A compression-native format that is also a language.
ATLAS puts the whole AT-1 moat into a single self-describing artifact: compressed, queryable, verified, tamper-pinpointing and erasable — all at once. 15 typed field kinds let a file describe its own shape, a parallel encoder runs ~57× faster, and it ties the best general codecs losslessly while winning on structured and genomics data.
Self-describing format
One file declares its own 15 typed field kinds — integers, deltas, timestamps, categoricals, blobs and more — so a reader knows the shape of the data without a side schema. The format is the language.
Queryable in place
Filter, project and aggregate against the typed fields without a full decompress — a query reads only the blocks its predicate touches, then reconstructs exact original values.
Verified & tamper-pinpointing
Integrity is built into the container. Alter a byte and verification fails at exactly the field and block that moved — provably the original, or provably not.
Erasable
Per-subject crypto-erasure lives in the same artifact: forget one person by destroying their key, while the rest of the file stays queryable and byte-stable.
--fast path — the format scales across cores.One file. One command surface.
Encode to a single ATLAS file, then query and verify it in place — no rehydrate, no separate schema, no managed service.
# compile a dataset into a single self-describing ATLAS file at1 atlas compile schema.atlas trades.csv trades.atlas # byte-exact, integrity-sealed at1 atlas compile schema.atlas trades.csv trades.atlas --fast # parallel-column encode, ~57× faster at1 atlas query trades.atlas price ">=" 24000 # zone-map skip + projection pushdown at1 atlas verify trades.atlas # -> integrity: PASS (or pinpoints the field)
Nobody else carries all five
On natural media we're not the smallest file on this table — that's not the claim. ATLAS is the only format that is also queryable, verified, tamper-pinpointing, and erasable, in the same artifact.
| Format | Compressed | Queryable in place | Verified | Tamper- pinpoint | Erasable |
|---|---|---|---|---|---|
gzip / zstd / xz small and portable, but opaque — any query means a full restore, no integrity, no erasure | |||||
Parquet + zstd queryable & compact, but cannot prove integrity, pinpoint tampering, or erase a subject | |||||
A database queryable & live, but a rented service — not a sealed, portable, self-describing file | partial | partial | partial | ||
ATLAS all five, in one self-describing format |
On structured data — telemetry, ticks, logs, tabular exports — and on genomics, ATLAS's typed fields beat general-purpose codecs while staying queryable and verified. That's the data where the typed layout has real structure to exploit.
On natural media — images, audio, video — the entropy is already modelled by xz / WebP-LL / FLAC / H.265. ATLAS does notclaim to beat them; it delegates losslessly (never-worse) and adds the queryable, verified, addressable layer on top. We say so plainly because it's true.
Built for
Structured telemetry & CSV/JSON · genomics & scientific tables · regulated archives that must stay searchable, provable and erasable · any pipeline that wants one self-describing file instead of a format, an index, a manifest and a service.
Decoding and verifying an ATLAS file is always free and needs no account; encoding and querying are metered against a connected account.