🐺 i10 jackal: a calibration policy for a realistic device¶
🟢 active since 2026-09-26 · sim-to-sim gap measured (35×); trained on the realistic device 7.2e-3, closing the 5× margin to gradient-optimised knobs
Runs in runs/i10_jackal/, records in results/i10_jackal/.
Question. Does ibis's open-loop carrier policy, which needs only a routine frequency calibration, survive a realistic device model? And if it is trained on the simplified model one would write down first, how much does it lose on the realistic one (the sim-to-sim gap)?
The device model¶
rlquantopt/jx/device.py, rlquantopt/jx/env_device.py (--physics device).
| Element | Model | Source |
|---|---|---|
| Circuit | Two fixed-frequency transmons and a flux-tunable transmon coupler, 3 levels each | Sung et al., PRX 11, 021058 (2021) |
| Parameters (idle) | Qubits 4.16 / 4.00 GHz, anharmonicities −220 / −210 MHz; coupler 5.45 GHz (tunable 3.7-6.7 GHz), −90 MHz; g1c = 72.5 MHz, g2c = 71.5 MHz, g12 = 5.0 MHz | Sung et al. 2021, Table 1 |
| Coupler flux curve | Asymmetric SQUID, ωc(Φ) = ωmax [cos²(πΦ) + d² sin²(πΦ)]1/4, with ωmax = 6.7 GHz and d = (3.7 / 6.7)² from the quoted range (our assumption: the range is the full swing); couplings scale as √(ωc / ωc,idle) | Koch et al. 2007 |
| No rotating-wave approximation | Counter-rotating terms kept; parity is conserved, so the Hamiltonian splits into two blocks (7 and 10 states, up to 3 excitations) | Checked against the full 27-level space: fidelity within 1e-5 |
| Computational states | Dressed idle eigenstates; gate in the frame of the idle Hamiltonian; free virtual Z | |
| Control | Coupler flux offset from idle (±0.15 Φ0; knob steps up to 0.03 Φ0, 0.3 rad, 0.02 Φ0 per decision), a carrier at the measured dressed qubit detuning (159.5 MHz) steered by amplitude, phase and offset; 1 ns AWG hold; first-order flux-line filter, 0.5 ns | AWG 1 ns as in Braket's Rigetti example; filter constant assumed |
| Drift | Qubits ±5.7 MHz (recool, Zhang et al. 2022); coupler flux offset ±20 mΦ0 (Dai et al. 2021: the worst loop after 17 days; here about −178 MHz of coupler frequency); optionally couplings ±5.7 MHz and anharmonicities ±5 MHz, which the calibration does not see | |
| Calibration (the policy's only input) | Dressed qubit frequencies and coupler idle frequency from spectroscopy, errors 0.1 / 0.1 / 1 MHz | |
| Not yet included | Decoherence (T1 60 / 30 / 10 µs in Sung et al.), spectator qubits, measured flux-line distortions |
A first check. A parametric √iSWAP is reachable in this device: five constant carrier knobs optimised by gradient give 1 − F = 7.8e-4 in 60 ns (Sung et al.'s resonant iSWAP takes 30 ns).
The simplified model (--device-simplified) is what one would write down first: rotating-wave
approximation, no direct coupling, coupler frequency linear in flux around idle, couplings fixed, ideal
AWG and flux line. Same parameters, same calibration and actions, so a policy trained on it can be
deployed on the full model unchanged.
Experiments¶
All: carrier actions, calibration only, 64×64 GELU, clip λ = 1, 100 ns episodes of 20 steps of 5 ns,
3M env steps, qubits ±5.7 MHz and coupler flux ±20 mΦ0. Evaluation: scripts/jackal_eval.py,
the full model, 24 drifted devices, the full pulse.
| Run | Trained on | Unmeasured drift in training |
|---|---|---|
| A | simplified model | no |
| B | full model | no |
| C | full model | couplings ±5.7 MHz, anharmonicities ±5 MHz |
Results so far¶
Each policy's best checkpoint (by the full-pulse 1 − F on its own training model's nominal device), deployed on the full model: 24 drifted devices, 3 calibration-noise draws each.
| Run | Its own model, nominal | Full model, drift the calibration sees: median (below 1e-2) | Plus unseen coupling / anharmonicity drift | Env steps to 1e-2 (own model) |
|---|---|---|---|---|
| A, trained on the simplified model | 1.1e-3 | 3.7e-2 (0 %) | 3.9e-2 (0 %) | 0.15M |
| B, trained on the full model | 6.9e-3 | 7.2e-3 (75 %) | 7.9e-3 (69 %) | 1.3M |
| C, full model + unseen drift in training | 6.3e-3 | 8.1e-3 (69 %) | 1.0e-2 (49 %) | 1.3M |
A calibration error three times larger than in training changes these by less than 10 %.
- The sim-to-sim gap is large. A policy that reaches 1.1e-3 on the simplified model (the model one would write down first) is 35× worse on the realistic one. The rotating-wave approximation, the linearised flux curve, the direct coupling and the flux line all matter at this level, so a policy must be trained on the most faithful model available (and, on hardware, checked against it).
- Training on the realistic model works, with a 5× margin still open. B gives 7.2e-3 on drifted devices; gradient optimisation of the same 20-step knob schedule reaches about 1.3e-3 with the same calibration (an exploratory check by the hippogriff session), so the policy is not yet at the limit of its action space.
- Unseen drift costs little here, and training on it did not help at this budget (C).
An evaluation bug, fixed
Until the fix in this version, the periodic evaluation during device training ran on the paper
model's qubit frequencies (5.03 / 5.89 GHz) instead of the device's (4.16 / 4.00 GHz): reset_to
received them from the agents' evaluate(). The training itself was unaffected, but the logged
evaluation and the choice of the best checkpoint were wrong. scripts/jackal_eval.py now re-scores the
saved checkpoints on the device's nominal point, and reset_to uses it.
Next¶
- Close the 5× gap on the realistic model: longer training, more seeds, and the knob scale (a starting amplitude near 0.06 Φ0 worked for gradient optimisation).
- Decoherence (T1 of 60 / 30 / 10 µs): a gate-time penalty.
- A hardware contract (sampling, bandwidth, amplitude range) for a platform with pulse access.
Tests: tests/test_jx_device.py (flux curve, truncation against the 27-level space, idle gate,
parametric exchange, calibration and drift in the environment, filter, simplified model).