Prove your data is real. Disclose what’s synthetic.
The world is drowning in AI-generated data and the rules are catching up — the EU AI Act now demands synthetic content be disclosed, and every dataset’s provenance is suddenly a liability. AT-1 Reality is the layer that answers, with evidence: what is this data, where did it come from, and can you prove it?
- 0.95–0.99
- fabrication-screen AUC on real labelled data
- real / synthetic
- per-record composition, sealed + lineaged
- certified
- every finding carries a tamper-evident receipt
- EU AI Act
- turnkey synthetic-disclosure provenance
Authenticity tools are usually a classifier’s opinion. AT-1’s run on the same principle as the rest of the stack: byte-exact compression is a lie detector. Real, measured data and fabricated data have different structure, and structure is exactly what a lossless compressor sees. Every finding is bound to a tamper-evident, byte-exact certificate— so it’s evidence a third party can re-check, not a score to trust.
Four instruments, one layer
How was this file produced — real capture, simulation, model-generated, or hand-edited? A classifier reads the compression signature of the data itself, not metadata that can be forged.
ExploreFlag records that were made up rather than measured. On real labelled data the screen reaches AUC 0.95–0.99 and survives a spectrum-matched adversary that tries to look real — because fabricated data has the wrong kind of structure.
Seal a training corpus that proves its own composition — the exact % real vs synthetic, per-record lineage — so any value tamper or composition lie is detected. The provenance travels with the data.
ExploreMeasure how strongly a model memorized a given text — its compressibility under the model, controlled against a generic baseline so genericness alone can't fool it. An evidence-gathering instrument you run yourself, not a verdict.