ARCIFACT.
ARCIFACT · live demonstration

Watch certainty
get earned.

This is a real software project (PyTorch’s test pipeline) with fourteen facts hidden. Ask a question. The Oracle tells you exactly how unsettled it is and, where one exists, names the single fact that would settle it. Sometimes none does, and it says so. Reveal that fact and possibility collapses into proof, in front of you, with the count as the receipt under the declared model.

warming up…
about 1.2 billion schedules for every person on Earth still possible · computed exactly, on your device, in –

1 · Ask a question

Pick a question above.

2 · Reveal the key fact

Ask a question first, and the Oracle will point at what would settle it.

Or investigate yourself · all 22 candidate facts (8 are deliberately fake)

3 · Your investigation

Nothing revealed yet.
Bonus: draw one random valid schedule

One schedule, chosen exactly uniformly among the quintillions, in about a millisecond.

    For sceptics: audit everything

    The engine runs in a Web Worker handed only the 35 visible dependencies; the hidden truths never enter it and reach it one reveal at a time. Provenance is commit-pinned: source pytorch/pytorch @ 05dab25851a8. Workflow sha256 200cbb72cd21546e124fe349… · engine-input d78c3a40de5d25a80827d7b7… · truth-deck 282aea556e888630dab61cc4…. At build time I asserted, against that exact commit, that all fourteen hidden dependencies exist and all eight decoys do not. P here means the share of uniformly weighted valid orderings, not GitHub’s scheduler odds. (sha256 of the canonical payload, excluding the self-referential prov.data_sha256 field: c15e4803026fa3d776f93d0a…. Recompute: shasum -a 256 oracle_data.json)

    # extraction: pinned workflow -> edges
import yaml, hashlib, urllib.request
u=('https://raw.githubusercontent.com/pytorch/pytorch/'
   '05dab25851a8b0b2925aed717dc010ddc643280c/.github/workflows/pull.yml')
raw=urllib.request.urlopen(u).read()
assert hashlib.sha256(raw).hexdigest()=='200cbb72cd21546e124fe349635221e4bb0b48a177a0f65cbd7c042128b001e5'
doc=yaml.safe_load(raw); jobs=list(doc['jobs'])
edges={(p,j) for j,s in doc['jobs'].items()
       for p in ([s.get('needs',[])] if isinstance(s.get('needs',[]),str)
                 else s.get('needs',[])) if p in jobs}
print(len(edges),'true dependencies at the pinned commit')
    import json
d=json.load(open('oracle_data.json'))
jobs=d['jobs']; ix={j:i for i,j in enumerate(jobs)}; k=len(jobs)
pred=[0]*k; succ=[0]*k
for a,b in d['base']:
    pred[ix[b]]|=1<>x&1 and pred[x]&S==pred[x]:
            t=S|1<>x&1 and succ[x]&S==0)
print(F[max(ideals)])   # the world count, from scratch

    What you just used

    Arcifact’s core invariant, playable: while the evidence permits conflicting answers, the system reports the split and hunts the settling fact; only when every valid world agrees may it assert, and the assertion carries the count. The same gate governs agent actions, citations and policy claims. Certify your system →

    Exact integer arithmetic over the ideal lattice; Python and browser agree on the 19-digit count digit for digit. Arcifact · nothing asserted past evidence.
    // this is a laboratory. the product is next door.

    The same reasoning, pointed at your repository.

    The Oracle hides fourteen facts about PyTorch's pipeline so you can watch certainty being earned rather than asserted. It exists to show the method honestly, including where it says it does not know.

    Arcifact Gate is that reasoning applied to a real question: what does a green required check on your repository actually prove? It reads the declared workflow, finds the jobs that can fail while the check still passes, and constructs one concrete combination of results where that happens.

    You can run it on any public repository right now, with no installation and no account.

    Check a public repository What Gate does

    Static analysis, nothing executed, and unknown stays unknown. The same rule the Oracle demonstrates.