Terms of Service

1. These terms, and who they apply to

These terms cover your use of Cahier, A single-tenant model system for agentic automations in regulated enterprises, delivered into the customer's private cloud or run on GPU capacity we control in Paris, with a per-token price, a stated context window, and a governance console that issues an evidence pack per model version.. They are a contract between Cahier Ltd, of 20 Farringdon Road, London EC1M 3HE, United Kingdom (“Cahier”, “we”) and the company that signs up (“Customer”, “you”).

The service is for businesses only, not consumers. By signing up you confirm that you are acting for your business and that you have the authority to commit your company.

The person who accepts these terms accepts them for the company, not for themselves. If you don’t have that authority, please don’t go ahead.

If documents disagree. This order applies: (1) a signed order form or agreement between us, (2) the data processing agreement, (3) these terms, (4) our Privacy Policy, (5) anything else on our website.

2. What the service is

Cahier supplies the model version, the runtime, the policy gate and the governance console. The platform's evaluation harness scores each version on the customer's held-out cases and re-runs that evaluation automatically every quarter; the console generates the evidence pack and holds the review thresholds, which the customer's own team sets. On the private-cloud row Cahier hosts the tenant in Paris. The printed row on cahier.ltd is the schedule to the order form.

3. What the service is not

These limits are deliberate, and they are part of what you are buying. They are not defects.

Not a foundation model of our own. We fine-tune open-weight base models named in the model card. We do not pretrain, and a customer who requires a vendor that owns its model from pretraining up should say so before signing.

Not a certification. We do not hold a SOC 2 Type II report or an ISO 27001 certificate and we do not represent either. The evidence pack and the annual penetration test summary are what we supply, and a procurement rule that requires a certificate is not satisfied by them.

Not an automation builder or a governance service. We do not build, host or maintain the customer's workflows, and we do not run its governance. The customer's Center of Excellence builds the automations that call the model and sets their thresholds in the console; the policy gate governs what the model may propose, not what the robot does with it.

Not a hosted API. There is no multi-tenant endpoint, no hyperscaler-marketplace listing and no option to run the model on a public cloud region of the customer's choosing. The two deployments are the customer's hardware and our Paris tenant.

4. What you need to do, and why it matters

What you get from the service depends a lot on things only you control. Please read this section carefully, because sections 9 and 10 build on it.

The held-out cases. You assemble the evaluation set, at least three hundred cases per workflow with the decisions your own people made, and you are responsible for having the right to use them. Without it there is no evaluation and no evidence pack.

The allow-list and the thresholds. You set the allow-list of actions per workflow and the confidence thresholds in the governance console. A proposal the gate lets through at your threshold is not a judgment by us that it was correct.

The reviewers. You staff the review queue and decide the cases that fall below threshold. A queue nobody works is a workflow that stops, by design.

Your hardware, on the self-managed row. You provision the GPUs, the network path to our registry and the outbound path for the signed meter, and you apply the runtime security patches we publish within 30 days of release.

Version approval. A new model version is not deployed to your tenant or your hardware until you have received its evidence pack and approved it in writing.

Lawful use. You may not send the model a document you are not entitled to process, and you may not configure the gate to execute an action with a legal effect on a person.

5. Getting started, and what is free

Before the order form is signed, the evaluation harness scores the current model version on up to five hundred of your held-out cases and the governance console issues that version's evidence pack. No card, no NDA on our side, and the pack is yours to put in the memo whether or not you sign.

6. Fees

Per million tokens. Self-managed: $1.50 per million input tokens and $6 per million output tokens, metered by the signed runtime on your hardware. Private cloud in Paris: $3 per million input tokens and $12 per million output tokens. Evaluation runs are not metered.

Monthly minimum. $6,000 per month on either row, which covers the platform and its governance console, the evidence pack for each version, the automatic quarterly re-evaluation and a named support contact. Token charges count against it, and a month above the minimum is billed at the token rates alone.

Price hold. The per-token rates on the dated row you signed hold for twelve months from the date on it. A change to the published row is announced ninety days before it takes effect and never applies inside a held term.

Custom contract. Air-gapped delivery, a dedicated fine-tune beyond the customer's own labeled set, or a contract on the customer's paper are priced in the order form, and the per-token rates in it are never higher than the published row.

What is never billed. The pre-signature evaluation, the evidence pack, the security review of the container image, meter exports and the quarterly re-evaluation carry no price of their own.

7. Delivery, availability and support

The evidence pack. Issued by the governance console within ten business days of any model version being made available, and before that version is deployed to the customer.

Availability, private-cloud row. 99.5% monthly on the model endpoint in the tenant, measured on responses. A month below it is credited against the next invoice.

The quarterly re-evaluation. Every quarter the evaluation harness re-runs on your held-out cases without a request from you, and the console adds the new report to the evidence pack.

Support. A named support contact per account, answering a question about the product, a security questionnaire or a board question about the model within five business days.

The meter. A signed monthly token statement per workflow, exportable at any time and matching the invoice line for line.

Support. Support is by email at [email protected]. We aim to reply within one business day. That is a target, not a guarantee.

8. The evidence pack, versions and the meter

The evidence pack. One per model version: the model card, the evaluation report on your held-out cases, the NIST AI RMF mapping, the sub-processor list and the change log against the previous version, kept for the life of the contract plus five years.

The labeled set. Every reviewer decision beside the proposal it corrected, yours, exportable as JSONL at any time and at termination.

The meter and the gate log. Token counts per workflow per day, and every proposal the allow-list dropped, with the workflow and the timestamp, exportable as CSV.

9. What we promise, and what we don’t

We promise to provide the service with reasonable skill and care, and that we have the right to provide it.

We do not warrant that a proposal is correct. The evaluation report states the measured error on your held-out cases for the version it describes, and that figure is the warranty: a version performing within it in production is performing as sold, and a reviewer decision is the control for the rest.

Beyond that we give no other warranty. As far as the law allows, we exclude all implied warranties, including merchantability, fitness for a particular purpose and non-infringement.

10. Liability

10.1 Neither of us limits liability for death or personal injury caused by negligence, for fraud or fraudulent misrepresentation, or for anything else the law says can’t be limited.

10.2 Losses neither of us covers. Neither of us is liable for lost profit, revenue, expected savings, business, goodwill or reputation, or for any indirect or consequential loss, however it arises.

10.3 Specific exclusions. We are not liable for an action your automation executed on a proposal the gate let through at a threshold you set, for a decision your reviewer took on a case in the queue, for a workflow that stopped because its queue was not staffed, or on the self-managed row for hardware, network or patching the customer controls.

10.4 Cap. Our aggregate liability is capped at the fees you paid us in the twelve months before the claim.

10.5 You accept that the limits in sections 9 and 10 are a fair way to share risk, given the price and your part under section 4, and that we could not offer the service at this price without them.

11. Confidentiality and data

Your documents, cases, held-out sets, labeled set and evaluation results are your confidential information. The evaluation harness and your tenant process them; a person at Cahier reads one only to answer a support request you made in writing, and we never use them to train anything but your own model version at your written instruction.

How we handle personal data is set out in our Privacy Policy and in the data processing agreement between us, which forms part of these terms. If the DPA and these terms disagree about personal data, the DPA wins.

12. Who owns what

You own your data, your held-out set, your labeled set and the outputs your automations receive. The fine-tuned weights of your model version are yours to run for the life of the contract and are handed over on the self-managed row at termination.

The runtime, the policy gate, the evidence pack format and the published row are ours, and the row is published so that you can hold us to it.

You may not use our name or logo in public, and we may not use yours, without written permission first.

13. How long this lasts, and how it ends

13.1 The contract starts when you sign up and runs until one of us ends it.

13.2 You can end it. Cancel at any time. It takes effect at the end of the current billing month, unless section 6 sets a minimum term.

13.3 We can end it. We can end it with 30 days’ notice. We can end it immediately if you seriously break these terms and don’t fix it within 14 days of being told, if you become insolvent, or if your use exposes us to legal risk.

13.4 What happens then. We stop delivering and stop billing. You keep everything already delivered to you, and your right to use it continues. We delete or return our working copies as the DPA says. Sections 9, 10, 11 and 12 continue to apply.

14. Changes to these terms

We may update these terms. A change that matters takes effect 30 days after we email you about it. If you don’t accept it, you can end the contract before then. Using the service after that date means you accept the change.

15. Contact

Cahier Ltd, 20 Farringdon Road, London EC1M 3HE, United Kingdom
[email protected]

← Back