Run it again.
Get the same numbers.
A verdict is only worth something if you can check it. Load one here and your own browser checks its fingerprints. Then you can rerun the AI yourself and see the same numbers.
Load a result, pin the same model, feed it the same input and compare. WebCrypto verifies the Ed25519 signature and recomputes the payload, input and manifest SHA-256 here, so you do not have to trust this page either.
Choose a result
Or pick one of the newest:
Five steps from result to proof.
Five steps to rerun the AI on your own computer. The first time takes a while because it downloads the model.
The first run downloads the pinned checkpoint and verifies every file hash before it answers. After that, every replay runs on your own CPU with the pinned runtime.
Get the source and the manifest.
The source bundle carries the verifier, the chain adapter and the pinned runtime list. The manifest names the exact model revision and the SHA-256 of every weight file.
Install the pinned runtime.
Python 3.12 and Node 22. The Laya source at the pinned commit, and the same library versions as the publisher. The replay refuses to run on a version mismatch.
tar -xzf laya-source.tar.gz && npm ci
git clone https://github.com/NandhaKishorM/laya.git ~/work/laya
git -C ~/work/laya checkout SOURCE_COMMIT
python3 -m venv ~/work/laya/.venv
~/work/laya/.venv/bin/pip install -r ops/requirements-lock.txtFetch and verify the weights.
Downloads the open checkpoint and refuses to continue if any hash differs from the manifest.
~/work/laya/.venv/bin/python tools/prepare-model.pyReplay the result. One line.
Checks the Ed25519 signature, the input hash, the manifest hash and every runtime version, loads the verified weights in your own process, runs the exact signed input and prints every probability. It exits with an error on any nonzero probability difference.
~/work/laya/.venv/bin/python tools/reproduce-local.py RESULT_URLCompare in your browser.
Paste the JSON your replay printed, or load the file. The page compares each reported answer with the signed result, whose input hashes were verified above. Nothing leaves your browser.
Queue a server replay.
No computer handy? Ask the LAYA server to rerun it. It is a quick spot check; running it yourself is stronger proof.
POST /api/replay. The worker runs the exact signed input on the pinned model and reports max_probability_delta and identical. Useful as a spot check. Your own replay is the stronger proof.