Verify, inside the Compute Optimization Engine

Proof that your kernel is right and fast on the real chip

AI writes the code. Verify runs it on every chip you target and returns a signed pass or fail. Verify is the proving product inside the Compute Optimization Engine, which runs your whole environment from firmware to kernels, on any chip.

Start free Read the docs

Try it

Watch a kernel earn its pass

Pick a scenario. Each one plays a real Verify flow from intake to a signed proof pack.

computelab verify

# Pick a scenario to start.

Works with how you build

One verdict for agents and for people

For AI coding labs and agents

Call Verify from the loop

MCP server and REST API with a cost cap per task.
A machine readable pass or fail your agent can act on.
Batch submission from CI for thousands of candidates.
RL training environments built on the same judge.

For engineering teams

A pass in your pull request

One command, computelab fix, or a GitHub label.
A CI gate that warns first and blocks only when you choose.
A "Verified by Computelab.co" result page to share.
Full IP on delivered code. You click merge.

The risk

What goes wrong without a neutral judge

Code that passes a quick test and drifts on real shapes

Verify bounds the error against an FP64 answer across at least 64 shapes, with 20 percent held back. A kernel that looks right only on easy inputs fails.

A port that runs, but slower than the library

A pass requires 90 percent of the vendor library's median time on the real chip. Kernel gain and application gain are shown separately in your metric.

Fast on NVIDIA, wrong on AMD

Verify runs every target chip on real hardware. A qualified dispatcher then picks the proved variant per shape and chip.

A driver release breaks last month's speedup

Verify rechecks on every ROCm, CUDA, Neuron, JAX, PyTorch and firmware release. The console assigns every regression a named owner.

Results that change from run to run

1000 identical runs prove repeatability. The same input gives the same answer.

A claim your customer has to take on trust

Every pass comes as an Ed25519 signed proof pack with an offline verify script. Anyone can check it offline, on their own machine.

What Verify provides

Everything a buyer asks for, in one pack

Real chip runs

Every check runs on the hardware you ship on, from data center GPUs to phones.

Bounded accuracy

FP64 reference, 64 shapes or more. Whole model ports keep 99.9 percent of the task score, 99 percent for FP8 and INT8.

Speed you can quote

Timed against the vendor library median, with energy per accepted result beside it.

Signed proof pack

Ed25519 signature, offline verify script, shareable result page and safety standard evidence for MISRA, AUTOSAR and ISO.

How it works

Three steps from candidate to signed pass

mcp

# 1. Agree the acceptance contract
tool verify.contract { "target": "MI300X", "reference": "fp64", "max_cost_usd": 5 }

# 2. Submit the candidate kernel
tool verify.submit { "contract": "ct_8f21", "kernel": "attn_fwd.hip" }

# 3. Read the signed result
tool verify.result { "job": "vj_3c90" }
PASS  signed ed25519  proof: vj_3c90.pack

terminal

# 1. Connect your repo
$ computelab login

# 2. Verify a kernel on the target chip
$ computelab verify kernels/attn_fwd.hip --chip MI300X

# 3. Check the proof pack offline
$ computelab proof check vj_3c90.pack
signature valid  PASS