QUASI manages heterogeneous compute resources — CPU, GPU, QPU — and makes allocation decisions autonomously. Programs go in, results come back. What runs where, when, and on which hardware is the OS’s concern.
12,600 lines of Rust. 375 tests. 14 backend profiles. Verified on IBM quantum hardware.
The user submits an Ehrenfest program (a Hamiltonian + evolution time + noise constraints, encoded as CBOR binary). Everything after that is automatic.
Compiles Ehrenfest programs (physics-native CBOR) to executable OpenQASM 2.0/3.0. Pipeline: deserialize → Trotterize → type check → lower to ZX-IR → noise analysis → observable synthesis → QASM emission → optimize. 3,500 lines of Rust, 208 tests.
Filter-Score-Bind pipeline (Kubernetes pattern). Hard constraints first (gate support, qubit count, noise budget), then weighted scoring across 7 criteria: gate set fit, noise margin, connectivity, queue depth, cost, preference, cache hit. 14 backend profiles with real hardware specs. 1,599 lines, 54 tests.
Content-addressed using BLAKE3. The cache key includes the circuit, backend ID, parameters, and calibration version. Changed calibration produces a different key — no invalidation logic needed. L1 in-memory, L2 filesystem. 614 lines, 17 tests.
The Solvayeur is a new kind of kernel: a resource allocator that runs on a QPU. Existing quantum operating systems (QOS, HALO, IBM QCSC) use classical schedulers to manage quantum resources. The Solvayeur inverts this — the scheduling decision itself is a quantum computation.
Concretely: the Solvayeur is an Ehrenfest program, compiled by Afana through the same pipeline it dispatches user programs to. Its scheduling Hamiltonian encodes backend contention (ZZ coupling), learned preference (Z bias), and exploration (transverse field X). After each dispatching round, the bias fields update from the observed execution reward — fidelity, latency, cost. Over time, the ground state of the Hamiltonian converges toward the optimal resource allocation.
Verified on IBM Strasbourg: 10 ATW rounds, 2-qubit scheduling circuit, 100 shots per round. The kernel explored 4 backends through quantum measurement and began converging toward Huoma as the optimal target.
Before spending QPU time, the profiler estimates whether a circuit is classically simulable. Bond dimension below threshold → run on Huoma (tensor network, up to 1,000,000 qubits in 5.2 seconds). Exponential entanglement → route to QPU. QPU is reserved for problems that genuinely need it.
| Crate | Lines | Tests | Purpose |
|---|---|---|---|
| afana | 3,500 | 208 | Ehrenfest compiler: CBOR → AST → ZX-IR → QASM |
| quasi-scheduler | 1,599 | 54 | Filter-Score-Bind + 14 backend profiles + Huoma profiler |
| quasi-cache | 614 | 17 | BLAKE3 content-addressed result cache |
| quasi-solvayeur | 986 | 27 | ATW quantum scheduling kernel |
| quasi-demo | 900 | — | Pipeline demo + VQE orchestrator |
| quasi-senate | ~5,000 | 10 | AI governance daemon |
| Total | ~12,600 | 375 | All Rust. Zero Python in the critical path. |
Real hardware specs from vendor documentation. The scheduler selects automatically.
| Backend | Type | Qubits | Native Gates | Topology |
|---|---|---|---|---|
| IBM Heron / Torino | Superconducting | 156 | RZ/SX/X/ECR | Heavy-hex |
| IBM Eagle / Marrakesh | Superconducting | 127 | RZ/SX/X/ECR | Heavy-hex |
| IQM Garnet | Superconducting | 20 | PRX/CZ | Grid |
| IQM Sirius | Superconducting | 6 | PRX/CZ | Star |
| IonQ Aria / Forte | Trapped ion | 25–36 | GPI/GPI2/MS/ZZ | All-to-all |
| Quantinuum H1 / H2 | Trapped ion | 20–56 | RZ/RX/RY/ZZ | All-to-all |
| Rigetti Ankaa-3 | Superconducting | 84 | RZ/RX/iSWAP | Grid |
| AQT Pine | Trapped ion | 24 | RZ/RXX/R | All-to-all |
| Huoma ProjectedTTN | Tensor network | 1,000,000 | All | All-to-all |
| Statevector simulator | Exact | 30 | All | All-to-all |
Afana-compiled, 30 gates. IBM Torino (156q Heron).
50.9% |0⟩ / 49.1% |1⟩.
Job d7cui865nvhs73a53h70
10 Trotter steps, 94 gates. IBM Strasbourg (127q Eagle).
70.7% ground state. Consistent with Afana’s 0.763 fidelity estimate.
Job d7cuqjp4p4gc73f5o63g
10 ATW rounds on IBM Strasbourg. 2-qubit scheduling circuit, 100 shots per round. The kernel explored 4 backends (Huoma, IBM, IQM, Quantinuum) through quantum measurement. Measurement distributions near-uniform early (transverse-field regime), bias converging toward Huoma as optimal after reward feedback.
QUASI is named after the Quasi-Ergodic Hypothesis (1911) by Paul Ehrenfest and Tatiana Afanasyeva. The Afana compiler carries her name. The Solvayeur is named after the Solvay Conferences where Ehrenfest shaped the foundations of quantum mechanics.