Docs · Start

Example experiments

Commands that were run and checked on this checkout, with what they write.

Check the environment

$ qmint doctor$ qmint sources list --type paper | head -n 3

doctor prints dependency status, lane entrypoints (PRESENT or NOT_AVAILABLE) and ends with status: OK. sources list prints one JSON object per row of research/PRIOR_ART_MATRIX.csv (24 papers).

Reproduce published bounds

$ python -m qmint.proofs.reproduce --out /tmp/qmint-proofs

Recomputes every numerical example quoted by the registered sources and compares it with the paper (under a second, about one second). Each line reads [ok] name: paper=… qmint=…. Examples:

Check	Paper	QMint
W1 min length for ε ≤ 1e-9	73	73
W1_CV min length for ε ≤ 1e-9	273	273
S1 min length for ε ≤ 1e-9	131	131
P_noqub at μ = 0.09 (Kent 2022)	0.0038	0.003815
W1 n=2 SDP, certified upper bound vs exact dual	0.562500000007276	9/16

It writes reproduction_v1.json and example certificates checked by lane D's independent checker. Those are worked examples of the certificate format for the stated hypothetical parameters, not claims about a deployed system.

Attack suite and SDP certificates

$ python -m qmint.attacks.report --out /tmp/qmint-attacks.json --trials 2000 --seed 7

Runs the lane C attack suite (about 20 s at 2 000 trials): heuristic counterfeiters as Monte Carlo lower bounds, the n = 1 and n = 2 W1 SDPs (NUMERICAL_SDP), and exact rational dual certificates (CERTIFIED_NUMERIC, for example 3/4 ≤ OPT₁ ≤ 0.750000000003638). Add --clarabel-n2 for the Clarabel cross-check at n = 2, which needs about 6.5 GB of memory.

Classical baselines

$ python -m qmint.classical.run_baselines --out /tmp/qmint-classical --pqc-reps 1

Runs the classical comparators H0–H6 as deterministic state machines and writes 23 C6 BenchmarkRecords (MODELLED_LATENCY), ML-DSA timings and comparability_matrix.csv, which states where quantum and classical rows are not comparable.

Offline core run

$ python scripts/run_core.py --offline --seed 42 --out /tmp/qmint-core

Network is denied for the whole run. Stages whose lane has no entrypoints module yet are reported as SKIPPED with the reason; on this checkout that is every lane except L, so the run ends with PASS=2 FAIL=0 SKIPPED=9 and a verified manifest. Use --require-all to turn skips into failures.

Call the API

$ curl -s localhost:8787/api/v1/meta/tracker | jq '.data.gates[] | [.gate_id, .status]'$ curl -s localhost:8787/api/v1/research/sources/T-MVW12-THM1 | jq '.data.entry.title'$ curl -s -X POST localhost:8787/api/v1/lab/attack \    -H 'content-type: application/json' \    -d '{"attack": "measure_resend", "n": 4, "backend": "analytic"}' | jq '.status, .data.theorem_bound.value'

The attack call returns the MVW bound (3/4)⁴ = 0.31640625 as PUBLISHED_THEOREM and the exact measure-and-resend both-pass value (5/8)⁴ = 0.152587890625, each as a labelled Quantity. Its status and reason say which lane computed it.