AT-1 Artifact

Your data’s source code

To make bytes small, a compressor has to discover the structure in your data. AT-1 emits that structure as a first-class artifact — the invariants that hold across columns (c = a + b, f = 3·a, monotone counters), the marginal distributions, the category mix — then generates a synthetic twin from the artifact alone, and proves the twin’s fidelity information-theoretically.

Get the CLI Generator discovery →
your data's
source code: invariants + marginals recovered automatically
10%+
MDL gain — the twin, as a dictionary, shrinks the real data
beats random
twin dictionary beats a random one — it captured structure
invariants hold
every discovered rule holds exactly in the synthetic twin

Recover the model, mint a certified twin

at1 artifact build data.csv -o data.at1artifact
#   3 invariants; twin CERTIFIED (MDL gain 10.4%, vs-random
#   10.8%):
#     - total = a + b
#     - triple_a = 3 * a
#     - seq is monotone non-decreasing
at1 artifact twin data.csv -o synthetic.csv   # share a twin, not the data
at1 artifact certify data.csv synthetic.csv   # prove the fidelity

A real deliverable

The artifact is executable documentation: schema, invariants, distributions, a drift baseline. It travels where the raw data can’t.

Proven, not asserted

Fidelity gates on the robust signal: the twin used as a compression dictionary shrinks the real data, and beats a random dictionary. NCD is reported as a secondary diagnostic, honestly non-gating.

Honest scope

Invariant search covers sums, scalar multiples and monotonicity across up to a dozen numeric columns; marginals are per-column. Richer joint structure is on the roadmap, not claimed.

Billed per artifact/twin build — first 500/month free. See pricing.