The right codec for every column — chosen automatically.
AT-1 auto-selects the best transform and entropy back-end per numeric column and guarantees the result is never larger than a strong general compressor — so turning it on is pure upside. Validated across 27 real data tiers.
Auto-tuned per column
AT-1 tries a small battery of decorrelating transforms (delta, double-delta, linear-predictive, spectral, byte-plane) against several entropy back-ends — including brotli — and keeps the smallest. You don't pick a codec; the data does.
Never worse than zip — guaranteed
Every container is compared against a strong general-compressor baseline, and the smaller one always wins. There is no input — not even adversarial or already-compressed data — that makes AT-1 expand your file. It is enforced in CI on every release.
Signal-aware predictors
Sensor, vibration, biosignal and audio columns get a linear-predictive (AR/LPC) or spectral model that captures structure a general compressor can't see — the source of the double-digit wins on machine and instrument data.
Same verified container
Output carries AT-1's SHA-256 integrity trailer and stays queryable and addressable. You get the extra ratio without giving up tamper-evidence or in-place query.
Measured wins on real data
Extra reduction on top of a strong general compressor (xz/brotli), lossless and bit-exact. From the AT-1 compression lab — 27 tiers, 12,555 pipelines.
| Data type | Extra reduction | Auto-selected method |
|---|---|---|
| Sorted IDs / DB keys | −44% | delta + brotli |
| Tabular numeric | −38% | delta + brotli |
| Financial tick | −32% | delta + brotli |
| Hydraulic sensor | −31% | delta + AR/LPC |
| Genomic positions | −27% | delta + brotli |
| Multi-channel sensor | −26% | delta + brotli |
| LiDAR point cloud | −22% | AR/LPC + brotli |
| Smart-meter (AMI) | −18% | delta + brotli |
| EEG / biosignal | −13% | delta + AR/LPC |
| Model weights (int16) | −14% | byte-plane + brotli |
Tiers where structure isn’t present (random bytes, hashes, already-compressed data) fall back to the baseline — never worse, never a surprise.
Turn it on in one command
# auto-select transform + entropy back-end, never worse than baseline
at1 optimize compress readings.npy readings.at1o
at1 optimize verify readings.at1o # confirm the SHA-256 integrity trailerNeed a guaranteed-error lossy mode for scientific and sensor archives? See Bounded mode.