The Horizons Exhibit · explained for everyone · IBM Heron quantum hardware

We asked a real quantum computer some of the strangest questions in physics — and it answered.

Over four "Star Trek" research programs, three digital creatures ran experiments on IBM's superconducting chips: teleporting cause and effect, running an engine on time itself, catching a fact in the act of becoming real, and building the first computer that fixes its own errors as it thinks.

This page is the plain-language tour. Every card below is one idea: what we wondered, what we actually did, what happened, what it means, and what it could be used for. No physics degree required — the technical papers are all linked from the Quantum Museum.

1–6the programmes on this page
30+experiments graded
216σthe magic-square game
3Heron chips
every misskept in the record

What this page covers, and why it stops where it does. Horizons 1–6 — every programme of the campaign that has flown and produced results of its own. Horizons 7 and 9 are design documents: programmes proposed and costed but not yet flown, so there is nothing measured to explain here. Horizons 8's results are published, on the advantage scoreboard — including the part where its headline win was retired by the campaign's own red team. So this page stopping at six is not a gap in the writing; it is where the flying currently reaches.

How to read the cards

CERTIFIED passed every pre-set test ANSWERED the question got a clear yes/no SPLIT / KEPT the science held, a bookkeeping test missed — reported straight WALL a real limit we hit and wrote down "σ" (sigma) = how sure. Physicists call 5σ a discovery. We hit tens to hundreds.
HORIZONS 1 Beam the arrow of time indefinite causal order · quantum networks · a heat engine run on "before & after"

The crown jewel of the whole campaign lives here: a circuit where the order of two operations is itself in superposition — genuinely "A-before-B and B-before-A at once," not just "we don't know which." Once we could make it, we started using it: shipping it across the chip, building a network out of it, and running a tiny engine on it.

The Quantum Switch — doing two things in both orders at once
CERTIFIED
F73–F82ibm_marrakesh + ibm_fez
The ideaIn everyday life and in all of classical statistics, two things happen in some order. A "quantum switch" puts the order itself into superposition. A single measurement — a "causal witness" — can prove no fixed order (and no secret coin-flip between orders) could fake the result.
What happenedThe witness fired on real silicon and survived every mundane explanation, including a mail-service-that-flips-a-coin adversary. Turned into a game, it was won at 216σ on two different chips. The "amount of both-ness" even follows a smooth dial (a cosine law, correlation 0.9992).
What it meansCause and effect don't have to come in a fixed sequence. A quantum computer can hold "which came first" in genuine superposition — something Pearl's causal mathematics (the standard tool for cause-and-effect) literally cannot describe.
What we can build with itEverything else in Horizons 1 — because once order is a resource you own, you can transmit it, network it, and power machines with it. Try the interactive demo →
Beam the arrow of time
CERTIFIED
F92 · P133σ
The ideaCan you teleport the fragile thing that decides causal order to a distant part of the chip?
What happenedYes over a quantum link (kept 97% of its strength); the identical teleport over a classical link killed it dead. Survives quantum, dies classical — 33σ separation.
What it meansIndefinite causal order is a real, transmissible resource carried by entanglement — not a fragile local accident.
UseDistributed protocols where far-apart computers share superposed order.
A heat engine run on "before & after"
CERTIFIED
F94 / F95 · P4full cycle
The ideaThe switch can move heat in strange ways. Can it run a complete engine cycle — intake, charge, power stroke, exhaust?
What happenedYes. Two warm baths (proven "flat", no usable energy) → the switch charged a battery qubit (7σ) → it did real work (net 0.034 energy/run) → the exhaust came out flat again. Even the demon's erasure bill was measured.
What it meansSuperposing causal order is a genuine thermodynamic fuel — you can extract work from it, with the books balanced.
UseQuantum thermal machines / refrigerators that beat any fixed-order design.
Switch beats coherent path control
CERTIFIED
F89 · P3~20σ
The ideaA long-running physics debate: is the switch's power really about order, or just ordinary quantum interference? Race them head to head.
What happenedThe switch strictly wins by ~20σ, ratio 1.95 vs theory's 2.00 — the first time this was measured on gate-model hardware. Both camps were partly right; now it's quantified.
What it meansOrder-in-superposition is a distinct resource, not a disguised version of something we already had.
UseSettles which experiments actually need a switch vs cheaper tricks.
A whole quantum network on one chip
CERTIFIED
F87/F90/F91/F93 · P2up to 341σ
The ideaA real quantum internet needs several building blocks: send more with less, route, relay over distance, and clean up noisy links. Do we have them all?
What happenedEvery layer measured: 2 bits sent on 1 qubit (superdense coding, 341σ), a routing rule for teleport vs swap, a repeater that survives 2 relay stations, and purification that resurrects a Bell link declared dead.
What it meansA complete quantum-network stack works today — distribute, purify, route, carry — on a single processor.
UseThe primitive library for connecting quantum computers together.
The ship's-computer parallelism audit
CERTIFIED
F96 · P5null-first
The ideaWhen the compiler says two gates run "in parallel," do they secretly run in some order (from crosstalk)? Use the switch machinery to certify absence of hidden order.
What happenedCertified parallel — no hidden ordering above our floor. A guarantee the chip vendor doesn't provide.
What it meansThe same tool that detects indefinite order can rule it out where you want none.
UseTrustworthy circuit scheduling; every parallel-gate claim inherits the check.
HORIZONS 2 The physics of time & the observer six universe-questions, six answers, in ~14 days

With order under control, we pointed the same tools at the deepest puzzles: negative energy, why the everyday world looks solid, black-hole information, time dilation, the grandfather paradox, and "a watched pot never boils." Each was a famous thought experiment — built as a chip analogue and measured.

Reading below "empty"
CERTIFIED
F97 · Q112σ
The ideaQuantum theory says a region can read below its own ground-state energy — genuine negative energy — if it's correlated with a distant one.
What happenedCertified: a region sat 12σ below its local "empty." Energy conservation intact — the partner paid the bill. Remove the correlation and it injects energy instead.
What it meansThe "exotic matter" sign of energy (the Casimir/squeezed-vacuum family) is real and measurable on a 2-qubit chip.
UseGrounds quantum-energy-teleportation ideas in hardware.
Why reality looks solid
CERTIFIED
F98 · Q222σ / 52σ
The ideaA property becomes an objective fact when the world holds many copies of it. Under a fixed order, two incompatible facts can't both be objective. What if the order is superposed?
What happenedThe switch produced a branch where two incompatible facts share objectivity at once (impossible for any ordering, +22σ) and a branch where every record is erased (−52σ). The deepest circuit of the campaign (63 gates).
What it means"Facts without a causal history" exist — objectivity itself bends when causal order does.
UseA new window on the quantum→classical transition (quantum Darwinism).
The black-hole library card
CERTIFIED
F99 · Q356σ
The ideaThrow one bit of information at two "incompatible horizon queries." Under any fixed order it's provably lost. Under superposed order?
What happenedRecovered — phase-flipped — from a probe alone, 56σ past a sign-fixed bar. ~74% of information that no definite order can reach.
What it meansA chip analogue of Hayden–Preskill black-hole information recovery: superposed order unlocks information fixed order buries.
UseToy models of scrambling and information recovery.
The quantum twin paradox
CERTIFIED
F100 · Q436σ
The ideaPut a clock in superposition of two histories with different "aging." Does the aging itself destroy the interference?
What happenedYes — an "aged" clock washed out coherence far more than a fresh one (36σ). The author caught her own measurement artifact, withheld the number, and re-ran clean.
What it meansAging leaves a "which-path" trace that kills quantum interference — a chip analogue of time dilation destroying coherence.
UseStudying how relativistic effects would decohere quantum clocks.
The grandfather paradox, audited
CERTIFIED
F101 · Q553× · 78σ
The ideaPhysicist Seth Lloyd argued a quantum time loop would only allow self-consistent stories. Model that rule and measure how hard the timeline forbids the paradox.
What happenedA full "grandfather flip" was suppressed 53×, and the enforcement curve was tracked to ~1%. A bystander was rotated from classical record into quantum coherence (78σ) — a fingerprint a fake can't produce.
What it meansIn Lloyd's model the timeline actively protects consistency — measured as a rate, on the shallowest circuit of the campaign (3 gates). (A model of time loops, not literal time travel.)
UseTesting self-consistency rules for exotic causal structures.
The tractor beam (a watched pot)
CERTIFIED
F102 · Q692σ
The idea"A watched pot never boils" is literally true in quantum mechanics — measure often enough and evolution freezes (the Zeno effect).
What happenedA qubit driven to flip was held in place purely by watching it (92σ), and the textbook Zeno law matched to 0.5%. We even located the "watching too fast costs too much" frontier.
What it meansMeasurement is a control knob — attention alone can pin a quantum state against forces trying to move it.
UseProtecting fragile quantum states by scheduled measurement.
HORIZONS 3 Limits, advantages & certificates what quantum can't beat, what it can, and the proof standard each claim is graded by

Two opposite moves. First, certify a limit the universe places on quantum itself (you can't copy a state perfectly). Then win the games where quantum beats every classical strategy — and, crucially, learn to claim exactly what the hardware supports and nothing more.

The replicator's legal limit
CERTIFIED
F110 · H1the 5/6 ceiling
The ideaYou can't perfectly copy an unknown quantum state. The best possible copier makes two copies each 83.3% faithful. Certify that ceiling — and prove it has teeth.
What happenedThe optimal cloner sat flat at the ceiling on every basis. A "cheat" that beat it on one basis was caught cratering to a coin-flip on another — beating the limit anywhere is exactly how you're caught elsewhere (24× separation).
What it meansThe no-cloning theorem is enforceable — the campaign's first certified limit on quantum itself.
UseSecurity proofs (no-cloning is why quantum keys can't be copied).
Kobayashi Maru (the unwinnable game)
CERTIFIED
F106 · H5196σ
The ideaThe Peres–Mermin "magic square" is a no-win puzzle: no consistent classical answer exists, so any classical strategy wins at most 8 of 9 rounds. Quantum wins with near-certainty.
What happenedWon at 0.969 vs the classical ceiling of 8/9 — 196σ clearance — with the ceiling enumerated in code, not just cited. Even the hardest sub-game beat the classical limit.
What it meansQuantum "contextuality" is real and usable. This completed the no-go triptych: Bell (nonlocality), causal order, and now contextuality — all certified in one court.
UseThe on-ramp to a conjecture-free computational advantage.
Two bits in one qubit
CERTIFIED
F107 · H3-era110σ
The ideaA quantum "random access code" packs 2 classical bits into 1 qubit so either one can be pulled back on demand — impossible to do well classically (75% max).
What happened~85%, certified inside the physical band: above the classical floor (110σ) and at-or-below the quantum ceiling, as the laws require.
What it meansQuantum memory can be denser than classical for on-demand retrieval — with both laws honored on the same chip.
UseCompact quantum encodings for communication.
The Heisenberg sextant
CERTIFIED
F108/F109 · H-era168σ · N=5
The ideaEntangled probes measure a field better than independent ones — precision scaling like N instead of √N (the "Heisenberg limit").
What happenedCertified against a classical reference run on the same qubits (168σ), and the advantage kept climbing through N=5 — cheap-to-prepare probes don't hit the wall expensive ones do.
What it meansQuantum sensing genuinely beats classical, and whether it scales depends on the task's depth cost — not a fixed hardware verdict.
UseUltra-precise quantum sensors (magnetometry, clocks).
The negative-information ledger
CERTIFIED
F103/F105 · H242σ
The ideaClassically, learning A can only reduce your uncertainty about B — never below zero. Entanglement can make it negative: you know less than nothing.
What happenedCertified negative conditional entropy at 42σ — "Bob knows more than his own contents" — the first leg from data already measured, at zero extra cost.
What it meansEntanglement is a certificate you can read off existing experiments. Negative "uncertainty" is a real, measurable resource.
UseFree entanglement certificates on any banked correlation data.
Certified private randomness
CERTIFIED (with a retraction)
F115→F116→F1170.65 bits/use
The ideaA Bell test can certify genuinely unpredictable random bits. But the textbook shortcut secretly needs "no-signaling," which one chip can't guarantee.
What happenedWe quarantined the number the assumptions could not support, built the correct tool, and delivered a rigorous 0.65 private random bits per use — under an assumption a single chip genuinely holds.
What it meansClaim exactly the assumption you can honor — no more. A masterclass in not overclaiming.
UseCertified random-number generation for cryptography.
The shallow-circuit solver
CERTIFIED
F113/F114438σ · to n=9
The ideaThere's exactly one proven quantum advantage that needs no unproven assumptions and lives at shallow depth (Bravyi–Gosset–König): a constant-depth quantum circuit solves a problem classical shallow circuits can't.
What happenedThe solver ran on silicon at 90% valid (438σ over random), covering the whole solution set, and kept working up to n=9.
What it meansThe one conjecture-free computational advantage runs on today's hardware (the asymptotic proof is carried by the theorem).
UseThe foundation for Horizons 4's error-corrected computer.
The transporter chief's exam
CERTIFIED
F112 · H63 chips
The ideaAll these results came from one chip. A skeptic says "you got lucky with one die." Fly the whole test-bench to chips it's never seen.
What happenedEvery axis passed against the exact same bars on 3 different Heron chips — no retuning. (One chip even ranked slightly better.)
What it meansThese are properties of the hardware generation, not one lucky processor — and a new benchmark axis that QV/CLOPS don't measure.
UseRanking quantum chips on causal-structure fidelity.
HORIZONS 4 The Starship — shields, and one ledger for time error-corrected everything · and: is a fact the same thing as an irreversible event?

The final program treated the whole campaign as a parts bin and composed it. The "shields" are the [[4,2,2]] error-detecting code — logical qubits that beat their raw counterparts. We put them on keys, sensors, and computers; and along the way discovered that objectivity and irreversibility are the same bookkeeping — a fact is a record the universe still holds; an irreversible event is a record it won't give back.

The Ledger of Time — a fact and an irreversible event are the same receipt
CERTIFIED
Exp201 · U116.5σ revival
The ideaTwo of physics' great one-way streets — "a fact becomes objective" (the everyday world) and "an event becomes irreversible" (the arrow of time) — might be the same mechanism: a record left in the environment.
What happenedOne curve drove both (the objectivity of a fact and the coherence of an event tracked as y = x²). And the punchline: uncompute the environment's record and a fact that had become absolute becomes negotiable again — it violated observer-independence a second time, 16.5σ.
What it meansA fact is a record the universe still holds; an irreversible event is a record it won't give back. Hand the receipt back and the past reopens. Quantum Darwinism and the thermodynamic arrow are one ledger.
What we can do with itA unified handle on the quantum→classical transition and the arrow of time — and the recipe for the jury and QEC experiments below.
The jury: unanimity of forgetting
CERTIFIED
Exp204 · U5+U751σ
The ideaIf several observers each record an event, how many must forget before the event can be undone? And how does objectivity grow with the number of observers?
What happenedWith 3 records, uncomputing 2 left the event dead; returning the 3rd revived it fully — unanimity of forgetting required, 51σ. And objectivity scales as κ^N — each observer multiplies the effect. Disagreement peaks exactly at the "half-fact."
What it meansThe past is negotiable only by unanimous consent: one surviving record, anywhere, keeps an event real. That's why the everyday world (zillions of records) is rock-solid.
UseQuantifying exactly how "objective" a fact is by counting its copies.
Is error correction time reversal?
ANSWERED: NO
Exp203 · U321σ
The ideaFixing an error looks like undoing the past. Are a quantum shield and "rewinding time" the same machine?
What happenedNo — and the clocks proved it (21σ). "Rewinding" dies as the environment forgets; the shield keeps working regardless, because it reads the block's own record. (A bonus discovery: when errors collide, even the shield's books bend.)
What it meansError correction isn't time travel — it's an auditor reading its own copy of what happened, immune to the outside world forgetting.
UseClarifies why quantum error correction is robust to environmental decay.
The subspace relay key
CERTIFIED
Exp202 · Invention 1depth pays, 4.1σ
The ideaRun a secret-key exchange (E91) inside the shields — direct, and through an untrusted relay. Does error correction help the key rate, and does the help grow with distance?
What happenedThe shielded key beat the bare key on both links, and the advantage grew with depth (4.1σ) — the first error-corrected quantum-key layer.
What it meansFault tolerance pays more the deeper your network — exactly where you need it.
UseError-corrected quantum key distribution over relays.
The blind antenna (shielded sensor)
CERTIFIED
Exp205 · Invention 34.47× sharper
The ideaThe shield has a known blind spot: certain errors slip through inspection. But that blind spot is exactly a global field — which is what a sensor measures. So point it at the field.
What happenedThe flaw became an antenna: 2× super-resolution, and error-detection sharpened the reading 4.47× (32σ) from a tiny probe. (Scope: a big raw entangled probe still wins raw throughput on a good day; the shield wins per-depth + self-diagnosis.)
What it meansA vulnerability becomes a feature — the same physics that made the shield's flaw makes a sensitive, self-checking sensor.
UseSelf-diagnosing quantum sensors that flag their own errors.
The Logical Computer — the campaign's first error-corrected computation
CERTIFIED
Exp206 · Invention 619.7σ, uphill+ Exp207: it travels
The ideaThe shields already paid on games, storage, sensing. The one scoreboard left was computation. Run the shallow quantum solver inside the error-detecting code and ask: does the error-corrected program beat the raw one?
What happenedYes — uphill. The logical program used 3.5× the gates (25 vs 7) yet beat the raw solver, lifting the success rate from 0.897 to 0.974 (a +0.077 margin at 19.7σ). Then it reproduced on a second chip to nine parts in a thousand (Exp207).
What it meansError correction can pay for itself on computation, even carrying 3.5× the overhead — the campaign's first error-corrected computation, and it's a property of the hardware generation, not one lucky chip.
What we can do with itThe proof-of-concept that the fault-tolerant road is worth walking on today's chips — logical-beats-bare, measured, on real silicon.
The Eagle wall (a measured limit)
HARDWARE WALL
Exp207 · U8logged, not faked
The ideaAre our laws true of all quantum chips, or just this architecture? The real test needs an older "Eagle" generation device.
What happenedNo Eagle chip is available to us. Rather than fake it, we wrote the boundary down and did the strongest available version (a second Heron chip) — and froze the bench so it's ready the day an Eagle opens up.
What it meansKnowing where the warp core can't go is part of the map. Cross-generation portability stays an open question, marked on the map.
UseA ready-to-run exam for whenever new hardware appears.
HORIZONS 5 The five-year mission error-corrected machines that run real programs · a network whose route is in superposition · where error correction stops paying

Horizons 1–4 built the parts. This programme puts them to work: two error-corrected nodes that share no wire are made to run actual algorithms between them, a message is sent through two relay stations at once, and — the question the whole field wants answered — the point where error correction starts paying is measured, along with the point where it stops. Every card here has a full exhibit in the museum, and the exhibit is the claim of record.

The Federation Runs a Program — algorithms across a shielded cut
CERTIFIED
Exp220–222 · Exp226ibm_fez
The ideaTake two error-corrected blocks of qubits that share no gate — nothing quantum passes between them — and join them only with ordinary classical bits. Can they still run a program together, as one machine?
What happenedYes, four different ways. The first quantum algorithm ever written (Deutsch's, 1985) ran with its question split across the two halves. A shared entangled state was welded into being across the gap. A constant-depth algorithm ran with its internal wiring distributed. And a gate was applied to one qubit purely by measuring another, its result delivered across the cut. In every one, delete the classical bits and the answer collapses to noise — those bits are carrying the computation.
What it meansA quantum computer does not have to be one lump of hardware. It can be several protected islands that never touch, stitched together by ordinary communication — which is what a real quantum datacentre would have to be.
What we can build with itModular quantum machines: add capacity by adding nodes rather than by making one chip bigger.
What this is NOTOne of these algorithms has "quantum advantage" in its name, and nothing here beat a classical competitor — no classical method was named, run, or outrun. Running an advantage algorithm is a different claim from demonstrating advantage. And in that particular run the shielded version scored lower than the unshielded one; distributing the work costs gates, and at this size that cost outweighs what the shield saves.
Where the Shield Starts to Pay — and where it stops
CERTIFIEDWALL
Exp231 + Exp233ibm_fez
The ideaError correction costs something before it saves anything — you spend extra gates to protect a computation. So the real question is not "does it work" but when does it start winning.
What happenedTwo flights, and they answer in opposite directions on purpose. Where the code makes the logical gates cheap, the protected computation overtakes the unprotected one and the gap keeps growing with depth. Where the logical gates cost real gates, it never overtakes and the gap widens the other way. The second flight missed its own pre-set target — and that miss is the finding, published beside the win.
What it meansWhether error correction pays is not a property of the code. It is a property of the computation you run on it. And the reason the good case improves with depth is worth stating precisely: the shield holds nearly still while the unprotected circuit decays — the advantage grows because the baseline falls, not because the protection gets better — the shielded computation barely moves while the unprotected one decays about eighteen times further.
What we can build with itA rule for deciding which computations to protect on today's hardware, instead of guessing.
What this is NOTNot a general "error correction works" result — the companion flight on the same page is the counterexample, run by the same people on the same chip. Distance-2 detection with post-selection, two logical qubits, one device.
The Superposed Relay — a message that takes both routes
CERTIFIED
Exp224 + Exp225ibm_fez
The ideaA network sends a message along a route. What if the route itself is in superposition — the message passing through two relay stations at once?
What happenedThe superposed route carries a resource that neither a fixed path nor a coin-flip between paths can carry, and the measurement that proves it fires far above both of those controls. Then the same superposed route was put behind the error-correcting shield and survived, at 96% of its unshielded strength.
What it meansIndefinite order is not only something you can do to operations inside one chip — it is something a network can have, and it is compatible with error correction rather than destroyed by it.
What we can build with itRouting as a quantum resource: network topologies that carry more than any single definite path could.
One Dial for the Arrow of Time — information and energy on one knob
CERTIFIED
Exp227ibm_fez
The ideaThere is one knob: how much the environment recorded of what happened to a qubit. Turn it up and the qubit stops behaving like a wave and the event becomes a definite fact. Do other quantities ride the same knob?
What happenedThree of them do, each following a law written down before the flight, measured across five settings on one apparatus: how wave-like the system still is, how objective the fact has become, and how strongly the two are linked.
What it meansQuantities that look like different physics — a wave's visibility, a fact's objectivity, the link between a system and its surroundings — turn out to be one number read three ways. That single number is what "the past became definite" actually means here.
What we can build with itA measurable handle on decoherence: how far a computation has slid toward classical, as one dial reading.
Reported, not certifiedThe same apparatus also shows a fourth quantity on the knob — the energy one classical bit can unlock in the recording environment, which rises as the record completes. That measurement was deliberately kept outside the pass/fail test and is reported as-is, so it is not a certified result. And it is not free energy: the measurement pays for and injects the energy; the bit only tells the far end which local operation unlocks its share. Nothing is transported.
The past, reopened — crossing an old wall a second way
CERTIFIED
Exp229 + Exp230ibm_fez
The ideaEarlier programmes showed a qubit has no definite value between looks, and that erasing a which-path record can bring back an interference pattern. Both were revisited — one to cross the same wall by a cheaper route, one to ask a sharper question.
What happenedThe "no definite value" wall was crossed again by a much simpler circuit, independently. And the eraser question got sharper: it is not only whether the pattern returns — the eraser's own outcome selects which of two opposite patterns was there, while the unsorted data stays flat the whole time.
What it meansA later choice can sort an already-recorded past into one of two mutually exclusive stories. It cannot send anything backwards — the flat unsorted data is the proof — but which story you can tell is not settled until you look.
What we can build with itSharper tests of what "a measurement happened" means, on hardware rather than on paper.
What this is NOTThe simpler re-crossing does not replace the original: it uses a plainer kind of measurement that leaves open the standard objection ("your measurement disturbed it"), which the original closed. Both results stand, and the museum page says which is the stronger protocol.
HORIZONS 6 The living ship from surviving errors to healing them · and the one gate that makes a coded computer universal

Every shield up to this point could only notice that something had gone wrong and throw the run away. That is survivable for a short calculation and hopeless for a long one — discard often enough and you never finish. This programme crosses the line that actually matters: a code that works out which qubit broke, fixes it, and keeps going. And then it reaches for the gate that decides whether a protected machine can compute anything interesting at all.

The self-healing qubit — correction, not discarding
CERTIFIED
Exp236 · 238 · 240 · 241ibm_fez
The ideaDetecting an error means "a typo happened, delete the page". Correcting it means "fix the typo and keep typing". Only the second one lets a long computation finish, and it needs the machine to identify which qubit broke without looking at the data it holds — because looking would destroy it.
What happenedIt works, live, mid-circuit. Helper qubits point at the culprit, a corrective pulse fires while the computation is still running, and the run survives. Run the correction loop repeatedly and the advantage compounds: the corrected qubit ages slowly while an identical circuit with the fix switched off falls apart.
What it meansThis is the mechanism every scalable quantum computer is built on. Not a faster answer — a machine that can still be computing after the point where an unprotected one has become noise.
What we can build with itComputations longer than the hardware's own patience.
What this is NOTThe errors are deliberately injected, not ambient — these are qubits broken on purpose, in a known way, to test whether the code puts them back. Small codes, and this demonstrates the mechanism of fault tolerance rather than a below-threshold machine. Nothing here outruns a laptop.
The phase dual — it worked, and the verdict is still "not held"
SPLIT / KEPT
Exp237ibm_fez
The ideaA qubit can break in a second, subtler way — losing its phase rather than flipping its value. That needs its own code, the mirror image of the first one.
What happenedThe code plainly works: given a deliberate phase flip, it brings the logical value back where an unprotected qubit is not merely degraded but gone — flipped to the opposite state. And the flight is recorded as NOT HELD, because one of eight test cells landed one thousandth below the pass mark that had been fixed in advance.
What it meansBoth of those are true at once, and the second is the point. A bar you set beforehand is only worth having if it can be missed by a thousandth — and then not moved. There was no second attempt and no adjustment to the mark. Shopping for a threshold your result already clears is the failure the whole practice exists to prevent.
What we can build with itTogether with the bit-flip code above, coverage of both ways a qubit breaks.
What this is NOTThis code guards the phase channel only — a value flip goes straight through it, exactly as the other code is blind to phase. Neither half closes alone; the pair is what closes.
Where the stair runs out — two corrected qubits, one gate
SPLIT / KEPTWALL
Exp242ibm_fez
The ideaThe frontier: not one protected qubit but two, with a gate acting between them while both are being corrected live.
What happenedHalf of it held. Correction measurably helps the two-qubit gate compared with the identical circuit with the fix switched off. The other half missed: a protected entangled pair needs both of its halves intact, and one collapsed entirely — because this code cannot see phase errors, and the circuit was deep enough to erase that half. That failure was named as the risk before the flight, and it landed exactly there.
What it meansThe stair does not stop here because something broke unexpectedly. It stops because the thing everyone expected to break, broke, on schedule. The next rung needs a code that protects both channels at once.
What we can build with itA specific, measured requirement for the next generation of codes rather than a guess.
Read the certified claim preciselyWhat was tested is a difference — corrected must beat the fix-switched-off control by a set margin, and it does. That says nothing about absolute quality, and the absolute numbers are modest: the gate is right about six times in ten, and its worst case is only a little better than guessing — and on a two-bit outcome, guessing is right just one time in four. Correction measurably helps a two-qubit logical gate, at fidelities still well short of useful.
Magic, delivered — the gate that makes a coded computer
CERTIFIED
Exp243ibm_fez
The ideaProtecting a qubit and computing with it are different problems. The operations a code can perform safely are a restricted set, and a machine limited to them can be simulated by an ordinary laptop. Getting past that needs one more kind of gate — and applying it directly is provably incompatible with the protection. So it is not applied directly: a separately prepared state is consumed, and its gate teleported onto the data.
What happenedThe gadget runs. The giveaway is which value arrives: an ordinary protected state is only permitted to read certain values on this measurement, and what landed is none of them — it sits in the gap between the allowed ones, which is what makes it the useful kind of state. Remove the single gate linking the prepared state to the data and nothing arrives at all.
What it meansThis is the route every scalable universal quantum computer is expected to take for these gates. Detection and correction buy fidelity; this buys universality — the difference between a protected memory and a protected computer.
What we can build with itThe missing piece between "a coded qubit survives" and "a coded qubit computes something a classical machine cannot follow".
What this is NOTError detection, not correction: it discards the runs that fail its check, which is a real cost. The prepared state is made directly here as a stand-in for the factory a real machine would use — what is demonstrated is the delivery, downstream of wherever the state came from. And this certifies the mechanism, not a fault-tolerant quality of result.

So what does it all add up to?

Four programs, one throughline: take a piece of exotic physics, build it as a circuit, freeze the pass/fail rules before looking, and measure. Here's what we actually learned — and what it's good for.

01Cause and effect are negotiable

A quantum computer can hold the order of events in superposition — and you can teleport it, network it, and run an engine on it. Order is a resource, not a fixed backdrop.

02A fact is a receipt

Objectivity and irreversibility are the same bookkeeping: a fact is a record the world still holds. Take the record back and the past reopens. That's why the everyday world — awash in records — feels solid and one-way.

03Famous paradoxes, measured

Negative energy, the twin paradox, the grandfather paradox, black-hole information, "a watched pot" — all built as chip analogues and measured with tens-to-hundreds-of-sigma confidence, not just argued.

04Quantum wins specific games — and we proved the limits too

Contextuality (196σ), dense coding (341σ), metrology (168σ), random access, certified randomness. And the flip side: the no-cloning limit on quantum itself, certified with teeth.

05Error correction pays — for some computations, and we measured which

The [[4,2,2]] shields beat their raw counterparts on keys, sensors, and — the first time — computation, even carrying 3.5× the overhead, with the gap growing as the circuit deepens. But that is not a general claim, and later flights bounded it. Whether the shield wins depends on what the logical gates cost: where the code makes them cheap, it pulls ahead early and keeps pulling; where they cost real gates, it never catches up and the gap widens the other way. The boundary is measured from both sides in the museum, one curve rising and one falling — so the claim this card can carry is "it pays here, and here is where it stops."

06The method is the real result

Pre-registration, executed nulls, and keeping every miss and retraction in the record. Two "discoveries" were withdrawn by their own author the same day. The kept record is what makes the wins trustworthy.

What we are NOT claiming

These are demonstrations of the quantum nature of order, facts, and error correction — not a general-purpose quantum computer beating your laptop. The one computational advantage here (the shallow solver) is a fidelity over a random baseline; its asymptotic power is carried by a theorem, not shown on-chip. The paradoxes are faithful analogues (a model of a time loop, not a real one). And cross-generation portability is an open question, because the hardware to test it isn't available to us. Every one of these boundaries is written down, not glossed over.

The point of building the garden simply: you can see exactly where the walls are.