PRICING v4 · 2026-10-09

One fixed-price first engagement. The spec stays free.

Start with one real delivery. The Acceptance Evidence Pack is the record of one acceptance decision: the plan you and your client approved before testing, the signed acceptance report, and what you agreed about later changes, linked to the system version and test runs and sealed with independent timestamps. Anyone reviewing the decision later can see who approved what, and which result rests on which plan, and can check it offline. The open standard underneath (PRML, Community Specification License 1.0, four MIT reference implementations) is free and stays free. The Pack is one piece of evidence inside your compliance file; it does not replace your risk management, your auditors, or their opinion.

Evaluation Evidence Pack criteria locked

Sample · PREP-Eval worked example · illustrative data, real receipts

Acceptance criteria

metric: instruction_set_accuracy
pass if: ≥ 0.85
test set: cphos-en-sample-v2 e4c01b52…7bb0
seed: 7

Timestamped commitment

f38e1825ae4f7ba7e6469026f054503bb6dfe0f0f3bdcb70d8af0053261de8b7

✓ registry receipt 22 Sep 2026 14:57:40 UTC
✓ RFC 3161 timestamp independent authority
✓ Rekor transparency log #120937972

Revision history

v1 2282ae19… pre-registered
v2 f38e1825… test set re-translated · threshold unchanged

Shows the bar existed, unchanged, by that time. Not that the score is right. Verify independently →

What the client or reviewer receives. The acceptance criteria, a timestamped commitment, the revision history, and a link anyone can use to check them without access to your systems. Built from the real receipts of our PREP-Eval worked example · see a full sample pack.

First engagement · Acceptance Evidence Pack

€4,500

fixed · first draft within 2 working days

For a team delivering an AI system to a client, and the person on the client side who signs acceptance. Also for evaluation and assurance firms that test against agreed criteria.

One AI system, one acceptance round, up to three measurable criteria
An acceptance plan approved by both sides and sealed before testing: the criteria, your written rationale, the named approvers
The criteria as a canonical record (PRML manifest) with an independent timestamp (RFC 3161 and a public transparency log)
A dated acceptance report, signed by your responsible persons with your own e-signature tool: the result per criterion, pass or fail as it came out, and the accept or reject decision
Links between plan, system version, test runs and report, with any missing reference pointed out
The change and re-review conditions you agreed at acceptance, recorded (we do not monitor them)
A record of the checks run on the bundle, and those not run, with the commands to repeat each one
An offline verification bundle both sides keep and can check without us
Deliverables are signed with a qualified electronic signature and carry a qualified electronic time stamp (eIDAS Art. 25(2) and 41(2))
A one-page map of where this evidence sits in your compliance file (AI Act, ISO/IEC 42001) and what it does not cover

You write the content and make the decision; we draft from your existing documents, link, seal and point out what is missing. Arranging a document is not vouching for its content. You need a working test process and a named decision owner on each side. The first draft arrives within 2 working days of receiving your documents (if testing has not started, that is the acceptance plan); the final record follows once both sides have reviewed and signed. See a completed example record.

Request an Evidence Pack

Developer

€0

Free, forever, no account

For researchers, individual engineers, and teams happy to self-host. Everything needed to implement and verify the PRML standard independently, on your own infrastructure.

Full PRML v0.1 spec + frozen v0.2 RFC (Community Specification License 1.0)
Reference implementations in Python, JS, Go, Rust
21 byte-equivalent conformance vectors
mlflow-falsify plugin (PyPI)
Self-host the registry anywhere
Public manifests on registry.falsify.dev
Community support via GitHub Issues

Read the spec

What the Pack does not do. It does not guarantee a passing test, and it does not certify that the result is correct. It shows which criteria were agreed, that they existed no later than the recorded time and have not changed since, and what happened when they were applied. If the scope grows (more systems, CI wiring, several acceptance rounds), we quote it in writing before any work starts.

In return for the fixed price we ask for one feedback conversation at the end. If the Pack is used in a real acceptance decision, we would like to name you and quote one sentence; a written case study is welcome but is never a condition.

Need more?

Three further options exist. None is needed to start.

Acceptance test preparation. If your team needs capacity on a specific delivery, we propose the criteria, build and run a reproducible test harness in a narrow, defined scope, and report the results and gaps. Scoped and quoted in writing; the acceptance decision stays yours.
Pro, managed registry. A tenanted private registry designed for ten-year retention (Article 18 floor), content-addressed with full export, re-verifiable without us. Early access, price on request.
Sprint engagements. Written deliverables such as an audit memo or an Annex IV section: Audit Review, Full Sprint and Enterprise Engagement, each scoped and quoted in writing. Each includes twelve months of Pro access.

See the Sprint tiers → Ask about Pro

Why the spec is free

The open spec answers one question: can a third party detect that the acceptance criteria were changed after the result was seen? That question is free to ask, free to answer, and free to verify. We do not gate it.

The Acceptance Evidence Pack is where we do the work with you: agreeing the record with both sides, linking it to the real test runs, and handing over something your client and your auditor can check. Pro is for teams that want that record kept in a private registry for ten years. A Sprint is for teams that need a signed written deliverable on top.

FAQ

Why no checkout? Why everything by email?

Every engagement starts with a scoping conversation, because the deliverable is written to your claims and your pipeline rather than sold off a shelf. Once a contract is agreed we invoice from Falsify OÜ (reg. 17574308, Estonia) in EUR, payable by SEPA transfer (clients outside the SEPA area can pay through any service that delivers EUR by SEPA).

Why is almost everything written?

Because the written artifact is the deliverable: detailed answers, sample Evidence Packs, and full technical detail arrive in writing so they stay citable, dateable, and re-readable when your auditor reopens the file. Alongside the written record, scheduled video calls are available at every stage (scoping, technical walkthrough, delivery Q&A), in English, at your preference; anything agreed on a call is confirmed in writing the same day.

What's the difference between Pro and a Sprint engagement?

Pro is an ongoing subscription to a hosted private registry. You log in, your team commits manifests, the platform keeps them for ten years. It is self-serve infrastructure, not an engagement: nothing about it is scoped to a person's calendar.

A Sprint is a one-off written engagement. The founder reviews your evaluation claims, writes a memo or a full technical-documentation package (Art. 11 / Annex IV), and signs the output for use with your auditors and, where your conformity route involves one, a notified body. Sprint pricing reflects scope, not volume. Every Sprint includes twelve months of Pro platform access bundled, so you do not pay twice.

How is this different from a GRC platform like Holistic AI or Credo AI?

It is not the same product. A GRC platform is a top-down dashboard that maps your AI portfolio to a hundred regulatory controls. Falsify is a bottom-up technical primitive for evaluation-claim integrity within Annex IV technical documentation (Art. 11) and Article 15: pre-committing the evaluation manifest cryptographically so retroactive edits become detectable. The two compose well. Many organisations adopting PRML will already have a GRC tool. Falsify is the thing the GRC tool's evidence field points at.

What does "early access" actually mean?

It means three things. First, Pro prices are locked in for your first contract term once signed. Second, we will work with you to shape the v0.3 spec and the Evidence Pack format, in writing. Third, early-access support target: response within two business days. Contractual service levels are set as part of a Sprint engagement.

What if the 2 December 2027 (Annex III) deadline gets pushed?

Article 99 penalties bind regardless of the politics around enforcement timing. Locked evaluation records are also sensible engineering practice independent of the AI Act. Our pricing is not premised on the deadline holding. If you want to wait and see, the Developer tier costs nothing and carries everything needed to implement and verify the standard yourself; what it does not carry is managed retention, organisational controls, or support, which is what production teams add later.

Can a notified body or audit firm use Falsify on behalf of their clients?

Yes. We are explicitly designing the Evidence Pack format with notified bodies in mind. If you are a notified body, an audit firm, or an insurance underwriter and want a technical briefing on how PRML maps to AI Act Article 9(8), Annex IV point 2(g) and ISO/IEC 42001 controls A.6.2.4 (verification and validation) and A.6.2.7 (technical documentation), email [email protected] with subject "PRML technical submission" and we will send the pass/fail vectors and schemas the same day.

Open core or open spec?

Open spec. The PRML specification, the four reference implementations, the conformance vectors, the JSON Schema, and the MLflow plugin are all permissively licensed and will stay that way. The Pro subscription adds managed infrastructure around the open specification, not proprietary format features. A Sprint engagement adds a signed written deliverable and the review behind it, not proprietary format features. Using PRML never requires a Falsify account.

Not sure where to start? Email [email protected] with a short description of the AI system you are delivering and who signs acceptance on the client side. We reply within 24 hours, in writing.