Postselect decoded QEC shots by decode confidence. Stim simulation: up to 68% fewer logical errors. Hardware (recomputed against protected-logical): ~14% fewer at 20% discard on a degraded-day Fez run with 927 real failures (single-job), plus 38.8% on IBM Torino via postselection. The earlier IBM Fez "41.3%" was an odd-parity readout-event reduction on below-threshold data with zero logical errors, not a protected-logical result.
Calibrated to hardware validation data from IBM Heron and Rigetti Ankaa-3 (Feb 2026).
All shots accepted (no veto)
,
Total Shots
,
Logical Errors
,
Logical Error Rate (LER)
High-risk shots vetoed
,
Accepted
,
Vetoed
,
Errors
,
Logical Error Rate (LER)
SafetyGate adapts to each vendor's noise profile. Calibration windows are hours, not days, so results are reported per regime, stable sessions deliver a larger uplift, drift-stratified sessions a smaller one. Same safety guarantee in both.
Hardware protected-logical
~14% / 38.8%
Fez 20% discard (single-job) / Torino postselection
Drift-stratified (IBM Fez)
7.7% odd-parity event reduction
confidence-ranked, stable blocks
Throughput retained
71-96%
Regime-dependent
Total validated
26,000+ shots
4 backends, 2 vendors
Hardware validation: IBM Heron (Torino, Fez, Marrakesh) + Rigetti Ankaa-3 (Feb-Apr 2026). Fez April 7 session showed intra-session drift (monotonic detection-rate ramp, R² = 0.992), expected physics for superconducting qubits between calibration windows. The earlier “23.1%” CW number is retired; the drift-stratified baseline is 7.7%. See SafetyGate guide for full drift-audit methodology.
Surface codes encode logical qubits in a 2D grid of physical qubits. Error measurements detect faults, and decoders attempt to correct them. However, correlated errors can create patterns that fool the decoder.
SafetyGate analyzes error indicators to compute a risk score. Shots likely to cause errors are vetoed before they can corrupt results.
Run on IBM Heron and Rigetti Ankaa-3 quantum processors,26,000 shots across 4 backends. The earlier "39-57% on real hardware" figure mixed an odd-parity readout-event proxy (IBM Fez/Marrakesh, below threshold) with a withdrawn Rigetti number; the protected-logical hardware reduction (recomputed) is ~14% at 20% discard on a degraded-day Fez run with 927 real failures (single-job), plus 38.8% on IBM Torino via postselection. Below-threshold data has zero logical errors to reduce. Larger protected-logical reductions are shown in the Stim-simulation matrix below.
~14%
protected-logical, 20% discard, Fez single-job (+38.8% Torino)
26,000
Hardware Shots
2
Vendors
4
Backends
SafetyGate adapts its behaviour to the current calibration regime automatically. Calibration windows drift; SafetyGate responds.
| Regime | Backend | Shots | Baseline LER | Gated LER | Action |
|---|---|---|---|---|---|
| Below threshold | Marrakesh | 2,000 | 0.00% | 0.00% | 99.9% allow |
| Near threshold | Torino | 18,000 | 0.34% | 0.21% | 39% reduction |
| Near threshold | Ankaa-3 (Rigetti) | 8,000 | 0.21% | 0.09% | 57% reduction |
| Above threshold | Torino (IBM) | 12,000 | 38-49% | n/a | 100% block |
d=3 repetition code. IBM Heron: 6 jobs, 62 corrupted shots. Rigetti Ankaa-3: 1 job, 17 corrupted shots.
SafetyGate vetoes 29% of shots, retaining 71% throughput. Among allowed shots, logical error rate drops 39%. Net effect: higher reliability at moderate throughput cost.
Multi-round syndrome analysis is the key differentiator. Detection accuracy degrades significantly without temporal context, confirming that single-round heuristics miss correlated failure modes.
When the backend drifts outside its published calibration window, SafetyGate blocks 100% of shots rather than returning results with understated error bars. This prevents compute waste and protects downstream applications from results they cannot defend.
100,000 Stim-simulated episodes across 4 code distances and 5 noise regimes. Wilson 95% confidence intervals. Every cell shows 88-100% fewer logical errors in simulation.
88-100%
LER Reduction
100,000
Episodes Validated
20
Test Configurations
<1%
Typical FAR-bad
| Distance | STABLE | AGING | STEP CHANGE | GLITCH | LEAKAGE |
|---|---|---|---|---|---|
| d=3 | 97.4% FAR: 1.4% Allow: 51% | 95.8% FAR: 1.8% Allow: 42% | 93.8% FAR: 2.4% Allow: 39% | 91.4% FAR: 4.7% Allow: 54% | 100.0% FAR: 0.0% Allow: 52% |
| d=5 | 96.7% FAR: 1.0% Allow: 31% | 97.3% FAR: 0.4% Allow: 15% | 96.0% FAR: 0.6% Allow: 14% | 95.9% FAR: 1.1% Allow: 28% | 98.4% FAR: 0.5% Allow: 28% |
| d=7 | 96.6% FAR: 0.7% Allow: 19% | 99.0% FAR: 0.1% Allow: 6% | 95.8% FAR: 0.2% Allow: 5% | 96.1% FAR: 0.6% Allow: 16% | 94.8% FAR: 0.9% Allow: 17% |
| d=9 | 94.6% FAR: 0.7% Allow: 12% | 88.3% FAR: 0.2% Allow: 2% | 91.8% FAR: 0.1% Allow: 2% | 95.3% FAR: 0.4% Allow: 8% | 96.5% FAR: 0.3% Allow: 10% |
LER Reduction %. FAR-bad = fraction of allowed shots that carry logical errors. Higher allow rate = more throughput retained.
As code distance grows, temporal correlation awareness becomes essential. SafetyGate maintains 88%+ fewer logical errors in simulation at every distance, with FAR-bad dropping below 1% at d=5 and above.
88% floor across all configs
Below 1% at d=5+
Explore the tradeoff between throughput cost and safety benefit. Each point is a validated noise regime. Slide to see how SafetyGate adapts.
Current Operating Point
d=5, STABLE
96.7%
LER Reduction
69.5%
Block Rate
30.5%
Throughput Retained
0.99%
FAR-bad
96.7% of errors eliminated
Click points on the chart or slide to explore how SafetyGate adapts to different noise conditions. Higher block rate = more aggressive filtering = better safety at throughput cost.
No. SafetyGate runs alongside existing decoders and does not modify correction logic.
Yes. SafetyGate is a conservative safety layer. It may veto shots that would decode correctly in order to reduce overall logical risk.
SafetyGate uses vendor-specific noise profiles but the approach is hardware-agnostic. Validated on IBM Heron (3 backends) and Rigetti Ankaa-3. Vendor-specific noise profiles with <0.5% cross-backend variance within IBM.
SafetyGate computes a risk score from error indicators.
High scores indicate error patterns that statistically correlate with failure.