// Reproducibility

Every result,
from a fresh copy of the code.

Each result has a step-by-step page that takes you from a fresh clone to the published numbers, tested by us from a fresh clone at the commit it names. This page lists them all.

// How it works

Clone, check out,
run, compare.

Every step-by-step page has two paths. Reproduce the published run: clone the repository and check out the commit below; the answers must be the published ones. Run the latest: the same without the checkout; the answers must still be the published ones, and what changed since is in the repository's changelog. Each page ends with "Tested": when we last ran it from a fresh clone, where, and how long it took.

Everything runs in Docker on any Linux machine, x86_64 or aarch64. Our tests ran on aarch64 (an NVIDIA DGX Spark). The base images are named by digest, and every published result set is tagged in its repository (for example results/comparisons-2026-11-15). The full list, with what each comparison requires of the answers and why: REPRODUCE.md.

Done
  • A step-by-step page for every published result, tested from a fresh clone
  • The commit each was tested at; tags for the result sets
  • Base images pinned by digest
  • The tolerance of every comparison, stated
  • A citation file for PQCF
Missing
  • An archived release with a DOI
  • An independent review of the physics, and a paper
  • An independent reproduction by someone outside Quanterall
  • Runs on x86_64 for the comparisons with external programs
// Every result

Every result,
with its page and commit.

resultstep-by-step pagetested atyou should seechecked againsttime · needs
Symbolic computationpqcf · comparison/form92e2aebevery task ok; the term counts publishedFORM 51 min (4 the first time) · 5 GB, 1 core
Feynman-integral reductionpqcf · comparison/fire92e2aebevery case equalFIRE 6.5.21.5 min (+6 to build) · 5 GB, 1 core
Four-loop integralspqcf · comparison/forcer92e2aebevery check equal; PQCF's results unchangedForcer; FORM's assertions8 min (+4 to build) · 8 GB, 8 cores
Photonic head-to-headpqcf · comparison/photonic92e2aeb50 of 50 checks agree (the count depends on the sizes reached)five photonic simulators14 min (+4.5 to build) · 5 GB, 1 core
The three-loop β-functionpqcf · comparison/betaeef0e6fbeta functions: as knownthe published coefficients15 min · 8 GB, 8 cores
Holographic QCDpqcf · comparison/holographicd77ead6each result file unchanged but for the commit, date and timesthe papers; SciPyseconds to 4 min each; IHQCD 40 min on 16 cores · 2.5 GB
Quantum opticspqcf · comparison/quantum-opticsaef6143every digit count and difference as publishedQuTiP, QuantumCumulants.jl, PIQS, closed forms75 s; 7 min for PIQS (+ images) · 8 GB, 1 core
One-loop integrals, Dirac tracespqcf · comparison/oneloop7570e62the reference files rewritten byte for byte; the tests passmpmath, LoopTools, QCDLoop, FORMtests 3.5 min; references 2 min, plus 30 for C0 · 10 GB
Feynman diagrams and rulespqcf · comparison/diagrams7570e62the 24 cases; the reference file byte for byte; 23 tests passFORM's generatorseconds (+ images) · 5 GB
The running of QCDpqcf · comparison/qcd7570e62mpmath and CRunDec agree; 5 tests passmpmath, CRunDecseconds (+ images) · 4 GB
Genome researchpqcf · research/genomebb8a5dd6 of 6 checks hold; every result as publishedproven optima6 min on 8 cores · 4 GB, network
GenHub2, phase Agenhub2 · acceptanced6a1733all 20 checks; every result as published (a reduced run without GRCh38: 14)NCBI Virus, Nextclade20 min on 10 cores · 25 GB, 16 GB memory
GenHub2, phase Bgenhub2 · phase B3a68ed520 of 20 checks holdone node's results10 min after phase A
Repetition-code sweepsqcs · sweeps3c31766PASS, all 88 rowsthe closed formseconds · Python, NumPy
Surface code under noiseqcs · noise3c31766every count as publishedStim7 min (+7 to build) · 4 GB
Benchmarks against Qiskit Aer and Stimqcs · benchmarks3c31766the circuits as published; your timesQiskit Aer, Stim7 min · 4 GB
Photonic schemesqcs · photonic3c31766the output files identical to the published onesthe papers, Perceval7 min (+5 to build)
The ten demo circuitsqcs · demos3c3176639 passed, 0 failedthe circuits' exact predictions2 min to build, 11 s · 2 GB

pqcf: github.com/quanterall/pqcf · genhub2: github.com/quanterall/genhub2 · qcs: github.com/quanterall/quantum_computing_simulations. If a result does not reproduce, please open an issue in its repository with your machine and the output.

// Methods and development practices

How the work is done,
and who is responsible for what.

AI-assisted implementation. Most of the code since 28 October 2026, and the checks around it, were written with an AI coding assistant (Claude, by Anthropic), directed by our team. That is why so much is dated within days, after a project begun in March 2026.

  • Human research direction and review. Our team chooses the problems, the methods and the references, reads the results, and decides what is published and how it is worded.
  • Independent reference implementations. We rely on other people's tools rather than on the assistant: every result is compared with an established program (FORM, FIRE, Forcer, QuTiP, LoopTools, CRunDec, Stim, Qiskit, The Walrus, Piquasso, Perceval and others), a published result, or an exact formula.
  • Automated testing. Each repository's tests run on every change, on x86_64 and aarch64; a comparison that disagrees fails.
  • Reproducibility. Every published result has a step-by-step page, tested by us from a fresh clone at the commit it names (above).
  • Still needed: independent scientific review. No independent physicist has reviewed the published physics work yet. Independent review is planned as part of the preprint process.