The model
Name, version, weight digest, quantization, and license.
OPEN-SOURCE PROJECT · MODEL RECEIPT
Know what
actually ran.
ModelReceipt is an open-source format for recording the model, software, hardware, evaluation, and evidence behind an AI deployment.
WHY IT EXISTS
“We used Model X” leaves out the important parts.
The same model can behave differently when its version, runtime, settings, hardware, tools, or evaluation changes. Those details often live in separate documents and systems.
ModelReceipt gives engineers, security teams, buyers, and leaders one shared answer to a basic question: what did we run, and what evidence supports it?
WHAT A RECEIPT CONTAINS
One record for the full deployment.
A receipt captures enough detail to identify a deployment and compare it with another one.
The software
Inference runtime, container, settings, and application or agent harness.
The machine
Hardware, accelerator count, memory, driver, and hosting environment.
The evaluation
Test set, evaluation setup, results, logs, and supporting evidence.
HOW TO READ THE EVIDENCE
A receipt says how well the claim is supported.
A signature can prove who published a record. It cannot prove that the deployment worked.
The publisher describes the deployment.
The receipt links to configuration, logs, digests, or other evidence.
Another party ran the declared stack and linked its result.
PROJECT PRINCIPLES
Made to live with the work.
Portable. Keep the receipt with the project, deployment, or evaluation that produced it.
Inspectible. Use plain structured data that people and tools can review.
Evidence-aware. Separate a reported claim from one backed by artifacts or reproduced by another party.