Independent software evidence

Show the evidence behind your code before someone asks.

Independent evidence on whether AI coding agents and vendor software do what they claim.

Before you raise, sell to enterprise or accept a delivery, Barg Labs checks whether the evidence in your repository supports the technical claims being made about it, and gives the other side a record they can rerun. Fixed scope, fixed fee.

Book 20 minutes to pick the decision and the commit

Fixed feeA record anyone can rerunLimits statedNot a pentest or a certification

claims vs. record

Illustrative

Tests: 42 passed

JUnit record: 42 passed

Closes #117

Closing relation: #117

Files changed: 3

PR changed 4 files, 1 undeclared

Merged 14:02 UTC

1 of 4 claims not in the recordre-runnable

Independent consultancy · Vancouver and London

NVIDIA Inception member Claude for Startups member

Services

One question, one repository, a fixed fee.

The Repository Evidence Review is the core of the work. Instrument Review and fractional technical leadership use the same discipline: write the claim down, check it against the record, state the limits.

Repository Evidence Review

Available now · C$7,500 fixed fee

One repository. One named decision.

Raising soon? We check whether the evidence in your repository supports the technical claims in your deck. Selling to enterprise? We check what your security questionnaire says about how your code is built, tested and released. Either way you get a record the other side can rerun, and a two-page memo for whoever signs.

Decision, commit and scope agreed before we start; two to three weeks
Reproducible evidence review with Cejel, run offline in your environment
Adversarial reading by a senior engineer against your written claims
Design-partner places at C$3,500: five places, closing at five or on 30 November 2026
Not a penetration test, a certification or a guarantee

Scope and price

Instrument Review

Available now · from C$12,500

Your AI judge, measured by someone else.

People accept what your scanner, code reviewer or AI judge says without seeing inside it. We test it on cases where the right answer is already known.

Cases, predictions and scoring frozen before anything runs
A record a third party can re-run
You see it first and choose whether it is published, not what it says

How it works and pricing

Fractional technical leadership

Part-time, by arrangement

Senior technical judgement, a few days a week.

For firms adopting AI coding agents or modernising their integration estate. Houman spent 24 years building B2B integration platforms at GXS and OpenText: EDI and partner onboarding, secure file-transfer gateways, and migrations that do not break the partners on the other end.

Fixed days a week, contracted through Barg Labs
AI coding agents: what to let them change, and how to check what they did
Integration: EDI network migrations, OFTP2 and AS2 upgrades, partner onboarding backlogs

Talk about fractional work

Who it’s for

Someone is about to rely on your code.

An investor, a board, a customer’s security team, an acquirer or your own engineering lead. Boards and investors now carry the risk of software written or changed by AI; a review gives them independent evidence of what the code actually does, which they can check for themselves.

Founders before a round

Let diligence start from evidence.

Your investor’s technical diligence starts from a record they can rerun, not from your word.

See the review

Vendors selling to enterprise

Answer the security review with a record.

The build, test and release answers in your questionnaire, backed by evidence the buyer can rerun.

See the review

Acquirers and diligence teams

Check what you are buying.

A target’s code and its team’s account of what was tested, checked at letter of intent and at closing.

Watch the 2.5-minute walkthrough

Engineering leaders

Accept a delivery on evidence.

Work from contractors or coding agents, judged against the record rather than the handover report.

See the review

How a review runs

Two to three weeks, about two hours of your time.

Twenty minutes

Pick the decision and the commit, agree the scope and sign a mutual NDA.

Your claims, in your words

Read access to the frozen commit, and the claims you want checked, written before we read the code.

The review

The evidence review in your environment, the adversarial reading and a reproducibility rerun.

The readout

Report, evidence session, readout and a two-page memo. Two to three weeks end to end.

50% at kickoff. The rest only if the readout gives you a decision you can act on; an adverse or inconclusive result counts. Full scope and terms: barglabs.ai/review.

The method is open

The tools behind every review are free and open source.

So the other side can rerun the evidence without trusting us.

Cejel

Free and open source

Make repository evidence inspectable.

Examines engineering signals in a repository at a pinned commit and produces a reproducible report and trust certificate. Deterministic: no model call, no network, no signup. A trust certificate is not a security certification.

cejel.dev GitHub

Dunstan

Free and open source

Check what the agent says it delivered.

Checks the claims in a coding agent’s report (head commit, changed files, references, test and check counts) against the repository and CI record. Fails closed; checks declared facts, not whether the work meets the requirements.

About Dunstan GitHub

Research you can inspect

Read the method. Inspect the result.

Our published studies show what was tested, how it was measured and where the conclusions stop.

20 Sep 2026

Jev, by TypeSafe AI

A known-answer study of a commercial AI judge. Read the frozen method, results and limits before relying on its verdicts.

Read the experiment

29 Sep 2026

Laya, by Convai Innovations

An open-model study using the same cases, with the context-window limitation stated alongside the results.

Read the experiment

1 Oct 2026

skill-scanner 2.1.0, by Cisco AI Defense

A study of constructed credential-transfer instructions across agent skills. Methods, results and the study’s limits are published.

Read the experiment

These constructed studies do not estimate real-world error rates.

How we work

Evidence that survives someone else checking it.

Offline by design

Cejel runs where the code lives: no network, no model call, nothing sent to us.

Checkable records

Cejel reports and Dunstan records can be reproduced. Studies state their inputs, method and limits.

Limits stay visible

A missing source, an unverified claim and an observed contradiction mean different things. The record should say which.

Predictions frozen first

We write down what we expect before we run, and publish when we are wrong.

Company

Led by an engineer who spent 24 years building B2B integration platforms.

Barg Labs is led by Houman Azimi-Nejadi. Before founding it, he spent 24 years building enterprise B2B integration platforms at GXS and OpenText: EDI and XML mapping engines, trading-partner onboarding and secure file-transfer gateways.

Sareh Asadi Saravi is a director of Barg Labs and leads its accountancy.

Barg Labs is a small consultancy based in Vancouver and London. It drafts with AI and answers for every word, and the evidence it produces is built to be checked by someone else.

Barg Labs is a member of NVIDIA Inception, NVIDIA’s program for startups, and of Claude for Startups, Anthropic’s program for startups building with Claude.

Claude for Startups member

Bring the decision. We’ll pick the commit together.

Twenty minutes is enough to name the one question a review should answer, the commit it should answer it at, and whether a review is the right tool at all.

Book 20 minutes