Parameter sweep as a first-class object

One `SweepRunner(params, kernel_template, post_process)` driver replaces the hand-rolled loops in g1.py + g3.py.

Results · real Selene shots

sweep_runner_drives_g1 · verdict_matches_reference · metrics_within_mc_error

Migrated G1 driver runs through quantum/sweep.py (SweepSpec + SweepRunner). Output matches nadarasa_g1.json: refutation verdict identical (`consistent_with_g1`), Δ worst_violating_excess = 0.021, Δ min_srp_minus_violating = 0.001 — both within Monte-Carlo error at 400 shots/slope. ~80-line slope/template/JSON boilerplate eliminated.

verdict parity · Δ headline metricMATCH · |Δ| ≤ 0.021

JSONsrc/data/demos/nadarasa_g1_via_sweep.json

Open full Nadarasa page →

Why this matters

G1 and G3 share a kernel shape — coset state, per-qubit controlled phase, terminal H + measure — and differ only in the host post-processor. Today each script duplicates ~80 lines of slope-loop, template-render, JSON-write boilerplate. A SweepRunner makes adding G4 (and every future Gn) a 20-line file.

What the repo already has

quantum/sweep.py (SweepSpec + SweepRunner) plus the migrated quantum/nadarasa_g1_via_sweep.py driver that consumes it.

What's missing

Nothing required by the card. nadarasa_g1.py and nadarasa_g3.py are kept as the authoritative reference drivers; the migrated G1 (and any future Gn) goes through SweepRunner.

Smallest experiment

Build it or kill it

Qubits

n/a — pure refactor

Ancilla pattern

n/a

Shots

Re-run G1 through the runner; verify headline metrics match within Monte-Carlo error.

Predicted outcome

Migrated driver emits nadarasa_g1_via_sweep.json with the same refutation verdict as the reference, and the headline metrics agree within Monte-Carlo error at 400 shots/slope.

Refutation criterion

If the migrated driver's refutation verdict differs from the reference, or the headline metrics drift beyond Monte-Carlo error (~0.05), the SweepRunner changed semantics — revert.

Kernel sketch

Untested — sketch only

# quantum/nadarasa_g1_via_sweep.py (verified migration)
spec = SweepSpec(
    name=f"g1_N{N}_p{p}_{branch}",
    params=[{"s": s, "n": n, "N": N, "p": p} for s in slopes],
    template=lambda pp: render_kernel(pp["n"], pp["s"], pp["N"]),
    n_qubits=lambda pp: pp["n"] + 1,
    post=lambda shots, pp: {"s": pp["s"], "hist": residue_histogram(shots, pp["n"], pp["p"])},
    shots=shots,
)
result = runner.run(spec)

Host pipeline

Per (N, p, branch): one SweepSpec → SweepRunner.run() → per-slope rows. Aggregation is the only experiment-specific code; render/import/build/shot-loop boilerplate is gone.

Files in this repo

· quantum/sweep.py
· quantum/nadarasa_g1_via_sweep.py
· src/data/demos/nadarasa_g1_via_sweep.json
· quantum/nadarasa_g1.py

More in Compilation

2 cards

Verified

JSON-as-IR: a versioned selene_run schema