Bench Addendum to the Terms of Service and Privacy Policy
Effective 28 October 2026. This addendum applies to Bench (MCP and REST), a document-extraction benchmark service, and supplements the Terms of Service and Privacy Policy. Where this addendum and the main text differ, this addendum governs for Bench only.
1. Scope and BYOK
Bench measures how a model endpoint you choose extracts fields from our versioned reference sets. You supply an OpenAI-compatible endpoint URL, model name and authorisation header value ("your endpoint"). Inference runs on your endpoint, under your key and at your cost. Bench does not run a model of yours and does not send documents of yours to your endpoint.
You confirm that you are permitted to use your endpoint and key for this purpose and that you accept your provider's terms. Bench is self-serve: you start runs and quotas are enforced automatically.
2. Data we send and receive
For every run, Bench sends documents from its reference set to your endpoint, one request per document, with a fixed prompt owned by the set. Reference sets consist of publicly filed documents and sets prepared by us from public documents only; each set's manifest describes its source, licence and PII status. Bench receives the model responses and scores them.
Before a run begins, you must expressly confirm that the reference documents will be sent to the endpoint you supplied. We retain the endpoint URL (with credentials removed) and model name with the run record so runs can be told apart. We do not include your authorisation header value in run metadata or reports.
3. Retention
Reports are available for seven days after a run. Raw model responses for DEV items are retained for up to 30 days, whether those DEV items were run in a DEV-only run or in a full run. Raw model responses for blind holdout items are never retained. Holdout reports contain aggregates only and do not expose individual holdout cells or reference values.
After those periods, the relevant records are deleted from the active database. This statement does not claim secure erasure or deletion from backups.
We store Bench credentials to authenticate requests. Credential retrieval after checkout is one-time; save your credentials when they are shown. A page reload or a lost response cannot recover that retrieval.
Request metadata and billing data follow the main Privacy Policy.
4. Plans, billing and cancellation
Prices are in USD and billed by Stripe. Free smoke ($0): one run per customer, five documents, no accuracy score. Dev pack ($199, one-time): up to 30 DEV runs within 30 days of purchase; unused entitlement expires. Blind set ($490/month): 20 full runs per billing period. Team ($1,990/month): 100 full runs per billing period, shared by up to five credentials. Unused runs do not roll over. Custom sets (from $12,000) are provided under a separate written agreement.
The Dev pack is non-refundable once an access credential has been issued, except where consumer law requires otherwise.
You may cancel a monthly subscription at any time through the billing portal. Cancellation takes effect at the end of the current billing period; the subscription remains usable until then and no partial-period refund is made. Access credentials linked to the subscription are revoked when the subscription ends.
5. Privacy and transfers
The reference corpus contains public documents only. When you provide an endpoint and expressly confirm before a run, you instruct Bench to send the reference corpus to that endpoint, including when the endpoint is outside the EU. For Bench reference-corpus transfers, this is an exception to any EU-only transfer restriction in the main Privacy Policy.
Bench processes runs and stores records on servers in Germany (EU). Inference runs on the endpoint you choose, including outside the EU under the reference-corpus transfer exception above.
6. Liability
Bench reports are measurements of your endpoint on our reference set on the day of the run. They are not a guarantee of your model's accuracy on your own documents or in production, not a certification, and not legal, financial or compliance advice. Public filings are suitable for comparing models, prompts and routes; they are not proof of accuracy on your document class. The limitation of liability in the main Terms applies; in particular we are not liable for charges your endpoint provider bills you for benchmark traffic, or for decisions taken on the basis of a report.