Precursor Detection of Systematic Bias Drift in Cryogenic Quantum Control: A Fail-Closed Estimator Validated in Simulation and on Live Hardware Calibration
Cryogenic control electronics for quantum hardware accumulate systematic, non-zero-mean bias that integrates in a feedback loop until the actuator saturates; a threshold controller detects this only after saturation. We present a precursor estimator that detects the onset of systematic drift from the dynamics of the control signal, built as a finite observer (declared error budget, explicit competence boundary, fail-closed abstention) in the sense of a companion paper on the physical limits of finite observation. It uses path-functional invariants (a Fisher–Rao metric stability bound, a maximum-likelihood bias decoder, a cumulative-damage functional) rather than eigenvalue analysis, which we show is neither necessary nor sufficient for finite-time survival. In simulation, across 30 seeds × 7 sensor-environment conditions (including a 1/f dephasing spectrum), it detects 96.7% of drift-collapse scenarios before saturation at a 12 s median lead with 0% false positives over 210 healthy runs. On real data, the same governed outlier-and-drift detector, applied to live publicly accessible IBM quantum-hardware calibration time-series (a 156-qubit IBM Heron device), flags genuinely degraded qubits and catches a real relaxation-time decline (one qubit's T1 falling 88.0→19.1 µs, −78%, over 43 days) while staying silent on a qubit that is poor but stable and on the other 155 — detection of change, not of mere badness. The work is TRL 4: an advisory detection layer, not a deployed controller, and it does not emulate entanglement, non-local correlation, or contextuality. The contribution is the integrated, calibrated estimator and its operating discipline, demonstrated in simulation and against real hardware-calibration data.
Cryogenic control electronics for quantum hardware accumulate systematic, non-zero-mean bias that integrates in a feedback loop until the actuator saturates; a threshold controller detects this only after saturation. We present a precursor estimator that detects the onset of systematic drift from the dynamics of the control signal, built as a finite observer (declared error budget, explicit competence boundary, fail-closed abstention) in the sense of a companion paper on the physical limits of finite observation. It uses path-functional invariants (a Fisher–Rao metric stability bound, a maximum-likelihood bias decoder, a cumulative-damage functional) rather than eigenvalue analysis, which we show is neither necessary nor sufficient for finite-time survival. In simulation, across 30 seeds × 7 sensor-environment conditions (including a 1/f dephasing spectrum), it detects 96.7% of drift-collapse scenarios before saturation at a 12 s median lead with 0% false positives over 210 healthy runs. On real data, the same governed outlier-and-drift detector, applied to live publicly accessible IBM quantum-hardware calibration time-series (a 156-qubit IBM Heron device), flags genuinely degraded qubits and catches a real relaxation-time decline (one qubit's T1 falling 88.0→19.1 µs, −78%, over 43 days) while staying silent on a qubit that is poor but stable and on the other 155 — detection of change, not of mere badness. The work is TRL 4: an advisory detection layer, not a deployed controller, and it does not emulate entanglement, non-local correlation, or contextuality. The contribution is the integrated, calibrated estimator and its operating discipline, demonstrated in simulation and against real hardware-calibration data.
Cryogenic control electronics for quantum hardware accumulate systematic, non-zero-mean bias that integrates in feedback loops until actuator saturation. We present a precursor estimator that detects the onset of systematic drift from control-signal dynamics before saturation. Designed as a finite observer, it operates to a declared error budget, carries an explicit boundary of competence, and abstains—fails closed—outside it. The estimator uses path-functional invariants rather than eigenvalue or linearisation analysis. In simulation across 30 seeds and seven sensor-environment conditions, it detects 96.7% of drift-collapse scenarios before saturation, with a 12-second median lead and 0% false positives across 210 healthy runs. Applied to live, publicly accessible IBM quantum-hardware calibration time-series from a 156-qubit IBM Heron device, it flags degraded qubits and detects a real relaxation-time decline while raising no alarm on a poor-but-stable qubit or the remaining 155 qubits. This is a TRL-4 detection layer: advisory, validated in simulation and on hardware calibration data, and not a deployed controller.
Cryogenic control electronics accumulate systematic, non-zero-mean bias — from thermal cycling of mounts, flux leakage, and dielectric memory — that a feedback loop integrates until its actuator authority is exhausted and the integrator saturates. A controller that alarms on amplitude thresholds sees this failure only after correction is no longer available. This paper reports a precursor estimator that reads the dynamics of the control signal for the onset of systematic drift, built deliberately as a finite observer in the sense of the first Finite Cognition paper: a declared error budget, an explicit competence boundary, and fail-closed abstention. Its evidence is presented in two declared strata. First, a fresh, fully archived Monte-Carlo run of record on a declared control-loop model: across 30 seeds × 7 sensor-environment conditions × two drift polarities, with every threshold frozen before the batch, the estimator detected drift before saturation in 246/354 eligible collapse runs (69.5%, 95% interval 64.5–74.1%), with per-condition detection ranging from 25.5% (all disturbances combined) to 100% (dropout; 92.4–100), median leads of 7–17 s, and zero false interdicts across 210 healthy runs within the declared calibration window (0.0–1.8%) — a bound the long-horizon batch shows is conditional on the mandated recalibration schedule. Second, an archived read of live IBM quantum-hardware calibration telemetry (a 156-qubit Heron device, June 2026): the same governed detector family flagged the two genuinely degraded qubits and the sole sustained relaxation-time decline within that window (T_1 falling 88.0 → 19.1 µs across 43 returned calibration snapshots spanning 42 calendar days) — an event a longer archived record later resolves into the first, longest and deepest of four recurring excursions on the same qubit, at a cost to this detector family the paper states while staying silent on a bad-but-stable qubit and the 154 others. The estimator holds advisory-only authority under the programme's interface contract: it emits verdicts and writes nothing to the plant. This record supersedes an earlier version whose simulation table could not be traced to a quantum-specific run; the run of record reported here, code and data archived, is the evidence this paper stands on.
A companion paper reported a fail-closed estimator — the Quantum Control Kernel (QCK) — that detects the precursor of systematic bias drift in cryogenic quantum control before actuator saturation, validated in simulation and against live, publicly accessible IBM hardware-calibration data. This paper reports what the estimator is for: the Quantum Control Architecture (QCA), a governed measurement-and-control layer in which the demonstrated kernel is the sensing floor of a larger construction. The construction is not free design. It is the quantum instantiation of a result established generically in the Finite Cognition series: a finite observer serving both a committing execution demand and a wide-field vigilance demand is forced to partition the two and re-couple them at the point of finalisation, so that no output is committed that vigilance has not verified. We show that quantum control information forces exactly this construction — telemetry is consumed by every layer of the control stack yet governed by none of them; observation of a quantum plant must be non-perturbative by discipline, not by hope; and an estimator that is fluent but ungoverned reproduces, on the control substrate, the same confident-and-blind failure the series identified generically. We then state the architecture that discharges the constraint, in its public units: a non-perturbative observer (NPO); the QCK detection floor; fail-closed gates that allow, abstain, or refuse; a state-coherence and equilibrium discipline over commits; provenance-checked memory that refuses known-failure patterns; a lifecycle and loop supervisor; a measurement plane of stability, coupling, margin, and cost features; and an integrity plane that audits the architecture’s own calibration model and gates its portability to new hardware on provably leak-free relocation. We are explicit about maturity throughout: the kernel is demonstrated at Technology Readiness Level 4 under published bounds; the architecture beyond it is compile-ready geometry — derived, specified, and internally consistent, not validated as live middleware on partner hardware. The architecture is advisory by construction: the operator owns the plant at every depth.
A companion paper reported a fail-closed estimator — the Quantum Control Kernel (QCK) — that detects the precursor of systematic bias drift in cryogenic quantum control before actuator saturation, validated in simulation and against live, publicly accessible IBM hardware-calibration data. This paper reports what the estimator is for: the Quantum Control Architecture (QCA), a governed measurement-and-control layer in which the demonstrated kernel is the sensing floor of a larger construction. The construction is not free design. It is the quantum instantiation of a result established generically in the Finite Cognition series: a finite observer serving both a committing execution demand and a wide-field vigilance demand is forced to partition the two and re-couple them at the point of finalisation, so that no output is committed that vigilance has not verified. We show that quantum control information forces exactly this construction — telemetry is consumed by every layer of the control stack yet governed by none of them; observation of a quantum plant must be non-perturbative by discipline, not by hope; and an estimator that is fluent but ungoverned reproduces, on the control substrate, the same confident-and-blind failure the series identified generically. We then state the architecture that discharges the constraint, in its public units: a non-perturbative observer (NPO); the QCK detection floor; fail-closed gates that allow, abstain, or refuse; a state-coherence and equilibrium discipline over commits; provenance-checked memory that refuses known-failure patterns; a lifecycle and loop supervisor; a measurement plane of stability, coupling, margin, and cost features; and an integrity plane that audits the architecture’s own calibration model and gates its portability to new hardware on provably leak-free relocation. We are explicit about maturity throughout: the kernel is demonstrated at Technology Readiness Level 4 under published bounds; the architecture beyond it is compile-ready geometry — derived, specified, and internally consistent, not validated as live middleware on partner hardware. The architecture is advisory by construction: the operator owns the plant at every depth.
In realistic hardware for quantum computation that possesses fault-tolerance, non-stationary noise and stochastic drift lead to logical failure from the temporal accumulation of errors, not from independent events. Static decoding and fixed calibration techniques are structurally incompatible with this situation because they do not take into account temporal correlations between errors or control-induced back-action of errors. These effects motivate control policies that must track noise evolution across correction cycles, rather than respond to individual syndromes in isolation.
We treat fault-tolerant quantum computation as a stochastic control problem, modelled using reduced quantum dynamics in which Pauli error processes are governed by latent noise parameters that vary temporally. From this perspective, logical failure arises through the accumulation of a hazard variable, and the corresponding control objective depends on the full history of observations.
Operating under these conditions, a Chronological Deep Q-Network (Ch-DQN) maintains an internal belief state that tracks both noise evolution and accumulated hazard. During training, backward refinement of trajectories is used to sample slowly drifting modes of operation, while runtime inference remains strictly causal. A fractional meta-update stabilizes learning in the presence of non-stationary, control-coupled dynamics.
Through multi-distance simulations that incorporate stochastic drift and feedback from decision-making, Ch-DQN suppresses hazard accumulation and extends logical survival time relative to static and recurrent baselines. Error correction in this regime is therefore no longer a static decoding task, but a control process whose success is determined over time by the underlying noise dynamics.
We report systematic experimental observation of coherent, topology-dependent inter-horizon correlations in thermofield double (TFD) states implemented on IBM quantum processors. Using a 100-qubit ladder geometry mapped onto the heavy-hex topology of ibm_fez (156 qubits, Heron r2 architecture), we demonstrate signal ratios R_fit spanning 1.2–202.7 across runs/seeds, with best single-seed R_fit = 23.93 (Run #11) and record multi-seed median R_fit = 102.7 (Run #2, 5 seeds, range 63.5–202.7), validated by pre-registered GO/NO-GO criteria across 14 experimental runs spanning January 24-30, 2026. Our experimental framework incorporates two rigorous anti-artifact controls (see Fig. 6-7): (1) SCRAMBLE control applies identical circuit depth with incorrect topology, destroying the signal (reduction factor > 10×); (2) ECHO control (Loschmidt reversal) applies the coupling layer twice, annihilating coherent contributions (reduction factor 8.6-50.5×, consistent with expected finite-size echo saturation for N~50 effective subsystems (see Gorin et al. [14] for saturation plateau theory)). Both controls maintain > 95% local fringe retention, confirming they do not trivially destroy coherence. A critical discovery emerged from our January 30, 2026 validation campaign (see Fig. 5): quantum processor calibration drift can cause signal collapse within less than one hour on hardware with identical offline calibration snapshots. Runs #11-#12 achieved R_fit = 23.93 and 18.4 respectively (GO-1 PASS, GO-3 PASS), while Run #13 executed only 53 minutes later on the same backend collapsed to R_fit = 1.20 (GO-1 FAIL). This finding establishes that offline Health Scores do not reliably predict experimental success, and defines an operational envelope requiring Health > 80% (Grade B or better) for empirically associated with high-probability signal stability. The observed second-harmonic (2ω) dominance in the phase-dependent signal (A₂ω/A₁ω > 40×, R² > 0.99, see Fig. 9) is consistent with theoretical predictions for bilinear left-right correlations. A stride sweep (s = 2-5, see Fig. 4) reveals a signal-coherence trade-off, with s = 3 identified as optimal. Total quantum processor time across all 14 runs: approximately 45 minutes. These results establish a validated, falsifiable experimental framework for probing analog quantum gravity signatures on NISQ hardware, with explicit calibration-conditional bounds. The methodology may provide a validated framework applicable to analog Hawking radiation platforms where topology-controlled correlations play a central role (see Section 5.4). Note: Results demonstrated primarily on IBM ibm_fez; multi-device validation on additional backends is recommended to establish cross-platform generalizability.