DON'T TRUST
THE OUTPUT.
TRACE IT.
Run a transparent model. Capture its inputs and arithmetic. Hand the record to someone else to check. Every claim stays attached to its evidence.
Run a trace.
Read every step.
These small built-in models calculate in your browser. The lab records the exact inputs, coefficients, intermediate sum and output. No file leaves your device.
Take the record.
Challenge the claim.
Paste a record or load a JSON file. This verifier recomputes the local model and SHA-256 digest, then explains exactly what matched or failed.
From opaque answer
to inspectable path.
The experimental loop is complete locally. A future EZKL adapter would replace the record digest with proof artifacts and a verification key.
Set the inputs.
Three transparent sample models; no uploaded ONNX is executed in this build.
calculation.
Weighted terms and activation are computed in the browser and shown plainly.
evidence.
Export a portable JSON record with a SHA-256 digest over canonical fields.
independently.
Import any record and compare model, output, terms and digest with a fresh run.
Inspired by the open ZKML workflow.
EZKL describes ONNX model export, circuit generation, proof generation and verification. Witness Trace is an independent front-end experiment. It does not include EZKL code, a prover, an ONNX runtime, an EVM verifier, or a token contract. Repository license metadata was unavailable at review; no source code was copied.
A ledger of
local attempts.
Saved records stay in this browser. Open one to inspect it or export a copy. Clearing browser storage removes them.