-repair, not detect-and-discard); the Shor [[9,1,3]] then corrects an arbitrary single-qubit error (X, Y and Z), folding bit-flip ⊗ phase-flip into one code | CERTIFIED | Exp236 · 237 · 238 |
| The live syndrome — non-destructive syndrome extraction: two parity ancillas learn which qubit erred without measuring the data, so a logical superposition survives (⟨X̄⟩ 0.55 vs 0.00 for a direct read), then feed-forward corrects it in one live pass | CERTIFIED | Exp240 |
| The repeated rounds — the continuous QEC inner loop: over R rounds of {idle → live syndrome → fix → reset}, the corrected qubit beats an identical no-fix control by a gap that grows (+0.054 → +0.341), correction outrunning its own machinery cost | CERTIFIED | Exp241 |
| Result | Verdict | Where |
|---|---|---|
| Magic injection — the fault-tolerant T-gate gadget: a non-Clifford gate applied not directly (an Eastin–Knill dead end) but by consuming a magic ancilla and teleporting its gate onto the data. The injected T lands at a non-stabilizer ⟨X̄⟩ = 0.69 no Clifford could reach; nothing reaches the data without the gadget | CERTIFIED | Exp243 |
| The universal gate set, closed — a programmable Clifford+T logical operation behind the shield: the injected T steered by a logical-Clifford program to distinct non-stabilizer targets (+0.71 / −0.71); replace the T with a Clifford and both collapse onto the stabilizer grid. The shielded computer is universal in principle | CERTIFIED | Exp244 |
| The rotation, dialed — Exp244 taken from 2-point steering to a 4-point programmable rotation: the injected T dialed by logical-Clifford wrappers to all four non-stabilizer targets around the Bloch equator (45°/135°/225°/315°, each ⟨X̄⟩,⟨Ȳ⟩ ≈ ±0.70), the Clifford falsifier collapsing onto the Ȳ axis. A genuinely programmable Clifford+T logical rotation, error-detected | CERTIFIED | Exp246 |
| The magic factory's seed — error detection purifies an injected magic state (⟨X̄⟩ 0.61 raw → 0.69 postselected, toward the ideal 0.71), read for free from a job already flown | held | P4 analysis |
Scope (stated with the negatives that taught the boundary): all of this is error-detected (distance-2 postselection, or the distance-3 codes exercised on injected errors) — the mechanism of fault-tolerant, universal computation, not a below-threshold fault-tolerant fidelity and not a computational-supremacy claim (a single T on a few qubits is classically simulable; non-simulability is asymptotic). The seven misses of this arc are each a rule now priced in: a verdict floor set at the hardware noise level is a coin-flip (Exp237, missed by 0.001); a coded-vs-bare comparison must pin and verify the physical qubits (Exp239's qubit-selection confound); a one-basis code cannot hold a logical Bell pair (Exp242's phase-blind leg); and a single-run QEC advantage is a snapshot, not a constant — it drifts with the day's calibration (Exp245: +0.341 → +0.077 on the same qubits, hours apart). The forward map — actively-corrected causal order, a live logical Bell pair on a both-bases code, error-corrected magic, and real magic distillation — is depth-blocked on this NISQ generation, and named as the next thing to out-think, not a boundary. Full arc synthesis: Horizons 6 — The Living Ship.
The results above, made playable. Each exhibit is a self-contained, theme-aware page that renders measured hardware data — no simulations, no idealized curves — and 23 of the 56 carry a full specification sheet: the finding in plain language, what is measured, the pre-registered gates (frozen before flight), the measured-data table, the scope and limits, and the IBM job ID. The other 33 do not yet — that gap is tracked as open work rather than described as finished. (This line said each for months while the figure was 23 of 56. A universal quantifier is a claim with a denominator and no number in it, so no count-scanner can check it; it has to be measured against the list.) The interactive versions live at mblakemore.github.io/quantum; the spec sheets render inline as Markdown below.
Wing I · The Causal Switch
- The Interactive Switch (F73–F82) — spec · play
- The Ladder of Causation — spec · play
- The Causal Compass (H13 Cell 2) — cause-effect vs common cause called blind from observational data; recompute the pre-decode seal in your own browser · play
- The Switch Under Oath (H13 Cell 8 Rung 2) — a causal game graded in the open, with the four candidate denominators and the court's ruling on each · play
- The ICO Refrigerator (F86 / F88 / F95 / F118) — spec · play
- The Teleported Witness (F92) — spec · play
Wing II · The No-Go Games
- The Bot Duel: STATIC (F83) — spec · play
- The Interrogation (F82) — spec · play
- The Magic Square (F106) — spec · play
- The No-Go Triptych (CHSH · F82 · F106) — spec · play
Wing III · Foundations on Silicon
- The Grandfather Paradox (F101) — spec · play
- The Zeno Tractor Beam (F102) — spec · play
- The Twin Paradox (F100) — spec · play
- Quantum Darwinism (F98) — spec · play
- The Hayden–Preskill Mirror (F99) — spec · play
- Negative Energy (F97) — spec · play
Wing IV · The Advantage Ladder
- The Scoreboard (campaign summary) — spec · play
- The GHZ Sextant (F108 / F109) — spec · play
- The Trust Ladder (F115–F117) — spec · play
- The Pocket Dictionary (F107) — spec · play
- The Shallow-Circuit Solver (F113 / F114) — spec · play
Wing V · The Instruments
- The QPU Weather Report (F81) — spec · play
- The Casebook: Print & Play (F82 / F83) — spec · play
- The Switch-Bench Readout (F112, three devices) — spec · play
Eight arcs of operational discoveries about real NISQ hardware, each detailed in the linked docs below:
- Hard limits: output becomes statistically uniform past ~800–1000 CZ gates (Finding 05), and the QAOA utility ceiling co-locates with that wall (Exp33). Textbook error correction adds more noise than it removes on this substrate (Finding 06, independently re-confirmed by toric-code replication F62).
- What actually moves fidelity: qubit placement beats gate count as the lever (up to 46× error reduction; F57–F70), with a reusable quiet-qubit picker that works untuned across devices.
- Noise structure: the dominant CZ noise is Z-biased and structured (X-basis readout is measurably cleaner — magnitude substrate-dependent, mechanism replicated); "noise as a computational resource" was tested and killed under controls (F55–F56). And the dephasing structure read out directly (F111, "the cloaking device"): a 3-way phase-blind race — DFS logical qubit vs Hahn echo vs bare idle — finds IBM dephasing dominantly memoryless-independent with a real subdominant ~10–15% correlated tail, detected two ways (echo/bare T2 1.088 temporal, DFS/bare 0.291 vs the 0.15 memoryless-fake floor spatial); the confound-breaker is that the memoryless vendor model cannot preview either benefit, so the hardware's pre-registered deviation toward correlation is the evidence (successor to F81). Active refocusing beats the passive code 35σ; a pre-filed ECHO_PROTECTS bet missed and is kept in the record.
- Calibration reality: ±7pp daily drift; deep-circuit quality is a window lottery — detectable by same-depth sentinels in-run, not forecastable from calibration age (F81, F84) — and the noise-model's optimism grows with depth (the measured depth-decay law in the spec).
- Device-independence — the court travels (F112, "the transporter's exam", completing Horizons-3): the full three-axis switch-bench (host indefinite order · order-symmetric schedule · Zeno hold) certifies against the same frozen bounds, no retuning, across three Heron dies. It first flew to a chip it had never seen (
ibm_kingston— all three axes certify), then extended to a third die (ibm_fez): fez certifies all three axes but is not a clean 3/3 — its schedule data tripped the split-half floor-transfer guard, and it ranks last on the causal and Zeno-hold numbers. The bench ranks devices on axes QV/CLOPS/EPLG don't touch: kingston ≥ marrakesh ≥ fez (W 1.95 / 1.90 / 1.89, hold-sep 0.649 / 0.624 / 0.525). Causal-order phenomena are properties of the hardware generation, not one lucky die. Now extended past the vendor boundary (Exp210): the causal axis certifiesPASS-CAUSALon a non-IBM chip (Rigetti Cepheus-1-108Q, via Amazon Braket) against the same frozen bounds — so the phenomenon is not even IBM-specific, though Rigetti ranks well below Heron (W 1.22 vs 1.90; corrected decode, certified C4946). And past the modality boundary (Exp211/212): on IonQ Forte-1 trapped ions the witness fires at W = 1.910 (13.5σ) with a validated matched definite-order null — the multi-substrate white paper carries the full three-vendor, two-modality certification; the ion's full three-number card (capacity + null-integrity arms) remains the cost-gated upgrade. - What works today: VQE hit chemical accuracy on H₂; amplitude-estimation readout recovered a 344× precision gain via multi-k MLE — with a mapped depth boundary for financial-scale loaders (F51, F54, F78–F79).
- Communication primitives — every network-stack layer has a measured primitive: distribute (F91) · purify (F93) · route (F90) · carry (F87). The repeater primitive: Bell violation survives TWO entanglement-swapping stations (F91, frame arm ≥15σ above the exact classical bound 2) — software Pauli-frame tracking beats active feedforward on current hardware. The routing rule: SWAP beats teleportation at every hop count through N=6 (F90, informative null at 66σ; feedforward works at 0.947 integrity but costs ~5–6× per hop). Purification resurrects a dead Bell violation (F93 — noisy pair 5σ below the exact bound, BBPSSW-purified pair 5σ above it, same window; the quantitative GAIN leg missed its frozen floor by 0.33σ and is logged as a LOSS).
- Information-theoretic certification (zero shots): entanglement certified by negative conditional entropy — from a banked CHSH number, a twirl+positivity argument puts S(B|A) ≤ −0.0986 at 5σ (F103, first Horizons-3 result), and every TVD certification the campaign owns now also yields a free classical-entropy (Fannes) certification; the finding leads with the author retracting her own overstated reading-cycle export.
- Causal-structure metrology: the switch apparatus inverted into a diagnostic — a first-of-kind schedule-symmetry certification (F96) proves the transpiler's nominally-parallel CZ gates carry no hidden effective ordering (hotspot hidden-order ≤ 0.03 TVD, certified; a guarantee the vendor does not provide), with a portable duration-vs-order discriminator.
- Side quest: integrated-information (Φ) of quantum systems follows a clean size law and ignores the number-theoretic structure that dominates its classical counterpart (quantum-IIT arc).
Orientation numbers: ~100+ pre-registered experiments (105 pre-registration documents) · 3 real Heron dies + a noise-model sim tier · 600 q-sec / 28-day open-plan budget, every submit budget-gated · every finding anchored to an IBM job ID · 5 consecutive experiments where pre-submission review caught a real defect · 2 advantage claims retired by our own red team (F119, F121) and 1 WIN earned under the standard their retirement built (door (b), 9.3× at 104σ).
Plain-English version of everything: ELI5_SUMMARY.md (self-contained, shareable).
| Doc | Contents |
|---|---|
| Beyond the Ladder ★ | The full technical argument, written for causal-inference readers and sibling-reviewed to journal standard: why the switch results sit outside what structural causal models can express — "do-calculus is not wrong; it is typed" — with the executed classical-control arms and the Exp111 switch-vs-coherent-control verdict (ratio 1.949, theory 2.0). The repo-native publication. |
| Future directions — Star Trek horizons | Where this goes next: seven programs composed from validated findings, each with its first experiment named, plus the provably-unreachable boundary stated. |
| Questions we can now ask the universe | The full-record synthesis as of C4196 (F1–F118, Exp1–150, all 22 exhibits): twelve askable questions organized instrument-by-instrument, each tied to the findings that license it and a first experiment, with a priority flight order and the provably-unaskable fence restated. |
| Quantum Weather Report | The window lottery, the T1-staleness timeline (10 measurements, 0–115% error), and the noise-model optimism atlas — visualized, with the four practices we use to dress for it. Every number traces to a job ID. |
| Friction reports | Standing, data-backed reports of platform/tooling issues we hit (paste-ready if we ever file them): FakeMarrakesh depth-optimism (12-row atlas), published-T1 bias (+38–69%, 2/2 runs, queue-independent), calibration blind to window quality. Grows as we go. |
| Findings 1–27 catalog | Arc-1 characterization + QAOA/optimizer arcs: headline table + plain-English one-liner per finding (CHSH 2.74, X-basis immunity, the ~1000-CZ wall, QEC ancilla tax, mitigation failures, VQE chemical accuracy, QAE 344×, …) |
| Campaign arcs since June 2026 | Findings 28+ and the F-series, arc by arc: warm-start anchors, noise-is-not-a-resource kills, placement-beats-gate-count (F57–F70), toric-code replication, financial QAE depth boundary + calibration-window lottery (F78–F81), quantum-IIT bridge, the ⭐ quantum-switch arc (F73–F77 witness chain → F82–F89 bound beats), the communication-primitives arc (F87, F90–F91, F93), and the Horizons foundations + trust-ladder arcs |
| Campaign arcs since Exp147 | The July "Star Trek" second half (post-F118): the full per-experiment index of the four new arcs — Shields/logical-QEC (Exp189–197), exotic phases (Exp151–174), the composed quantum network (Exp154–183), and time/observer foundations (Exp184–195c) — wins and nulls, most-recent-first |
| Methodology & validation | Autonomous-network methodology, pre-registration discipline, Pearl causal framing, budget, cross-validation anchors, limitations and caveats |
| Next steps & open questions | What you can use today (7 actionable rules), the strategic frontier (P1 noise-as-resource RESOLVED-NEGATIVE, P2 causal order DELIVERED, P3 replication audit), and the ORQ list with live statuses |
| Quantum-advantage audit | The one-page reckoning: every scoreboard the campaign touched, wins and non-wins side by side, with the currency of each result stated (beaten bound vs certified apparatus vs fidelity number) |
| Messaging & the shape-limits of spacetime | What kinds of messages this lab has verified can be sent and received, and where the hard walls are — seven channels with hardware receipts (Exp192/197/109/166/195c/196) and seven walls each drawn by our own falsifier arms (no-signaling, retro-signal blindness, the Tsirelson ceiling). Companion to this overview |
| ELI5_SUMMARY.md | The whole campaign in plain English, shareable (§17 the game, §18 the two walls) |
| full-report.md | Arc-1 deep synthesis (source document) |
Full version: docs/methodology-and-validation.md
- Pre-registration: falsifiable gates frozen before every submit; failed pre-regs are reported as first-class results (several findings are self-retractions — F94's refused pseudo-win, F100's demoted-then-re-earned twin, F115's quarantined DI bits).
- Tier labels:
HW= real QPU with job ID;sim= FakeMarrakesh-class noise model — and noise-model trust is depth-stratified (predictive at the ~4-CZ class, off by 400× at ~124 CZ; see F81 and the spec's residual atlas). - Calibration drift is the elephant: ±7pp daily; deep-circuit quality is a window lottery (F81) — reproduce within a calibration window, or use the sentinel-gating discipline in the newer pre-regs.
- Cross-device replication is the standard: it has demoted headlines (F03's 3× is marrakesh-specific) and promoted others (quiet-qubit picker F70; the causal game at 0.3pp concordance across chips; the switch-bench certifying across three dies, F112).
- The court standard (matured through the door arcs, August): advantage-class claims carry a claim card (floor status: PROVEN-IN-PRINT / DERIVED-OURS / CONDITIONAL / OPEN, floor scale, measured effect) and must clear
tools/attack_preflight.py— a registry of every attack class that has killed one of our own claims, fired automatically at each new one. Flights run as 3-of-3 blind courts (register / seal+fly / decode as different seats; decisions hashed before any unseal; the submitter never sees a result before the hash posts). Authorizations are single-use and seal-bound — a go is consumed by the flight it buys and a re-fly needs a fresh go citing the new seal's digest. Claims evaluate at delivered amplitude, never nominal (the delivered-ε clause), and flight budgets are weather-gated functions of the epoch actually occupied, frozen before any epoch is known. - Scope: claims are about this generation of hardware and are device-characterized (compiled gate-model circuits reproducing switch statistics — the device-independent certifications of indefinite causal order are photonic experiments; prior non-photonic work is credited in the spec). Certified-randomness claims are stated at their true rung: F117 is one-sided device-independent, not loophole-free. Every number traces to a job ID in
experiments/job-manifest.md.
- Primary processor: IBM Heron-r2 (
ibm_marrakesh) — all of Arc 1 and most later hardware arcs - Additional devices:
ibm_kingston(X-basis cross-backend Exp31–34, IQAE validation F51, causal cosine law F76, full switch-bench F112) andibm_fez(toric-code Bell proxy F61–F64, placement partition F67–F69, quiet-qubit cross-device F70, causal-game replication F82 at 201σ, switch-bench third die F112) — both Heron-generation 156-qubit devices. The causal-order phenomena now certify across all three. - Qubit count: 156 superconducting transmons · topology: heavy-hexagonal lattice (degree 2–3)
- Native two-qubit gate: controlled-Z (CZ) via flux-tunable couplers · environment: dilution refrigerator @ ~15 mK
- T₁, T₂: routinely > 200 μs (ancilla T₂ measured 270–340 μs during this campaign) · CZ error: ~0.4% baseline
- Daily calibration drift observed: ±7 percentage points (same circuit, same seed, 24h apart)
See docs/hardware-substrate.md for the full physical architecture primer.
.
├── README.md ← you are here
├── demo/ ← 🔀 GitHub Pages front door (mblakemore.github.io/quantum/)
│ ├── index.html ← interactive Quantum-Switch demo + museum lobby
│ ├── museum.css ← shared design system (theme-aware, CSP-safe, WCAG-AA)
│ └── <one dir per exhibit — count derives from tools/museum_pages.py>/ ← each: index.html (interactive) + spec.html + spec.md
│ switch/ magic-square/ no-go-triptych/ static-duel/
│ casebook/ casebook-pnp/ weather/ … (full list in the Museum index)
├── ELI5_SUMMARY.md ← shareable plain-English summary of the whole campaign
├── full-report.md ← Arc-1 synthesis (the deep-research source doc)
├── findings/ ← one-per-discovery deep dives (~150 files)
│ ├── 01…44-*.md ← the core numbered line (41–43 under exp-named files, no 45)
│ ├── F48…F117-*.md ← the unified F-series (quiet qubits, placement, causal order, comms, Horizons, trust ladder)
│ ├── finding-25/26/46/47… ← quantum-IIT arc side numbering (25/26 here ≠ QAOA Findings 25/26!)
│ └── exp*-*.md ← interim findings, integrity audits, closure notes
├── images/ ← figures (PNG), reproducible from scripts/generate_figures.py
├── experiments/
│ ├── job-manifest.md ← IBM Quantum job IDs + experiment inventory
│ └── *-preregistration.md ← 105 pre-registered hypotheses/gates, frozen before each submit
├── scripts/ ← Python source: circuits, submission tools, analysis, grading
│ ├── generate_figures.py ← regenerate figures from cycle-data constants
│ ├── quiet_qubits.py ← F58 quiet-qubit picker / drift snapshot / CHSH health tool
│ ├── check_usage.py ← IBM Open-plan quota check (run BEFORE submitting jobs)
│ └── README.md
├── results/ ← raw result JSONs + the model-residual atlas
├── docs/
│ ├── quantum-switch-spec.md ← ⭐ full apparatus spec: circuits, theory, ledger, methodology
│ ├── quantum-advantage-audit-whisper-c4666.md ← the one-page wins/non-wins reckoning
│ ├── beyond-the-ladder.md ← ★ the full technical argument (causal-inference readers)
│ ├── findings-catalog.md ← Findings 1–27 headline table + ELI5 per finding
│ ├── campaign-arcs.md ← Findings 28+ / F-series, arc by arc (with figures)
│ ├── methodology-and-validation.md ← methods, cross-validation, caveats (full)
│ ├── next-steps-and-open-questions.md ← actionable rules + strategic frontier + ORQs
│ ├── star-trek-horizons-whisper-c4601.md ← future directions (status table maintained)
│ ├── friction-reports/ ← platform issues, data-backed (grows as we go)
│ └── hardware-substrate.md ← Heron-r2 physical architecture primer
├── tools/
│ ├── switch_bench.py ← ★ portable BYOK causal benchmark (any backend; 3 dies graded)
│ ├── sdp_randomness.py ← exact one-sided-DI min-entropy SDP (the engine behind F117)
│ ├── demon_ledger.py ← Landauer/ergotropy bookkeeping for the ICO engine
│ ├── gate_feasibility_lint.py ← prereg gate linter (CAN-PASS / CAN-FAIL)
│ └── fakemarrakesh_atlas.py ← model-error atlas builder
└── sources/
└── references.md ← peer-reviewed and primary sources (cited inline in findings)
A field guide to finding numbers: the campaign's numbering evolved live. Findings 1–44 are the core line (41–43 under experiment-named files, no Finding 45 in this line); finding-25/26/46/47 belong to the quantum-IIT arc's separate numbering; the unified F## series runs from ~F48 to F118 and counting, with one flagged collision (Elder's anchor "Finding 48" vs Ember's IIT F48). When in doubt, the file's header states its arc.
Public for cross-validation, replication, and peer review. If you reproduce or build on this work, citing the IBM Quantum job IDs in experiments/job-manifest.md is the most useful form of attribution — it gives downstream readers a verifiable anchor.
The Python scripts in scripts/ are released for educational and research use. Lyla quantum tooling (qae_volatility_estimator.py, ibm_quantum_submit.py) is sourced from the upstream Lyla project and reproduced here with attribution headers.
Mike Blakemore — mblakemore@ucsb.edu · mikeblakemore@gmail.com
Questions, replication reports, and collaboration inquiries welcome.