Terms of service

1. Who we are

Agentic Evals is a service of Agentic Thinking Ltd, a company registered in England and Wales, company number 17152930. Our address is Office 9, Sycamore Trading Estate, Squires Gate, Blackpool FY4 3RL, United Kingdom. We are registered with the Information Commissioner's Office. Contact us at sales@agenticthinking.uk.

2. Who these terms are for

Our packages are sold to businesses. By buying, you confirm you are buying for a business, including as a founder of a product you are building, and not as a consumer.

3. What you buy

You buy one fixed-price package from the buy page: a standard pack or an advanced pack, each with or without one re-test. The price on the buy page is the whole price. No VAT is added. The price shown is the total you pay. You pay in full, in advance, through Stripe. The fee does not depend on the result.

Plans, proposals and test specifications

A tailored evaluation plan or proposal describes our proposed approach, boundary, candidate cases, evidence aims, prerequisites and limits. It may identify feasibility questions to resolve. A plan or proposal is a planning deliverable; it is not the agreed test specification, an executable test kit or a completed evaluation. Candidate coverage in a proposal is not confirmed coverage.

The test specification is a separate written record of the agreed target and commit, evaluation boundary, selected cases and expected behaviours, relevant configuration and state, evidence and reproduction deliverables, exclusions and delivery date. We agree it with you before testing begins. A proposal does not itself authorise testing or establish a finding about your product.

Preparing and supplying a tailored plan or proposal after your order is substantive paid preparatory work within the package. It is not conditional on a test specification subsequently being agreed. If the engagement is cancelled after that work starts, section 8 applies. Work carried out before your order does not retrospectively become paid work.

4. What an evaluation is, and what it is not

An evaluation tests your product at one fixed commit, in our isolated environment, against the checks and cases in your package. The report describes what we observed under those conditions, with its limits stated. Severity is our judgement.

An evaluation is not a certification, a penetration test, an audit opinion, a guarantee that your product is secure, or an endorsement. A result holds only for the commit and configuration we tested. You must not present a report, or any part of it, as a certification or endorsement by us.

5. What you agree to do

• Only submit software you own or are authorised to have tested, and tell us anything we must not do during testing. You will cover any third-party claim that says you were not authorised.
• You authorise us to run, probe, attack and attempt to subvert the submitted software in our environment. You confirm testing will not reach any live system or any third party's system unless that party has given written consent you have shown us, and you will set up the intake (stubs, mocks, test endpoints) accordingly. You will cover any loss or claim we suffer because this confirmation is wrong.
• Give us a complete, runnable intake: the repository link, the exact commit, how to run it, and one action your guardrails must block. That is what "complete runnable intake" means in these terms.
• Never send us production secrets, live service credentials or customer personal data. Read-only repository tokens scoped to the evaluation are fine; revoke them when we confirm deletion. We plant our own fake secrets for testing.
• Implement any fixes yourself. We do not change your code.

6. Delivery

We aim to start testing within 2 hours of payment and receipt of a complete runnable intake, during UK business hours (9am to 6pm, Monday to Friday, excluding bank holidays). If both arrive outside those hours, we aim to start within 2 hours of the next business-hours opening. The baseline report arrives within 5 working days after payment and receipt of a complete runnable intake. For advanced packs we confirm the test scope and delivery date in writing from your complete intake before testing starts. Paid preparatory work described in section 8 may begin before the detailed test specification is agreed; testing itself starts only after scope is confirmed. If we cannot agree scope within 10 working days, either party may cancel. Refunds for business cancellations follow section 8, including where preparatory work has started. If we determine that we cannot provide the purchased package, we cancel and refund in full. If we miss a delivery date by more than 5 working days, you may cancel for a full refund.

If your product cannot run in the agreed environment, we contact you to agree the next step. If we cannot agree one, we refund the package in full.

7. Re-tests

A package with a re-test includes one re-test of the original agreed cases and controls at one updated commit, submitted within 60 days of your first report. New workflows, material rewrites and further rounds are outside the package. A re-test not used within 60 days expires. Because we delete your source code when we deliver the report, a re-test needs fresh read access to the updated commit.

8. Cancellation and refunds

You can cancel by emailing sales@agenticthinking.uk. For business purchases, we refund the fee in full if you cancel before substantive work on your paid engagement has started.

Work starts when we begin substantive work on your paid engagement. This includes intake or repository review, evaluation planning and preparation of a tailored plan or proposal, the separate development of a test specification, test design, synthetic fixtures, adapters, environment setup, rig preparation, feasibility checks and test execution. Taking payment alone, and work carried out before you placed your order, do not count as the start of paid work.

Once work has started and before results are delivered, any refund for a business cancellation is at our reasonable discretion. We consider the work already carried out, materials prepared or supplied, non-recoverable costs and the reason for cancellation. Starting preparatory work can therefore affect your refund even if testing has not yet begun; there is no automatic 50% refund. Once results are delivered, the fee is not refundable, subject to section 6 and any rights that cannot lawfully be excluded.

The full-refund provisions in section 6 still apply if we cannot provide the purchased package, cannot agree a way to run your product in the agreed environment, or miss a delivery date by more than 5 working days. If we have no complete intake within 30 days of payment, we may cancel; any refund follows this section, taking account of whether substantive work has started.

If you are nonetheless a consumer, the business discretionary-refund policy does not replace your statutory cancellation rights. If you expressly request work to start during the 14-day cancellation period and then cancel before the service is fully performed, any charge must be proportionate to the service actually supplied, as permitted by law. Loss of a statutory cancellation right on full performance requires the consent and acknowledgement required by law. Nothing in these terms limits your statutory rights.

9. Your code and data

• We never use your code to train models. Where AI is used, we use frontier models configured not to use your input for training. We list the providers on request.
• We delete our working copies of your source code when we deliver your report, and confirm it in writing.
• We keep the report and its evidence files for 12 months for comparison and re-tests, then delete them unless you ask us to keep them longer. Evidence can contain source excerpts and command output. You can ask us to delete it sooner.
• For private code you accept our confidentiality agreement (NDA), already signed by us, before you share access.
• Where your code or logs contain personal data, we process it as your processor, only to perform the evaluation and on your instructions in these terms. We keep it confidential and secure, use only the sub-processors we list on request (AI model and hosting providers) and tell you before changing them, make sure everyone who handles it is bound by confidentiality, impose the same data protection obligations on our sub-processors, tell you without undue delay if we become aware of a personal data breach affecting it, help you with data subject requests where relevant, delete it as set out in these terms, and give you the information reasonably needed to show compliance. Transfers outside the UK use the UK International Data Transfer Addendum or an adequacy decision. Subject matter and duration: the evaluation, for the evaluation period and the retention periods above. Data: whatever personal data is contained in the material you submit (typically names, email addresses and usernames of your staff or users).

10. Confidentiality and publication

Your report is private, whether your code is private or open source. We publish it, or name you as a customer, only with your written agreement. We may use anonymised, aggregate statistics across evaluations, such as how many products failed a check, in our publishing, marketing and research. We keep them in groups large enough that they never identify you or your product.

If we find a vulnerability in someone else's software while testing yours, we tell you first and consider your views. After at least 14 days, or longer if you show us a contract with that vendor that requires it, we may report it to that vendor under Agentic Thinking's published disclosure policy, without naming you or including your confidential information unless you agree in writing.

11. Ownership

You keep all rights in your code. The report is yours to use, including the fix prompts and regression tests in it. We keep the rights in our methods, tools, probes and templates, and may use what we learn from evaluations to improve them. You may not republish our methods, probes or tooling as a product or service.

12. Liability

Nothing in these terms limits liability for death or personal injury caused by negligence, for fraud, or for anything else the law does not allow us to limit.

Subject to that, our total liability for a package is limited to the fee you paid for that package. We are not liable for loss of profit, revenue, data or goodwill, for indirect or consequential loss, or for decisions you or anyone else make based on a report. For private code, liability for breach of confidentiality is set by the confidentiality agreement instead.

13. Changes and law

We may update these terms. The version in force when you paid applies to your package. These terms are governed by the law of England and Wales, and its courts have exclusive jurisdiction. Nobody other than you and us may enforce these terms (Contracts (Rights of Third Parties) Act 1999).

We may decline or stop an evaluation and refund pro rata if the software appears unlawful, is designed to attack, defraud or harass third parties, or breaches our AI providers' acceptable use policies. We do not sell to persons or countries under UK sanctions.