webgpu-q — experiments

Research protocol for the six-level breakthrough ladder. Protocols live in experiments/level-N-<slug>/protocol.md. Artifacts go to experiments/results/YYYY-MM-DD/level-N/. See RESEARCH.md for the master doc.

Level 1 — Statevector

Hypothesis: full-featured statevector sim in a browser tab, reaching bandwidth-bound performance. Pass bar: F ≥ 1 − 1e-5 vs CPU reference, ≥ 40% of peak memory BW at N ≥ 22, T(N) slope ≈ 1.0 at N ≥ 18.

Level 2 — MPS

Hypothesis: a CPU-only MPS simulator in a browser tab reaches ≥ 72 qubits for low-entanglement circuits, with χ-bounded error that tracks bipartite entropy. Pass bars: (E5) F ≥ 0.999 vs CPU statevector at χ = 64 on every cell; (E6) no OOM and median sweep ≤ 1 s at N = 72, χ = 32, depth = 4; (E7) fit slope of log₂(χ_req) vs depth within 1.0 ± 0.2.

Level 3 — Kernel fusion

Hypothesis: fusing k single-qubit gates on the same qubit into one dispatch collapses α — the fixed dispatch cost from E4 — by a factor of k. Pass bars: (E8) fused = unfused at F ≥ 1−1e-5 on all cells; (E9) α_eff(32) ≤ α_eff(1)/16 AND ≤ 20 μs; (E10) best speedup ≥ 2× AND N=20 regression-free (≥ 1.05×). Scope note: only same-qubit chains are fused; layer fusion across different qubits (brick-wall) is deferred to a future kernel.

Level 6 — Chemistry

Hypothesis: VQE on H₂ in STO-3G with a hardware-efficient ansatz reaches chemical accuracy (|ΔE| ≤ 1.6 mHa) at every bond length across the dissociation curve, from 10 random classical inits per R. Pass bar: median |ΔE| ≤ 1.6 mHa AND max |ΔE| ≤ 5 mHa at every R. Hamiltonian is built from STO-3G molecular integrals on the fly — each R gets its own 16×16 dense matrix; reference E_FCI is the Jacobi-diagonalized spectrum of THAT matrix.

Levels 4–5 — deferred