Reproduce & drift
Reproduction is QuantumLedger's "aha": take a run you recorded, re-run its exact circuit against a device state that has drifted over time, and measure how much the result changed — with a single score and a plain-language verdict.
Running a reproduction
CLI (offline):
ql reproduce <run_id> --days 90 --profile bad_day --html report.html
In the app: open a run at /app/records/<run_id> and use the Reproduce form — pick a number of
days to drift and a profile, then submit. The reproduction (and its report) attach to the run.
Drift profiles
Drift is modelled deterministically (seeded), so the same inputs always give the same result — no flakiness. The number of days advances the calibration snapshot's timestamp and degrades T1/T2, readout, and gate errors accordingly. Three profiles:
| Profile | Meaning |
|---|---|
typical |
Gradual, realistic day-to-day degradation. |
bad_day |
Aggressive degradation — a stress test. |
recalibrated |
The device improved (a fresh calibration). |
Scoring: how similar are the two distributions?
The primary metric is Hellinger fidelity — a 0–1 similarity between the original and reproduced measurement distributions, where 1.0 means identical. Secondary metrics give more texture:
- Total Variation Distance (TVD) — the largest single probability gap; lower is more similar.
- Jensen–Shannon divergence — a symmetric information-theoretic distance.
- Shot-noise confidence intervals — whether the difference is within what sampling noise alone explains.
Verdicts
The fidelity is bucketed into a human verdict so you don't have to interpret a raw number:
| Verdict | Meaning (approx. Hellinger fidelity) |
|---|---|
| reproducible | ≈ identical, within shot noise (≥ 0.99) |
| drifted | close but measurably changed (≥ 0.90) |
| divergent | noticeably different (≥ 0.70) |
| irreproducible | far apart (< 0.70) |
Reading the diff report
The report explains why the distribution changed, not just that it did:
- Calibration drift — per-qubit T1 / T2 / readout changes, with percentages.
- Gate errors — per-gate error changes (CX, SX, …).
- Transpilation delta — changes in compiled-circuit depth, size, and CX count.
- Backend substitution — flagged if the reproduction ran on a different device/vendor.
- Top bitstring shifts — the measurement outcomes whose probabilities moved the most.
Together these turn "the result changed" into "the result changed because qubit 3's readout error doubled and the transpiler added two CX gates" — an explanation you can act on or cite.
Next: Result cards & badges.