Data Processing Addendum

The terms on which Margin processes data on a customer’s behalf. This forms part of the agreement between Margin and the customer.

Version 1.5 · Last updated 17 October 2026

The short version

  • Margin’s meter receives no prompts and no model responses, by construction. We measure what a call cost. There is no field for its content.
  • Two things you start do send content to a model: the console’s AI tools and measurement runs. Margin routes that content only to the nine model hosts named in Section 8, or, for a measurement run on a project that chose “any zero-retention host” (Section 8), to a host on OpenRouter’s zero-data-retention list. Every one of them keeps none of it, and a host that keeps none of it cannot train on it.
  • In most deployments Margin receives no personal data at all. It reaches us only if you put it in a free-text label. Section 4 names every field where that can happen.
  • We never train on your data. Other customers only ever see model results combined across at least five customers, never yours alone, and you can opt out (Section 9).
  • You get it back or we delete it, within 30 days, on your instruction.

A summary, not the agreement. Sections 1 to 14 and the Annexes govern.

1. Scope and definitions

This Data Processing Addendum (“DPA”) forms part of the agreement between Margin and the customer (the “Agreement”) for the provision of the Margin service (the “Service”). Where this DPA and the Agreement conflict on the processing of personal data, this DPA governs.

“Customer Data” means the call and outcome records, and any associated labels, that Customer or Customer’s systems transmit to the Service; the content Customer submits to the Service’s AI tools, described in Section 3; and the model calls a measurement run makes, prompts and responses included.

“Personal Data”, “controller”, “processor”, “data subject”, “processing” and “personal data breach” have the meanings given in Applicable Data Protection Law.

“Applicable Data Protection Law” means all privacy and data protection laws applicable to the processing under this DPA, including the EU General Data Protection Regulation (“GDPR”), the UK GDPR and Data Protection Act 2018, the Swiss Federal Act on Data Protection, and US state privacy laws including the California Consumer Privacy Act as amended (“CCPA”).

“Sub-processor” means a third party engaged by Margin to process Customer Data.

2. Roles of the parties

Customer is the controller (or, where Customer is itself a processor for a third party, the processor) of Customer Data. Margin is the processor, and where Customer is a processor, Margin is the sub-processor.

Under the CCPA, Margin acts as a service provider. Margin does not sell or share Personal Data, does not retain, use or disclose it for any purpose other than performing the Service, and does not combine it with personal information received from another source except as permitted for a service provider. The Model Index under Section 9 is part of the Service; its statistics are aggregate, built only from model, task and cost figures, and contain no Personal Data. Margin certifies that it understands these restrictions and will comply with them.

Margin is the controller of the information described in its Privacy Policy: website visitors, people who email us, and account contacts. That is a separate relationship and this DPA does not govern it.

3. Subject matter and nature of the processing

The Service is a meter, not a proxy or a gateway. Customer’s systems call their model providers directly. Margin’s SDK measures a call after it completes and transmits a record of its economics. Margin is not in the request path. Through the meter, Margin does not receive, and has no technical means of receiving, the content of a request to a model or the content of a model’s response.

Measurement runs are the first exception, and Customer starts each one. When Customer connects a GitHub repository in the console and presses Execute, Margin clones that repository and runs the eval harness it contains in a throwaway container on Margin’s own server. Margin pays for the harness’s model calls through a key minted for that run alone, capped in dollars and revoked when the run ends. Those calls, prompts and responses included, are made from Margin’s server to OpenRouter. The container and the checkout are deleted when the run ends, though a daily server backup taken during a run can hold a copy until that backup expires, within 7 days. Margin keeps the run’s call and outcome records as it keeps any metered call. Nothing in a measurement run touches Customer’s production traffic.

Where Customer drops a folder into the console instead of connecting a repository, Margin stores that folder, encrypted with a key held only by Margin’s API and run server, against Customer’s project alone, so a later run can use it. Files that look like keys or credentials are left out before storage. Margin deletes the folder on request, or after 30 days with no activity in the project’s workspace; database backups can hold the encrypted copy for up to 35 days after that.

Every key Margin mints for a run is bound, before the run receives it, to a restriction that admits only the model hosts named in Section 8 (or, if Customer chose “any zero-retention host”, the hosts on OpenRouter’s zero-data-retention list, as described in Section 8) and only their zero-data-retention endpoints. A key that cannot be bound is revoked and the run does not start. Where Customer runs a measurement on its own model key, Margin mints no key, and Customer’s own account settings with that provider govern where those calls go.

The console’s AI tools are the second exception. When Customer asks the console a question, drafts evaluation criteria for an agent, or has the console grade an agent’s output, Margin sends the text that tool needs to a model: the question; the agent’s description, sample outputs and, where Customer has connected a repository, the source the draft reads; or the output being graded and its criteria. Margin sends it through OpenRouter, only to the model hosts named in Section 8, only to their zero-data-retention endpoints, and only to hosts that do not train on it. If no named host can serve a request, the request fails rather than going anywhere else. Margin keeps the answer and a cost record of the call, as it keeps any metered call.

Nature and purpose: recording, storing, aggregating and analysing the cost and outcome of Customer’s AI-agent operations, to produce cost-per-outcome measurements, comparisons between routes, and recommendations, and to make these available to Customer; and, unless Customer has opted out, contributing Customer’s comparison results to the Model Index under Section 9, which is part of the Service and is made available to every customer.

Duration: for the term of the Agreement, plus the deletion window in Section 11.

Categories of data subject: ordinarily none. Where Customer includes identifying information in the fields described in Section 4, the data subjects may be Customer’s personnel or Customer’s own end users.

Categories of Personal Data: ordinarily none. See Section 4.

Special categories: none. Customer must not transmit special categories of Personal Data to the Service. See Section 4.

4. What Margin actually receives

This section is unusual for a DPA and is included because it is the most useful thing Margin can tell a customer’s counsel.

4.1 The complete record

A call record consists of, and only of:

  • identifiers and labels chosen by Customer: workflow_id, session_id, prompt_id, operation, environment, team, harness, task_key;
  • the provider and model name, whether the call was a retry, its billing tier, and the provider’s stop reason;
  • token counts (input, output, cache read, cache write, reasoning), computed cost in USD, and latency in milliseconds;
  • structural identifiers linking calls in an agent tree: span_id, parent_span_id, orchestration_pattern, branch_width, stopping_rule, substrate;
  • an outcome record: whether a unit of work passed, an optional quality score and the method and grader that produced it, an optional link, and an optional list of failing grading criteria.

There is no field for prompt text, response text, message content, system instructions, tool arguments, or any part of a conversation with a model. The ingest API validates against a closed set of fields and rejects unrecognised ones. The call and outcome tables have no column capable of holding such content. Customer can verify this before installing anything: the SDK is published publicly on PyPI and npm.

4.2 The fields through which Personal Data could reach Margin

Several fields above accept free text that Customer chooses. If Customer places Personal Data in one, it reaches Margin. The fields are:

  • task_key: intended for a topic or task class, and the most likely to carry something incidental such as a ticket subject;
  • link on an outcome: a URL Customer supplies, which may itself contain identifiers;
  • every identifier and label named in the first bullet of Section 4.1, each of which could encode a person (for example, a username used as a team name);
  • the failing-criteria list on an outcome, which carries grading text Customer supplies.

Customer’s instruction and undertaking: Customer will not place Personal Data, and will not place special categories of Personal Data or data relating to children, into these fields, and will configure its instrumentation accordingly. Margin has no way to detect Personal Data placed in a free-text label, and processes these fields as opaque strings.

Where Customer nonetheless transmits Personal Data through these fields, it is Customer Data and every protection in this DPA applies to it.

5. Instructions

Margin processes Customer Data only on Customer’s documented instructions, which comprise this DPA, the Agreement, and Customer’s configuration and use of the Service, including any instruction given through the Service’s controls.

Margin will tell Customer if, in its opinion, an instruction infringes Applicable Data Protection Law, and may pause processing under that instruction until it is resolved. Margin will not process Customer Data for its own purposes, other than computing the Model Index under Section 9.

If Margin is required by law to process Customer Data beyond Customer’s instructions, it will tell Customer first unless the law forbids it, in which case Margin will challenge the requirement where there are lawful grounds to do so.

6. Security

Margin implements and maintains technical and organisational measures appropriate to the risk, including:

  • encryption of Customer Data in transit over public networks;
  • encryption at rest, as provided by the managed database and hosting platforms named in Section 8;
  • authentication of every ingest request by a per-project key, stored only as a SHA-256 hash so that a copy of the database does not yield a working key;
  • server-side attribution of every record to the authenticated project, so one customer cannot read or write another customer’s data;
  • access to production systems limited to personnel who need it;
  • logging and monitoring sufficient to detect and investigate errors and abuse.

What Margin does not claim: Margin holds no SOC 2 report, no ISO 27001 certification and no HIPAA attestation, and does not represent otherwise. Margin will not process protected health information and the Service must not be used for it.

Margin ensures that personnel authorised to process Customer Data are bound by confidentiality.

7. Session recording and the Customer console

Margin’s public website uses session replay for product analytics, as described in the Cookie Notice.

Session replay is disabled, or the rendered values fully masked, on any view of the Service that displays Customer Data. Margin will not enable a recording technology that could capture Customer Data rendered in the Service. Where Margin introduces such a view, this protection is applied before Customer Data is displayed in it.

8. Sub-processors

Customer gives general authorisation for Margin to engage sub-processors. Margin imposes data protection obligations on each that are no less protective than this DPA, and remains liable for their performance.

Margin’s sub-processors for Customer Data are:

Sub-processorPurposeLocation
Vercel Inc.Application and API hostingUnited States
Neon Inc.Managed Postgres databaseUnited States us-east-1
Dedale AI, Corp. (boat)Prepares a Customer-started measurement run: clones the repository and hands it to the per-run Vercel sandboxFrance (an OVH server); the company is in the United States
OpenRouter, Inc.Routes the console’s AI tools and Margin-funded measurement runs to one of the model hosts belowUnited States
Functional Software, Inc. (Sentry)Error monitoring; may incidentally receive labels appearing in an errorUnited States

The model hosts that may serve the console’s AI tools and a Margin-funded measurement run are the following, each through OpenRouter and each only on an endpoint that keeps no prompt or response:

Model hostLocation
Microsoft AzureUnited States
Google Vertex AIUnited States
Amazon BedrockUnited States
DeepInfraUnited States
Mistral AIFrance
MorphUnited States (added 17 October 2026)
InferenceNetUnited States (added 17 October 2026)
ParasailUnited States (added 17 October 2026)
CloudflareUnited States (added 17 October 2026)

Hosts you opt in to. In the console, a signed-in member can choose “any zero-retention host” for the project. The model calls of a Margin-funded measurement run may then be served by any host on OpenRouter’s public zero-data-retention list, not only the nine above. Choosing it is Customer’s general written authorisation of the hosts on that list as it changes over time, for those runs. Margin therefore gives no per-host notice when a host joins or leaves the list; the 30-day notice below applies to the nine named hosts and to every other sub-processor. Customer may switch back to the nine named hosts at any time in the console, and doing so also stops the keys of the project’s runs in progress, which then fail and say why. The default stays the nine named hosts.

How the list is applied. When Margin issues a run key it reads the list (cached for up to ten minutes) and limits the key to the hosts on it; later issues refresh that limit. A key lives for up to four hours, so a host removed from the list in that time can remain allowed on a key issued before. OpenRouter’s own zero-data-retention enforcement is set on every key and is applied to each call, so a host that stops being zero-retention is refused regardless. Every call goes only to an endpoint that retains nothing, and OpenRouter’s documentation states that providers that retain no data cannot train on it. Margin does not set a separate no-training option on the key, because OpenRouter’s guardrails have none. If the list cannot be read, no key is issued and the run does not start. Each call records the host that served it, the console shows it per run, and each choice is recorded with the member who made it and when. A run on your own model key follows the settings of that key instead.

PostHog is not a sub-processor of Customer Data. It is used for public-website analytics only and receives no call or outcome records.

Zoho is not a sub-processor of Customer Data. It sends the console sign-in email, and receives only the recipient’s address and the sign-in link.

No other model provider receives Customer Data. Margin also uses model providers for its own work, such as grading its own estate. Customer Data is not sent to them, and Customer’s production prompts never reach any model provider through Margin, because Margin never holds them.

Margin will give Customer at least 30 days’ notice before adding or replacing a sub-processor. Customer may object on reasonable data protection grounds within that period. If the parties cannot resolve the objection, Customer may terminate the affected part of the Service and receive a pro-rata refund of prepaid fees.

9. Artificial intelligence, training and aggregation

Margin does not use Customer Data to train, fine-tune, or otherwise develop any machine learning model, whether its own or a third party’s, and does not permit any sub-processor or model provider to do so.

The Model Index. Margin may combine the results of Customer’s model comparisons with those of other customers into aggregate statistics about how models perform (the “Model Index”). A Model Index statistic is built only from which model was compared with which, the kind of task, how many cases each model passed, and the cost per task. It never contains prompt or response content, evaluation content, workflow, project or customer names, or anything that identifies Customer or any individual, and a statistic is shown only when results from at least five customers contribute to it. Customer may opt out at any time by written notice, and Margin then excludes Customer’s results from any Model Index statistic computed after the notice. Model Index statistics are not Customer Data.

Apart from the Model Index, Margin does not pool Customer Data with another customer’s data, and does not use it to produce benchmarks, priors, comparisons or any other output made available to another customer.

Margin may use aggregate operational telemetry about the Service itself, such as request volumes and error rates, to run and improve the Service, provided it contains no Customer Data and identifies no customer.

Margin’s automated processing of Customer Data produces measurements and recommendations. It does not produce decisions with legal or similarly significant effects on any individual. Where the Service recommends a change to Customer’s configuration, the change is applied only with Customer’s ratification.

If Margin ever proposes to change this section, it will give notice under the same terms as a sub-processor change in Section 8, with the same right to object and terminate. A change that widens what a Model Index statistic may contain, or lowers the five-customer threshold, reduces Customer’s protection and requires Customer’s agreement under Section 14.

10. Data subject rights, breaches and assistance

Given what Margin receives, most data subject requests cannot be answered from Customer Data. Where a request nonetheless relates to it:

  • Margin will not respond directly to a data subject, other than to direct them to Customer, unless legally required;
  • Margin will notify Customer of a request it receives without undue delay, and in any event within 5 business days;
  • Margin will assist Customer in responding, by appropriate technical and organisational measures, taking account of the nature of the processing.

Personal data breach: Margin will notify Customer without undue delay and in any event within 72 hours of becoming aware of a personal data breach affecting Customer Data. The notice will describe the nature of the breach, the categories and approximate volume of data affected, the likely consequences, the measures taken or proposed, and a contact point. Margin will provide further information as it becomes available, and will not delay notification to complete its investigation.

Margin will provide reasonable assistance with data protection impact assessments and prior consultations with a supervisory authority, to the extent they relate to Margin’s processing and Customer cannot reasonably obtain the information itself.

11. Return and deletion

On termination or expiry of the Agreement, or at any time on Customer’s written request, Margin will delete or return Customer Data at Customer’s choice, and delete existing copies.

Margin will complete this within 30 days of the request or of termination, and will confirm completion in writing on request.

Margin may retain Customer Data where required by law, for as long as required and for no other purpose, and will keep it protected by this DPA for as long as it is held. Backups are overwritten on their ordinary cycle and are not restored except to recover the Service.

Customer’s own definitions remain Customer’s. Outcome definitions, quality bars and task-class priors belong to Customer and are exportable. Margin holds none of it hostage.

12. Audits

Margin will make available the information necessary to demonstrate compliance with this DPA, and will respond to Customer’s reasonable security questionnaires, once in any 12-month period and on a change that materially affects the processing.

Where Applicable Data Protection Law gives Customer a right of audit or inspection beyond that, Customer may audit on 30 days’ written notice, no more than once in any 12-month period unless required by a supervisory authority or following a personal data breach. An audit is conducted during business hours, without unreasonable disruption, subject to confidentiality, and at Customer’s expense.

Margin does not hold a third-party audit report and does not offer one in place of the above.

13. International transfers

Margin processes Customer Data in the United States, except that a measurement run is prepared on a server in France (section 8). Customer authorises those transfers.

A Customer that chooses “any zero-retention host” (section 8) authorises, for the measurement runs of that project, the transfer of the model calls’ content to the country of whichever listed host serves each call, which may be outside the United States and the EEA. The Standard Contractual Clauses and the UK Addendum below apply to those transfers.

Where Customer Data is subject to the GDPR and is transferred to a country without an adequacy decision, the parties incorporate the European Commission’s Standard Contractual Clauses (Decision 2021/914), completed as set out in Annex 2, as follows:

  • Module Two (controller to processor) applies where Customer is a controller; Module Three (processor to processor) applies where Customer is a processor.
  • Clause 7 (docking) applies. Clause 9: Option 2, general written authorisation, with the 30-day notice period in Section 8. Clause 11: the optional independent dispute resolution language does not apply. Clause 17: the law of Ireland. Clause 18(b): the courts of Ireland.
  • Annex I, II and III of the Clauses are populated by Sections 3, 4, 6 and 8 of this DPA and by Annex 1 below.

United Kingdom: the UK International Data Transfer Addendum (version B1.0) applies to transfers subject to the UK GDPR, with Tables 1 to 3 populated by this DPA, Table 4 selecting neither party, and the Clauses read as modified by the Addendum.

Switzerland: for transfers subject to Swiss law, references in the Clauses to the GDPR are read as the Swiss FADP, the competent authority is the Federal Data Protection and Information Commissioner, and the Clauses protect data of legal entities until Swiss law no longer requires it.

If a transfer mechanism is invalidated, the parties will work in good faith to adopt a valid alternative, and Margin will suspend the affected transfer in the meantime if no alternative is available.

14. General

Each party complies with Applicable Data Protection Law in respect of its own processing. Customer warrants that it has a lawful basis for the processing it instructs, and that it has given any notice and obtained any consent required.

Liability under this DPA is subject to the limitations in the Agreement, except where Applicable Data Protection Law does not permit it.

If a provision is held invalid, the rest stands. This DPA is governed by the law of the Agreement, except that the Standard Contractual Clauses are governed as stated in Section 13.

Margin may update this DPA to reflect a change in law, a new sub-processor, or a change in the Service, on notice under Section 8. A change that reduces Customer’s protection requires Customer’s agreement.

Annex 1: Details of processing

ItemDetail
Data exporterCustomer, as identified in the Agreement. Role: controller, or processor where acting for a third party.
Data importerMargin. Role: processor. Contact: subh@trymargin.io
Categories of data subjectOrdinarily none. Where Customer includes identifiers in free-text labels: Customer personnel, or Customer’s end users.
Categories of Personal DataOrdinarily none. Where present: identifiers or free text in the label fields listed in Section 4.2, and anything Customer includes in content it submits to the AI tools or a measurement run (Section 3).
Special categoriesNone. Customer must not transmit them (Section 4.2).
FrequencyContinuous, on each metered call and recorded outcome.
Nature and purposeAs described in Section 3.
DurationTerm of the Agreement plus the deletion window in Section 11.
Sub-processorsAs listed in Section 8, for the purposes and durations stated there.
Competent supervisory authorityFor SCC purposes, the supervisory authority of the EEA member state in which Customer’s EU representative is established, or otherwise as determined under Clause 13.

Annex 2: Technical and organisational measures

The measures in Section 6 constitute Annex II of the Standard Contractual Clauses. They are: encryption in transit and at rest; per-project key authentication with keys stored only as hashes; server-side project attribution preventing cross-customer access; least-privilege access to production; confidentiality obligations on personnel; logging and monitoring; deletion on instruction within 30 days; and the recording restriction in Section 7.

Margin does not hold a SOC 2 report, ISO 27001 certification, or HIPAA attestation, and Annex II should be read as the complete statement of its measures.

How this is accepted

This DPA is incorporated into the Agreement by reference and takes effect on the earlier of the Agreement’s effective date and Customer’s first transmission of Customer Data to the Service. No signature is required for it to apply.

If Customer requires a countersigned copy, or a copy with the Standard Contractual Clauses executed as a standalone document, email subh@trymargin.io and we will provide one.

Data protection contact: subh@trymargin.io

Sub-processor notices, breach notifications and audit requests all go to this address.