Proof on real data

Measured, not promised

On real public datasets, AT-1 (TinyFiles) stores structured data 1.1–2.5× smaller than xz-9 and 23% smaller than Parquet-zstd — while keeping it queryable in place, verified byte-for-byte on every row. For example: 209× on 1000-Genomes VCF (2.46× smaller than native BCF), 22× on UCI smart-meter telemetry, and −23% on a 2.96M-row NYC-taxi Parquet table that stays block-addressable.

Every row is a real public dataset, accepted only after the encoder's gate confirmed a byte-exact reconstruction. “AT-1 ratio” is times-smaller vs raw; the last column is the margin over the format that industry actually uses today.

Data type	AT-1 ratio	vs the format you use today
Server logs Apache access · SSH auth · HDFS	19–37×	1.1–1.5× smaller than xz-9
Genomics 1000-Genomes VCF, chr22	209×	2.46× smaller than native BCF (2.95× vs .vcf.gz)
Telemetry / IoT UCI smart-meter power	22×	1.9× smaller than xz-9 · ~3× vs Parquet-zstd
Neurophysiology / EEG PhysioNet CHB-MIT scalp EEG	5.1×	1.6× smaller than xz-9, lossless
Medical imaging (DICOM) pydicom uncompressed CR / MR / US	1.6–16×	1.07–3.27× smaller than xz-9 (uncompressed / RLE / colour)
Financial ticksQUERYABLE Binance BTCUSDT aggTrades	18.7×	~3× smaller than Parquet-zstd · 54× less query I/O
Event JSON / NDJSON GitHub Archive events	21×	1.15× smaller than xz-9
Database exportsQUERYABLE Mongo / Elasticsearch NDJSON	~28×	~2× smaller than xz-9, and queryable
Network telemetry / SOCQUERYABLE CTU-13 NetFlow (qcolumnar)	7.8×	≈ xz-9 on size — but queryable in place (31× less I/O) + per-host erasure
Lakehouse tabularQUERYABLE NYC-taxi, 2.96M rows, Parquet round-trip	−23%	23% smaller than Parquet-zstd (36.3 vs 47.0 MB), still block-addressable · 13× less query I/O
Map / geo OpenStreetMap, Luxembourg	14×	1.32× smaller than PBF, 1.47× vs xz

Every row is reproduced from a real public dataset, and every codec reports a byte-for-byte lossless check — we cite no result that isn't verified lossless.

Radical honesty

Most vendors tell you where they win. We publish where we lose.

Read that again — a vendor that tells you when notto buy is the one worth trusting. The advantage was never a ratio number that changes quarter to quarter; it's the capability (query + byte-exact) and the economics. So here is exactly where AT-1 is the wrong tool.

Decode speed is xz-class

The decoder is fast enough for archival reads, not for hot paths — ~2.2× xz and ~12× zstd in CPU/byte. AT-1 is a cold / archival tier, not your hot storage.

Already-compressed media

JPEG, H.264, and other entropy-saturated media gain ~nothing from any compressor, AT-1 included. We don't pretend otherwise.

Monochrome DICOM under JPEG-LS

On monochrome pixel data already under JPEG 2000 / JPEG-LS, the image-domain codec wins. AT-1's imaging win is uncompressed / RLE / color DICOM only.

Numeric-heavy tabular

On dense numeric SMART/sensor columns, a trained OpenZL graph edges us ~1.06× — at the cost of minutes of per-format training. We're zero-config.

High-entropy network data

On NetFlow and Zeek conn logs, the data is near its entropy floor; AT-1 ties xz and the non-inferiority fallback correctly kicks in. No structural win to claim.

“Best ratio everywhere”

Ratio leadership is contested and unprovable, so we never claim it. We claim only what the verification gate measured on real data, per domain.

Reproduce any number on this page — every codec prints LOSSLESS (byte-for-byte): True/False and we cite no result that prints False. Each dataset is public and named; the methodology is published open-access (Zenodo DOI 10.5281/zenodo.20811282), and the CLI that produces these figures is public (@tinyfiles/cli) with an open Apache-2.0 decoder.

Methodology & reproducibility

Every figure above is measured on a real, public dataset and accepted only after AT-1's encoder gate confirmed a byte-for-byte identical reconstruction. We publish no result that isn't verified lossless, and we state the boundaries we don't claim rather than hiding them.

Baselines

AT-1 is compared against the format each field actually uses: xz-9 and zstd for general data, Parquet+zstd-9 for tabular/lakehouse, and native domain formats (BCF for genomics, PBFfor OpenStreetMap). “AT-1 ratio” is times-smaller versus raw bytes; the comparison column is the margin over that baseline.

Verification protocol

Encode → decode → byte-compare against the original; a single mismatched byte rejects the result. For the queryable rows, query outputs (counts, sums, filtered scans) are checked identical to the same query run on the source data — the win is real only if the answers match exactly.

Dataset sources

Public, canonical providers for the benchmark data — verify on your own copy.

Reproduce it yourself

The CLI is public. Install it and run the audit on your own copy of any dataset above — it reports the ratio and the byte-exact lossless check, no account needed to read.

npm i -g @tinyfiles/cli
at1 audit ./your-dataset.csv     # ratio + byte-exact lossless verdict

Cite as: AT-1 (TinyFiles) compression benchmarks — tinyfiles.io/comparison. Numbers are pinned to repository sources and are not edited without one.