Alien Math · on a real quantum chip

Seven jobs on a real quantum chip

Teleportation, a time-reversal echo, a bomb found without touching it, a spin that reads out bigger than it is, three pigeons in two boxes, and a qubit that refuses to have been anywhere in particular. Each one ran on IBM's ibm_fez with its pass and fail lines locked before the job was sent. This page puts all of it in one place, with what bent and who paid.

ibm_fez · IBM Heron · 156 qubits

7jobs, October 24 and 28, 2026

27 / 35locked claims held; every miss is on this page

0.158pigeonhole pair sum on fresh qubits; classical needs at least 1

35σLeggett–Garg violation, K₃ = 1.49

579 sof the month's 600 free seconds of chip time

The scorecard

Every claim, every job

One tile per pass or fail line, written before the job ran. Hover or tap a tile for the claim. The orange tile is a null result the chip rejected, which turned into the echo's headline.

Loading the runs…

heldmissednull rejected

The protocol

Locked first, graded after

Write the linesEach claim gets a pass or fail rule and a number from IBM's noise model of the chip.
Lock themThe prediction file is hashed and committed before submission. The result file records the same hash.
Submit onceEach job is sent one time. A repeat is a new job with its own locked lines.
Grade, keep the missesHeld or missed, by the locked rule. Readings made after the run are labelled as such.

Job 1 · PG-021 · teleportation

A qubit moved, and two bits paid for it

Where it bends

A qubit's state leaves one qubit and turns up on another with no quantum channel between them at that moment.

Who pays

Two ordinary bits must travel first. Without them Bob holds noise, 0.496 in simulation. One entangled pair is used up, and the original is destroyed.

Fidelity with mid-circuit measurement (dynamic) and with quantum-controlled corrections (deferred), each beside IBM's noise-model prediction. Both clear the 2/3 line. Deferred came in 0.104 below dynamic, past the 0.08 allowed, so the gap clause missed; the noise model had the two within 0.001 of each other.

Full write-up in the analog spacetime room →

Jobs 2 and 3 · PG-020 · the quantum echo

Run it backwards and the chip shows its hidden phase

Where it bends

Scramble 8 qubits with up to 84 two-qubit gates, run the circuit backwards, and they come home. Add a small deliberate phase to the reversal and, if the chip has a hidden phase error of its own, a "wrong" reversal beats the perfect one.

Who pays

The light cone held: a flipped qubit only reaches the qubits it could have. The chip's own coherent error showed up instead, about 4 milliradians per layer, and it doesn't scale the way one uniform phase would.

The best reversal phase per layer on hardware, against the noise-free model. The first job rejected its own "nothing hidden" null at depth 6 (z = 4.1). The confirmation job, with the layout pinned and the circuits shuffled and run twice, found it again at about 6σ (C1). Depth 8 came out nearly twice depth 6, so C2, one uniform phase per layer, missed.

Full write-up in the analog spacetime room →

Job 4 · QW round 1 · the Zeno bomb tester

A bomb found without touching it, 76.7% of the time

Where it bends

Rotate a qubit a little at a time and check it after every step. If a "bomb" would absorb it, the checking freezes the rotation, and the qubit reports the bomb without ever reaching it.

Who pays

Loss. Every cycle leaks about 0.7%, nearly the same at every N, so more cycles stop helping past 16. A qubit with a mid-circuit measurement behaves like a leaky optical loop.

Bomb found without touching it, against cycles N. The hardware tracks the lossy-loop model closely and peaks at 76.7% at N = 16, against 77.2% from the noise model and 25% for the textbook single pass. With no bomb, the tester answered correctly 99.2 to 99.5% of the time.

Full write-up in the Quantum Wizard room → · Animated, as an As Seen on TV episode →

Jobs 4 to 7 · weak values and the floor

A spin that reads bigger than it is, until the floor

Where it bends

Couple a spin weakly to a pointer and keep only the rare runs where it ends up nearly opposite. The pointer in those runs moves several times further than a spin-1/2 should allow.

Who pays

Postselection keeps fewer and fewer runs, and on this chip about 3% of runs always sneak through by error. That floor caps amplification near 1.5× on a bad day and about 2.8× on a good one.

Measured amplification against the target. Ideal grows with the target; the hardware turns over. Round 1's −2.66 at a target of 10 missed the noise model (W1), and the same circuit gave −1.66 four days later (B2 missed). One day's slope describes that day's calibration.

Taking the floor apart

Round 2 rebuilt the circuit one step at a time. Readout and single-qubit turns add almost nothing. The two-qubit gate step adds 3.2 points at about 20σ, where the noise model expected under 0.4 and had blamed readout. No calibrated rotation removed it (B3 missed).

Round 3 tested the same step on four pairs across the chip. Every pair came in at 6 to 20 times the gate error IBM reports for it (F1 held), and round 2's pair repeated its own number (F3). The floor doesn't add up gate by gate, and on one pair doubling the gates lowered it (F2 missed), so part of the error is coherent across both qubits.

Full write-up in the Quantum Wizard room →

Jobs 6 and 7 · the quantum pigeonhole

Three pigeons, two boxes, no two together

Where it bends

Three qubit "pigeons" go into two boxes. Some pair must share a box, so the three chances of sharing must add up to at least 1. Among the runs that pass a final check, they added up to 0.16, twice, on two different sets of qubits.

Who pays

The effect lives only in the postselected runs. Without the final filter, every pair shares a box half the time, exactly like coins (the controls).

The three pair probabilities stacked end to end. The shaded band is where no classical assignment can land. Round 2 (qubits 144, 142, 136) summed to 0.161 and round 3 (qubits 24, 37, 26) to 0.158; the shown pairs are rounded to three places. The controls sum to about 1.5. All six pigeonhole claims held.

Full write-up in the Quantum Wizard room →

Job 7 · QW round 3 · Leggett–Garg

A qubit that was never anywhere in particular

Where it bends

If a qubit always had a definite value that could be checked without disturbing it, a combination of three correlations called K₃ could never pass 1. Qubit 54 scored 1.490 ± 0.014, 35 standard errors over.

Who pays

Time. The violation was gone by 11.4 µs of idling, where the chip's own T1 and T2 predict 76.6. K₃ dives, recovers and dives again, which is the signature of a qubit drifting off-frequency while it waits.

K₃ against idle time. The shaded band is where only quantum behaviour can reach. The slate line is the prediction from the chip's reported T1 and T2 (LG2 missed). Drift of that kind can in principle be undone with a refocusing pulse mid-idle, which separates it from real decoherence; that is the next run.

Full write-up in the Quantum Wizard room →

The bill, itemised

Three costs this chip charges

~0.7%

Loss per cycle

A mid-circuit measurement leaks about the same amount every time. It sets the Zeno tester's ceiling at 77%, and one number predicts the whole curve.

6–20×

The two-qubit floor

Postselection lets about 3% of runs through by error, from the two-qubit gates, far above their reported error. It caps weak-value amplification.

6.7×

Clock drift at idle

An idling qubit drifts off-frequency and gives up its quantum margin 6.7 times sooner than T1 and T2 predict. The echo's hidden phase is the same kind of coherent error.

Misses

Eight lines that didn't hold

Teleportation gap. Deferred corrections came in 0.104 below dynamic, past the 0.08 allowed.
Echo, uniform phase (C2). Depth 8 doesn't fit one phase per layer.
Weak values, round 1 (W1). Amplification at a target of 10 fell short of the noise model.
Replication (B2). The same circuit four days later gave −1.66 against −2.66.
Calibration (B3). A tuned postselection pulse didn't lower the floor.
Leggett–Garg lifetime (LG2). 11.4 µs against 76.6 predicted.
Floor scaling (F2). The floor doesn't add up gate by gate.
A null that failed. The echo's "nothing hidden" null was rejected; that one became the finding.

Two leanings written down ahead of time also went the other way: round 2 expected the floor to be coherent and fixable by a rotation, and round 3 had the right direction for the Leggett–Garg collapse for a partly wrong reason.

Prior art. Teleportation, mirror-circuit echoes (Google's 2025 Quantum Echoes among them), Leggett–Garg violations and the quantum pigeonhole effect are all known, and most have been shown on IBM hardware before. What these runs add is the locked-prediction record and where the costs land on one chip on two days in October 2026.

Ledger

All seven jobs

date	run	headline	claims	qubits	chip time	job

Every number on this page is drawn from runs.json. Prediction hashes and commits are listed in each room's write-up. Chip time was not logged for the three October 24 analog-lab jobs.

Back to the Alien Math wing, or the rooms: analog spacetime and Quantum Wizard.