Point it at your data. It finds the law — and proves it.
AT-1 Discover recovers the generatorthat produced a sequence — the shortest program that reproduces every value — and proves it the only way that can’t be faked: by regenerating the data byte-for-byte. It’s automated science with a lie-detector built in.
- byte-exact
- regeneration is the proof — no vibes
- 35.6×
- a 320-byte CA recovered to a 9-byte rule
- no false laws
- random / CSPRNG data yields nothing
- .at1gen
- a few coefficients that regenerate everything
Every other AI “insight” is a claim with a confidence score. Discover’s claim is a program that regenerates your exact data, bit-for-bit — you can’t fake a byte-exact reconstruction. The compression ratio is the p-value.And it never invents a law that isn’t there: point it at random or encrypted data and it honestly reports no generator.
What it recovers
Linear and modular recurrences, LCG/LFSR generators — the value stream is a hidden state machine, recovered to its coefficients.
output[i] = f(seed + i) for a hash or arithmetic map (crc32, squares-mod, …) — the column is a pure function of its position.
row[i+1] = rule(row[i]) over a fixed width — sequences that look random collapse to a single 8-bit rule and a seed.
When the raw stream resists, it cracks a delta or xor transform of it — store the head plus the rule of the differences.
How it works
at1 discover series.txt -o rule.at1gen # RECOVERED: elementary-CA ratio=35.6x [9 B program for 320 B of data] at1 discover verify rule.at1gen # VERIFIED - regenerated 80 values byte-exact
You hand it a column of numbers; it hands back a tiny .at1gen — the recovered rule — that regenerates every value on demand, SHA-256 verified. A hypothesis only counts if it reproduces the whole sequence exactly; a rule that gets one value wrong is rejected, not rounded.
Discover is the productised core of AT-1’s Hypothesis Engine. It executes none of the offline model-written code that inventsnew hypotheses — the shipped engine runs only verified, deterministic recovery, so it’s safe to point at untrusted data. Pairs with generator recovery, the RNG auditor, and zero-scan queries — all the same act: recover the generator, then use it.