Unorganized Machines

Prototype · In testing · Paper in preparation

The file that carries
its arithmetic.

A photograph, a checkpoint, or a log, stored as the recipe that rebuilds it, byte for byte.

Our storage prototype carries the exact arithmetic needed to reconstruct supported data. Keep the recipe and its required ingredients. A compatible machine rebuilds the original, byte for byte, and a ledger records the check.

Become a design partner
01 / A receipt, not a copyA measured prototype example
The recipe164,538 bytes
Rules4,328 characters
hashed
PriorNamed by SHA-256
PayloadThe remaining data
needed to reconstruct

Original

1,048,576 bytes

Deleted. Reconstructed. Checked.

1.2553 bits per original byte. xz: 2.3102 on this workload. The recipe names its prior; that shared input must remain available and belongs in the total storage cost.

02 / The return journey

Delete the copy. Keep a way back.

First reconstruct and verify. Then remove the original copy, retaining the recipe and every dependency needed to reopen it.

  1. i.

    Original

    1 MB

  2. ii.

    Prior + rules

    Named by hash

  3. iii.

    Recipe

    164 KB

    Original may be deleted
  4. iv.

    Resolver

    Supported
    ARM · x86 · L4

  5. v.

    Original

    1 MB
    Same bytes

Recipe’s promiseLedger · opened → verifiedRecovered digest
The reconstruction is checked against the recorded digest. A mismatch is a failed recovery, not a near-enough result. Sizes are rounded in this diagram.

03 / What makes it portable

The rules travel
with the file.

The same arithmetic, written down.

The rules specify how to compute the answer, including the order and rounding that affect its final bytes. They are part of the recipe, with an identity of their own.

A resolver follows those rules instead of relying on two machines to happen to do the same thing. In the file shown above, the rules occupy 4,328 characters.

A machine that can read the recipe.

The resolver checks the rules, finds the required prior by its hash, and combines it with the payload to reconstruct the original. The prior must still exist; its name is not a substitute for its bytes.

Order-free integer matrix multiplication lets supported CPU and GPU implementations agree. Our tests include ARM, x86, and NVIDIA L4 hardware; support for other devices and file formats is still being developed.

A recovery leaves a record.

The ledger links recovery records in a hash chain. Background checks reopen recipes, compare the rebuilt bytes with the promise, and record whether they agree. These checks can be counted month by month.

In a prototype test, eight deleted files were reopened under that chain. The record is useful because the recovery can be repeated: keep the ingredients, rerun the resolver, and check the answer.

04 / What has run

The return, in bytes.

These are separate prototype tests. Each answers a different part of the question.

Across machines
The same 1,127 bytes on three architectures.
After deletion
Eight files deleted and reopened under a hash chain.
Without model weights
6.6 MB reconstructed from a 115 KB recipe.
Recorded prototype output
VERIFIED: 1048576 bytes from a 164538-byte file, the promise 4fb5fea9… kept.

Prototype test figures and output reported by Unorganized Machines. The new manuscript is in preparation; these are separate from the results in our earlier papers.

The present limit: this codec decodes at about 1.4 KB/s on one tested GPU.

An invitation

Bring us something worth recovering.

We are seeking design partners with real data: labs retaining checkpoints, and GPU or storage providers exploring the cost of recomputation. Tell us what you keep, how much arrives each month, and which hardware you use.

Become a design partner