Decisions No Correction of the Uncontrolled Circuit Can Reach: Real-Time Control on IBM Heron, Declared Before the Data
Year: 2026
Type: preprint
Abstract: A circuit that measures in the middle and acts on the result at once has a natural comparison: the same circuit with the actions removed. Whatever is done to that uncontrolled circuit after the run, including error mitigation at any shot cost, returns at best its own noiseless value; the distance from that value to the exact answer is a correction ceiling c. We ask whether real-time control on today's cloud devices gets closer to the exact answer than c, with every shot counted, no mitigation, and every test declared in git before the device record of its run was fetched. Over four nights on IBM Heron devices (57 declared predictions: of the 49 of 11-16 to 11-18, 33 held and 16 failed; of the 8 of 11-19, 4 held and 4 failed; every failed one is reported as failed), chains of repeat-until-success (RUS) neurons crossed the ceiling and the margin grew with the chain length (slope +0.024 per neuron on ibm_fez; +0.031 and, after a recalibration, +0.053 for a frame-balanced pair on ibm_marrakesh); iterative phase estimation crossed its ceiling at n = 2, 3 on ibm_fez and n = 4 on ibm_marrakesh (0.903 against 0.471 at n = 4). Getting there needed a diagnosis on the device: the first neuron design lost its value not to the conditionals, which cost nothing measurable when they did not fire, but to time spent in |1> during mid-circuit measurement waits; holding the likely value as |0> recovered 0.11-0.22. That one-sided decay can also push a design above its own noiseless value; a guard declared beforehand took those cases out, and a frame-balanced pair removes the lean by construction. Two limits are measured: fixed circuits can share a circuit with no loss (halving device time), but two items with conditionals in one circuit cost the control arm (0.20 on average with ten items; 0.06 on the first of two), which a delay did not remove and which is not explained by decisions waiting for each other (no decision latency above about 0.4 us was resolved). Finally, a prescription loop that diagnoses the device, applies a fixed rule and runs the chosen circuit ran unattended on the device and picked a different prescription after the device was recalibrated. On a fourth night, the loss between two conditional items followed the register's bit order, not the qubits: the item whose bits came first lost (-0.037, 2 SE 0.020), as on 11-18. Two more tests inside one item cost nothing measurable when they did not fire (repetition code), though the code decoded in real time stayed below its ceiling (0.812 against 0.875). Phase estimation crossed its ceiling again at n = 2, 3, 4 on one device and night, and, as a separate yardstick, was right per shot more often than the textbook circuit by +0.03, +0.07, +0.17. The loop, run from the package, refused the previous night's ledger after a recalibration, stopped spending where its prescription never helped, added one try where it did, and reached a fixed point. What the device keeps of the noiseless value varies by night (0.61-0.87 for the same balanced pair on two nights); no quantum advantage is claimed.
Author: TN.Holdings
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236797960
