A ANLAAgent-Native Lossless Archive 中文
Research preview · Open format · Apache-2.0

Let an AI plan the packing.
Do not let an AI rewrite the truth.

ANLA is an archive format for agents. A model may choose the chunking, the codec, the indexes and the order of work. A public, deterministic decoder still has to return every byte that was declared into the archive — with no model in the loop.

Open the workbench Read the specification
Local-firstThe workbench runs entirely in your tab: your files are read into memory and never sent anywhere. The page makes no requests of its own, and its content policy forbids outbound connections.
DeterministicSame input, same archive bytes — proven by two implementations in two languages producing identical output.
InspectableManifest, chunk map, codec choices and the planner's decision log are all readable JSON.
ExperimentalA research profile, frozen and tested — not yet somewhere to keep the only copy of anything.
Control-plane transition

Not another interface over ZIP.

Traditional formats assume a human picks the files and a level, and a fixed algorithm does the rest. ANLA turns the packing plan into a structured control surface an agent can produce, a validator can check and a writer can replay — while the format's invariants refuse information loss.

01

An AI planner, never an AI decoder

A model may analyse the tree and emit a packing plan. The writer validates that plan; the decoder restores without a model, ever.

Planner → Policy validator → Deterministic writer
02

Content addressing and deduplication

Every raw chunk is identified by the SHA-256 of its content, so identical content is stored once no matter how many files reference it, or which codec stored it.

chunk_id = SHA-256(raw bytes)
03

The intelligence layer is disposable

Search indexes, decision logs and model annotations live in a plane you can empty completely. What a decoder extracts does not change by one byte.

Decode(P, I) = Decode(P, ∅)
Two planes

Intelligence can be rebuilt. Original data cannot.

ANLA separates the preservation truth that must not be compromised from the AI-assisted data that may change freely. This is not a visual grouping; it is the format's trust boundary.

Preservation plane

Deterministic, required

Raw payload chunks, the chunk map, paths, metadata, content hashes and the snapshot root. Every conforming decoder must reach the same result from these bytes.

  • Raw content chunks
  • Canonical JSON manifest
  • SHA-256 integrity digests
  • Filesystem object model
Intelligence plane

Disposable, rebuildable

The packing plan a model produced, its decision log, full-text and semantic indexes, agent history. Replaceable, deletable, and never a precondition for extraction.

  • Packing plan
  • Planner decision log
  • Search and semantic indexes
  • Agent operation history
What is actually true today

The claims, and what backs each one.

A format's promises are worth what its tests are worth. These are the ones that run on every commit, on Linux, macOS and Windows.

T-XIM-3

Two implementations agree byte for byte

A Python writer and a JavaScript writer, written separately against the specification, produce identical archives for every reproducible fixture — not merely archives each other can read.

T-ORG-1

The first release's archive still verifies

The .anla file shipped with the original v0.1 browser build is checked into the repository as a compatibility vector, and both implementations still restore it exactly.

T-CHK · T-PTH · T-BMB

Corruption is refused, not tolerated

Wrong hash, wrong length, unknown codec, unsafe path, duplicate path, a compression bomb, an offset past the end of the file: each has a test asserting the decoder fails instead of guessing.

T-EXT-1

A filesystem that cannot restore an archive says so

Two paths differing only by case, or by Unicode normalization, are distinct here. Extracting onto a filesystem that folds them fails with both names — it never silently drops one.

Honest scope

What this profile does not do.

The whitepaper describes a larger format: BLAKE3, Zstandard, CBOR manifests, FastCDC, append-only snapshots, cross-platform metadata, encryption, signatures, parity. ANLA-MVP v0.1 implements the smallest subset that can be finished and verified end to end, and claims nothing else.

Implemented and tested

Single snapshot · ordinary files and directories · fixed-size chunking · Store and DEFLATE · SHA-256 chunk identity · canonical JSON manifest · cross-file deduplication · full round-trip verification · reproducible output · safe paths · resource limits · ZIP export.

Not implemented, not claimed

Symlinks · hard links · permissions and ACLs · extended attributes · alternate data streams · sparse files · FastCDC · Zstandard · BLAKE3 · encryption · signatures · parity · append-only snapshots · partial materialization.

⚠ This is a research profile. Do not make an ANLA archive the only copy of anything you cannot lose.
Get it

Read it, run it, or check it yourself.

Live workbenchBuild and verify a real .anla in your browser. No backend.HTML · no backend Standalone pageOne self-contained HTML file. Save it, open it offline, it still works.single file SpecificationThe normative definition of ANLA-MVP v0.1, byte layout included.EN PapersThe concept paper and the technical whitepaper behind the format.EN · 中文 Conformance vectorsFrozen archives and their hashes, for a third implementation to test against..anla + SHA256SUMS SourceBoth reference implementations, the test suite and the papers.Apache-2.0