Wave 4 — what the field's last six months change, and what they don't
Written 2026-09-22 from a review of quantum computing and quantum-for-AI results through September 2026. The roadmap's thesis (quantum-wave-roadmap.md): durable quantum advances live in one-way dependencies (entropy), verification artifacts (correspondence benchmarks), interface contracts (QUBO) and infrastructure — never in putting a QPU on the hot path. The last six months vindicated it: every claimed quantum-ML advantage on classical data has been dequantized (QCNNs classically simulable to 1,024 qubits, PRX Quantum Apr 2026; no unambiguous QAOA advantage, Quantum Sep 2026; IBM's advantage tracker lists no ML task), while AI improving quantum hardware (Google's RL calibration of Willow, Nature Jul 2026) and quantum-inspired classical methods (tensor-network LLM compression, simulated bifurcation) demonstrably work. So this wave adds nothing to the hot path and tightens what is already there.
| # | item | status |
|---|---|---|
| W4.1 | Bit-order cliff. qBraid's unreleased 0.13 flips measurement bitstrings to little-endian across providers; _measured_index encodes the per-device convention verified on real devices. Pinned qbraid>=0.12,<0.13, qbraid-core>=0.6.4,<0.7; the bench now runs a bit-order canary first (core.bit_order_canary: an amplified peak at index 1 whose bit reversal is index 8) and fails, even under --update-baseline, if it trips. | shipped |
| W4.2 | IonQ gate cap. Braket rejects >2,000 gates per circuit on IonQ on demand (Jun 2026) after the per-task fee; run_qasm refuses before submission. | shipped |
| W4.3 | Bell certificate per harvest. harvest runs CHSH on the same device first and records S in the provenance line (Quantum Origin's design: a Bell-test seed feeding an extractor). A device that does not violate the classical bound that day is refused as an entropy source; --no-certify records bell: null. With Open Quantum devices at zero qBraid credits (Jul 2026), a certified monthly harvest is free. | shipped; the harvest itself needs the hosts (T1.5) |
| W4.4 | Readout mitigation, tensored. bench --mitigate-readout runs two calibration jobs (|0…0⟩, |1…1⟩) and inverts the per-qubit confusion on every scenario's counts (core.readout_calibration, core.mitigate_readout). Two tasks, not 2ⁿ, because the per-task fee is the cost. | shipped |
| W4.5 | Calibration-aware bench. calibrate ranks candidate qubits by retained P(1) after a delay on the native Rigetti route (decay.rank_qubits); bench --layout places each recall circuit on the chosen qubits (core.place_on). The measurement is taken right before the run it informs and recorded with it. | shipped |
| W4.6 | The Q4 hardware run, pre-registered. Same 5-scenario subset as ledger row 1, on Cepheus (99.1% 2Q) and on IQM Emerald (99.5% CZ, $0.0016/shot on Braket, zero credits via Open Quantum), each with and without W4.4/W4.5. Kept if agreement ≥ 80% with a 95% interval excluding 60% on at least one device; declined if no configuration clears 60%; the ledger gets one row per configuration and the depth-versus-fidelity question is answered by the difference between devices, not by a story. Budget: under the standing $5/quarter. | registered here; runs next quarter |
| W4.7 | Quantum-inspired, the part that pays (kannaka-memory). (a) Simulated Bifurcation as a ConsolidationSolver beside ClassicalAnneal, judged by T3.5's one-week dream diff. (b) Tensor-network compression of kannaka-brain-7b-v1 for CPU serving, accepted only if E-005's anchored faithfulness stays above its 0.808 floor and latency halves. | proposed; separate repo |
| W4.8 | QuantumOS boot seed. scripts/qseed-from-reservoir.sh draws 256 bits from the reservoir and prints the qseed= kernel argument, so a weekly boot carries real provenance through the existing attestation path. | shipped |
What this wave deliberately does not do
- No quantum transformer, variational classifier or kernel on HRM data: dequantized.
- No QAOA on hardware for consolidation: no measured advantage; QAOA stays the simulator reference for the QUBO interface contract.
- No claim that recall-as-amplitude-amplification is a speedup. It is a verification artifact; Aaronson's input-loading objection stands and the first hardware QRAM is at 81% fidelity for four bits (Nature Physics, Jun 2026).