Estimate rare evidence
The first paper studies amplitude amplification when an observation is unlikely. Its query result belongs to a mathematical oracle model, not a measured runtime speedup.
Quantum research / IBM Heron studies
QANTIS
Quantum evidence for a classical decision.
The research asks whether a quantum processor can help estimate unlikely observations. It returns a belief update to a classical planner, which still owns the decision and action.
Two public arXiv preprints · QANTIS Community Edition (MIT) · research, not a deployed autonomy system.
01Research model
Where the quantum step fits.
A planner starts with a belief about the world. The quantum step estimates how likely a new observation is; the result is converted back into an ordinary belief before any decision is made.
- 01
Classical input
A prior belief and observation model define the update requested by the planner.
- 02
Evidence estimation
IBM Heron circuits estimate the rare-event evidence term under a calibrated operating envelope.
- 03
Ordinary posterior
The service returns a classical posterior rather than replacing the planner or policy layer.
- 04
Decision check
The study compares the resulting immediate action with the action selected by exact Bayes.
What the studies actually tested
These are bounded findings from the public papers. The papers below carry the methods and full conditions.
Return a usable belief
The later paper returns a classical posterior to a classical planner. In its reported Tiger POMDP checks, the hardware-derived and exact Bayes posteriors led to the same immediate action.
Test a tracking case
The first paper also tests an 11-variable tracking assignment on hardware. Classical baselines were faster and exact on the small tested case.
Name the operating limit
On the reported circuits, error mitigation helped some shallower cases and hurt much deeper ones. Its value depends on the circuit and backend.
02Public records
Two papers. One continuous research line.
The 2026 records separate foundational hardware feasibility from the later sequential belief-update study.
Public arXiv preprint · February 28, 2026
Foundational hardware feasibility
arXiv:2603.00785v1
31 pages · 4 figures · 12 tables · quant-ph + cs.AI
The foundational QANTIS paper reports a hardware campaign counted as 45 experiments across POMDP belief conditioning and multi-target data association on three IBM Heron backends; the count is not 45 independent replications.
Paper recordOpen +Close −
Full title: QANTIS: A Hardware-Validated Quantum Platform for POMDP Planning and Multi-Target Data Association
- Grover amplitude amplification for POMDP belief conditioning
- Closed-loop hybrid quantum-classical Tiger POMDP
- FPC-QAOA for multi-target data association
- 45 reported campaign experiments across IBM Heron backends, not independent replications
Public arXiv preprint · July 7, 2026
Sequential belief updates
arXiv:2607.06760v1
10 pages · 6 figures · cs.AI + quant-ph
CC BY 4.0
A controlled hardware case study that treats the quantum processor as a calibrated belief-update service: it receives a prior and observation model, estimates rare-event evidence, and returns an ordinary posterior to a classical planner.
Paper recordOpen +Close −
Full title: QANTIS: Hardware-Calibrated Sequential POMDP Belief Updates on IBM Heron
- Sequential Tiger POMDP runs at 8 and 12 primary steps
- 20-step and 32-step controls within the reported operating band
- Boundary-aware BIQAE near zero and one
- Immediate-action agreement with exact Bayes in every reported check
Publication record and research keywordsRecord +
Attribution
Author and citation details are maintained in the linked arXiv records.
Keywords
Quantum computing · POMDP · Amplitude amplification · Fixed-point amplitude amplification · BIQAE · Sequential belief updating · QAOA · Multi-target tracking · Data association · NISQ · Error mitigation · IBM Heron
03Hardware evidence
Report the operating point, not a universal promise.
Results below are reported observations from controlled IBM Heron studies. They are not wall-clock speedup, production readiness, or quantum-advantage claims.
- Probability Amplification
- 0.179 → 0.907
- one reported Grover operating point
- Hellinger Distance
- 0.0015
- hardware posterior vs exact Bayes
- Usable Shots
- 1463 → 7429
- accepted-event usable-shot yield
- QPUs Tested
- 3
- IBM Heron backends
Validated route
Problem → Simulation → Hardware → Public record
Reported backends
ibm_torino · ibm_fez · ibm_marrakesh
2 arXiv preprints · IBM Heron QPUs · Sequential posterior checks
B01The foundational study found ZNE useful below ~100 ISA gates
B02The same study found ZNE harmful above ~1000 ISA gates
B0311-variable FPC-QAOA hardware case; quality degraded at 19 variables
04Code and reading
Reproduce the work, then inspect the limits.
Three packages keep shared infrastructure, POMDP research, and tracking research composable. QANTIS can also use qmesh as an experiment-substrate layer without merging their claims.
quantum-common
Shared utilities, circuit primitives, error mitigation (ZNE, Pauli twirling), and backend abstraction layer for IBM Qiskit Runtime.
quantum-pomdp
POMDP belief-state oracle construction, Grover amplitude amplification, closed-loop hybrid planning loop, and Tiger POMDP reference implementation.
quantum-mht
Multi-target data association via QUBO formulation, FPC-QAOA solver, classical MHT baseline, and cost-matrix construction for tracking scenarios.
Synthesis4 field notes
Synthesis01
Two QANTIS IBM Heron papers, one increasingly precise research question.
QANTIS moved from a cross-backend campaign reported as 45 experiments to a controlled sequential belief-update service. Reading both papers together shows what survived hardware testing, what was narrowed, and what still needs stronger evidence.
15 min readQuantum research
Synthesis02
A critical review of QANTIS hardware evidence and its limits.
The QANTIS result is credible when read narrowly: calibrated shallow belief updates can remain close to exact Bayes in controlled hardware runs. The evidence does not establish wall-clock advantage, general scale, or end-to-end autonomy.
16 min readQuantum research review
Synthesis03
A QANTIS reproducibility checklist for quantum hardware evidence.
The right reproducibility unit is not a notebook screenshot or a best metric. It is a versioned campaign bundle that reconstructs the posterior, its hardware context, its classical reference, and the decision it affected.
14 min readQuantum reproducibility
Synthesis04
Where QANTIS quantum inference belongs in an autonomous system.
The credible role for near-term quantum hardware is not an autonomous brain. It is a replaceable inference service inside a classical perception-to-action loop, invoked only when its calibration and depth gates pass.
14 min readQuantum AI systems
Technical deep dive6 field notes
Technical deep dive01
QANTIS as a calibrated belief-update service.
QANTIS does not replace the planner. Its hardware-tested inference core accepts a prior and observation model, estimates the difficult evidence term, and hands a conventional probability distribution back to classical software.
14 min readQuantum decision systems
Technical deep dive02
QANTIS FPAA and sequential belief stability.
The headline FPAA run achieved a maximum Hellinger distance of 0.009 at 32,768 shots per step. At 10,000 matched shots, the maximum was about 0.033, so the evidence supports stability rather than equal-budget superiority.
13 min readSequential inference
Technical deep dive03
QANTIS boundary-aware BIQAE where amplitudes become fragile.
On the reported Pittsburgh backend calibration, boundary error fell from 0.6317 to 0.00224 at amplitude 0.01 and from 0.4890 to 0.00773 at amplitude 0.95.
13 min readQuantum calibration
Technical deep dive04
QANTIS rare-event evidence: an envelope, not an advantage claim.
The rarest tested row closely tracks the analytic accepted-event probability, but transpilation keeps the circuit shallow. The result maps a logical sample-complexity envelope rather than validating deep circuits or wall-clock speedup.
14 min readRare-event inference
Technical deep dive05
QANTIS posterior fidelity when the action stays the same.
The reported 8-step matched-shot, 20-step, and 32-step checks show the same immediate action as exact Bayes with zero scored cumulative value loss under the tested Tiger rule.
12 min readDecision assurance
Technical deep dive06
QANTIS on IBM Heron: circuit depth defines the current boundary.
The paper's exploratory scaling probes distinguish state-space size from compiled circuit depth. Optimized four-state cases map a shallow corridor; deeper chains become noise-frontier markers.
14 min readQuantum hardware
05Explore and access
Explore the decision example.
The interactive surface is didactic. It helps explain evidence routing; it is not the research implementation or an autonomous decision system.
01 · InferCalibrated belief
Turn noisy, partial, or rare observations into a calibrated posterior belief — the single source of truth that every downstream step depends on.
Grover-AA on POMDP belief
O(P(e)⁻¹) → O(P(e)⁻¹ᐟ²) query-complexity change under the paper's oracle model; not a wall-clock result
BIQAE
Boundary-aware Bayesian quantum amplitude estimation under bounded depth
Hellinger distance ≤ 0.0149
vs ideal distribution across T=8 hardware steps (Tiger POMDP)
qantis / decision-engine.infer
Inputs
- Raw sensor stream
- Prior model
- Observation noise model
Outputs
- Posterior belief
- Confidence intervals
- Calibration diagnostics
Stack
Can compose with qmesh → signed run manifest → offline-verifiable provenance chain.
Click a step above to inspect inputs, outputs, and the techniques QANTIS uses at that stage.
Inspect the public work.
The public repository carries the inspectable community surface. Collaboration scope is discussed separately and does not change the public claim boundary.
Compare editions
18 rows
| Feature | Community Edition · public · MIT | Collaborator Edition · private · partners only |
|---|---|---|
| Backend abstraction layerframework | Public connectors | Hardened, multi-vendor, optimised |
| Configuration & reproducibilityframework | Included | Included |
| Error mitigation pipelineframework | Baseline (ZNE, Pauli twirling) | Full mitigation & calibration stack |
| Benchmarking infrastructureframework | Illustrative | Full experimental harness |
| Infer — calibrated beliefengine | Basic surface | Calibrated, production-grade |
| Risk — event & tail-riskengine | Basic surface | Calibrated, production-grade |
| Optimise — feasible decisionsengine | Basic surface | Calibrated, production-grade |
| Verify — trust & diagnosticsengine | Basic surface | Calibrated, production-grade |
| POMDP planning (Tiger reference)applications | Included | Included |
| Multi-Hypothesis Tracking (MHT)applications | Included | Included |
| Quantum-Bio Intelligenceapplications | Not included | Included |
| CRISPR moduleapplications | Not included | Included |
| Sensor fusion · adversarial robustness · mission orchestrationapplications | Not included | Included |
| Hardware campaign artefactsops | Aggregate results in cited preprints; raw campaign artefacts not bundled | Available by governed engagement |
| Comparative benchmarks vs classical SOTAops | Not included | Included |
| Confidential datasets & mission profilesops | Not included | Included |
| Supportops | Community, best-effort | Dedicated engineering |
| Licenceops | MIT | Commercial / partner agreement |