Terms of Service
The agreement governing your use of the three.dev platform.
These Terms of Service (the "Terms") are between Round3 AI, Inc., a Delaware corporation ("three.dev", "we", "us") and the entity that accepts them ("Customer", "you"). They govern your access to and use of the three.dev platform and related services (the "Service").
If you accept these Terms on behalf of a company, you represent that you are authorized to bind that company. If you and three.dev have signed an order form, master services agreement, or similar document that references these Terms, that document controls where it conflicts.
1. What the Service is
three.dev is an evaluation and experimentation platform for teams shipping AI features. The Service lets you define scenarios and judges, run offline experiments against candidate prompts and models, confirm winners in live experiments, review results, and, where agreed, build specialized models for your own use.
We may change how the Service works. We will not make a change that materially reduces the security of the Service or materially reduces its core functionality during a paid term without giving you notice and a right to terminate for the unused portion of your term.
2. Accounts and acceptable use
You are responsible for your users, their credentials, and everything done under your account. Notify us promptly at legal@round3.ai if you suspect unauthorized access.
You will not: (a) use the Service in violation of law; (b) upload data you do not have the right to process, including personal data you have no lawful basis to send us; (c) reverse engineer the Service, except where that restriction is unenforceable under applicable law; (d) resell or provide the Service to third parties except as expressly permitted; (e) use the Service to develop a product or service that competes with the Service and that you make commercially available to third parties; or (f) attempt to circumvent usage limits, rate limits, or security controls.
For clarity, Section 2(e) does not restrict you from building, running, or operating evaluation, testing, or observability tooling for your own internal use, including tooling that overlaps in function with the Service.
You are responsible for obtaining any consents, notices, and lawful bases required for the data you send us, including from your end users. We provide the processing infrastructure. You decide what is lawful to process.
3. Fees
Fees, usage limits, and term are set out in your order form or the pricing page applicable when you subscribed. Unless stated otherwise: fees are non-refundable, invoices are due within thirty days, and usage above your committed volume is billed at the then-current overage rate. You are responsible for taxes other than taxes on our income.
We may suspend the Service for non-payment after giving you written notice and at least ten days to cure.
Service levels, support response targets, and any uptime commitment are set out in your order form. These Terms do not create an availability commitment on their own.
3.1 Trials and beta features. We may offer free trials, evaluation access, or features labeled beta, preview, or early access. These are provided as is, without warranty of any kind, without any availability commitment, and may be changed, limited, or discontinued at any time. Our total liability for them is capped at one hundred United States dollars. We may delete data in a trial or beta environment on expiry. All other terms, including Sections 5, 6, 7, and 8, apply to trials and beta features in full.
4. Term and termination
These Terms run for your subscription term and any renewals. Either party may terminate for material breach that stays uncured thirty days after written notice.
4.1 Customer Traffic and Customer Materials. On termination your access ends, and we will delete or return Customer Traffic and Customer Materials within 30 days of your written request, or within 30 days of termination if you make no request. Deletion is subject to backup cycles and any retention we are legally required to keep.
4.2 Evaluation Signal. This section states plainly what happens to Evaluation Signal, so there is no gap between what you own and what we return.
(a) Export. Before termination, and for 30 days after, you may export the Evaluation Signal held in your account through the Service or, if that is not available, on written request to us.
(b) Deletion. On your written request we will delete the Evaluation Signal held in your account on the same timetable as Section 4.1, and it is deleted with your account in any event.
(c) What is not deleted, and why. Two things survive: Evaluation Signal that has already been de-identified under Section 6.5(a), and any model, parameter, statistic, benchmark, or derived work that was trained on or calculated from Evaluation Signal before deletion. These are our property under Section 6.2, contain no Customer Traffic or Customer Materials, and cannot be unwound. We are not obliged to retrain, delete, or roll back any model or aggregate already built. Your ownership under Section 6.1 is ownership of the Evaluation Signal itself, not of our derived works.
(d) Backups. Copies in backups are deleted on the ordinary backup rotation schedule and remain subject to Section 7 until deleted.
4.3 Survival. Sections 4.2, 5, 6, 7, 8, 9, 10, 11, and 12 survive termination.
5. Data definitions
These definitions are used consistently across these Terms, the Data Processing Addendum, and the Privacy Policy. Where any of those documents appear to conflict on what data may be used for, this Section 5 and Section 6 control.
5.1 "Customer Traffic" means the production inputs and outputs of your AI features that you send to or generate through the Service, including end user prompts and messages, model responses, historical traffic replayed in offline experiments, and traffic captured during live experiments. Customer Traffic includes any personal data contained in it.
5.2 "Customer Materials" means the content you author and provide to the Service, specifically your prompts, your system prompts, your scenario and dataset definitions, and any model you provide to us or that we build for you under an order form. Customer Materials do not include Customer Criteria or Platform Artifacts.
5.3 "Customer Criteria" means the evaluation criteria, quality standards, rubric statements, and preferences you supply to the Service to tell it what a good output looks like for your use case. Customer Criteria describe your standard of quality. They are not your product's prompts and they are not Customer Traffic.
5.4 "Platform Artifacts" means the evaluation apparatus the Service creates, whether generated automatically, generated with your input, or configured by us, including judge prompts and their scaffolding, judge output schemas, reason-code vocabularies, default and suggested criteria libraries, rubric structures, scenario generation logic, cluster taxonomies, and the statistical methods the Service applies. Where the Service composes a judge prompt from Customer Criteria, the composed judge prompt is a Platform Artifact and the underlying Customer Criteria remain Customer Criteria.
5.5 "Evaluation Signal" means the data the Service generates about Customer Traffic, Customer Materials, and Customer Criteria in the course of running evaluations and experiments, specifically:
(a) AI judge outputs: scores, verdicts, labels, classifications, confidence values, reason codes, and other structured reasoning fields; (b) human assessments: labels, ratings, accept or reject decisions, selected reason codes, and annotations submitted by your users through the Service; (c) alignment data: agreement and disagreement rates between judges and human assessors, calibration curves, inter-rater statistics, and drift measures; (d) failure analysis: cluster assignments, cluster identifiers, cluster sizes, cluster taxonomy labels, the structural relationships between clusters, and the titles and descriptions generated to describe a cluster; (e) experiment results: effect sizes, confidence intervals, probability to be best, sample sizes, and win and loss records between variants; (f) operational telemetry: latency, throughput, token counts, error rates, and cost per variant, model, and run.
5.6 "Structured Field" means either of the following within Evaluation Signal:
(a) Controlled values. A field whose value is drawn from a fixed or Service-controlled set, including numeric scores, booleans, enumerated labels, classifications, confidence values, cluster identifiers, cluster sizes, statistical measures, telemetry values, cluster taxonomy labels drawn from a Service-controlled failure taxonomy, and reason codes drawn from a Service-controlled vocabulary. The failure taxonomy and the reason-code vocabulary are closed, versioned sets we maintain as Platform Artifacts. A value not drawn from one of those sets is not a controlled value, whatever it is named.
(b) Aggregate characterizations. A field that characterizes what a group of evaluated items has in common, rather than the content of any individual item, including the label, title, and description we generate for a failure cluster. These describe a recurring pattern across many requests. A characterization that describes, quotes, or paraphrases an individual item is not an aggregate characterization, and Section 6.4A applies to it.
5.7 "Free-Text Field" means a field within Evaluation Signal containing free-form natural language generated or written at evaluation time about an individual evaluated item, including judge rationale prose and free-form assessor notes. A Free-Text Field is not a Structured Field.
5.8 "Service Data" means data about your use of the Service that is not Customer Traffic, Customer Materials, Customer Criteria, or Evaluation Signal, including account records, user identifiers, configuration, application logs, usage metrics, billing records, and support communications.
6. Ownership and data rights
6.1 Your ownership. As between the parties, you own all right, title, and interest in Customer Traffic, Customer Materials, Customer Criteria, and the Evaluation Signal held in your account. Nothing in these Terms transfers ownership of any of them to us. Your ownership of Evaluation Signal does not extend to our derived works as described in Section 6.2, and is subject to the deletion and survival rules in Section 4.2.
6.2 Our ownership. We own the Service, our platform models, our software, all Platform Artifacts, and everything we create that is not Customer Traffic, Customer Materials, Customer Criteria, or Evaluation Signal. We own Service Data, subject to the Privacy Policy. We also own all models, model weights, parameters, aggregates, statistics, benchmarks, reason-code vocabularies, and other derived works we create using Evaluation Signal, Customer Criteria, or Platform Artifacts under the licenses in Section 6.3, provided they contain no Customer Traffic and no Customer Materials and comply with Section 6.5.
For clarity, a judge the Service generates for your use case is a Platform Artifact that we own, and you have a license to use it as part of the Service for the duration of your subscription. Your Customer Criteria remain yours.
6.3 What we may do with each tier.
(a) Customer Traffic and Customer Materials: your account only. We process Customer Traffic and Customer Materials solely to provide, secure, and support the Service for you, and on your documented instructions. We do not use Customer Traffic or Customer Materials to train, fine-tune, or improve any model, feature, benchmark, or dataset that is made available to any other customer or to the public. The only exception is the metadata carve-out in Section 6.4.
(b) Evaluation Signal: platform-wide license. You grant us a worldwide, non-exclusive, royalty-free license to use Evaluation Signal to operate, analyze, benchmark, secure, train, fine-tune, evaluate, and improve the Service and our platform models, including models and features made available to other customers. This license survives termination as to Evaluation Signal generated before termination, and continues for any model or system already trained on it. Section 6.5 limits how this license may be exercised.
(c) Customer Criteria: platform-wide license. You grant us the same license as in Section 6.3(b) over Customer Criteria, for the same purposes. Customer Criteria describe what good looks like, not what your users said, so they are not gated by Section 6.5(a). They remain subject to every other limit in Section 6.5, and to Section 6.4A if you paste Customer Traffic into them.
(d) Platform Artifacts: ours outright. We may use Platform Artifacts, including judges the Service generates for your use case, without restriction under this Section 6, subject only to Section 6.5(b) and Section 7.
(e) Service Data: ordinary operation. We may use Service Data to operate, secure, support, bill for, and improve the Service, and to produce aggregated usage statistics.
6.4 Metadata carve-out. Notwithstanding Section 6.3(a), we may use structural metadata derived from Customer Materials for the purposes in Section 6.3(b), provided that metadata contains no Customer Traffic, no Customer Materials content, and no personal data. Structural metadata means values such as token and character counts, number of variants or scenarios, model identifiers, sampling parameters, tool-call counts, schema shape, and run configuration flags. It does not include prompt text, system prompt text, scenario content, or any excerpt of any of them.
6.4A Contamination rule. If you place Customer Traffic, or personal data, into a field that would otherwise be Customer Criteria or a Structured Field, that field is treated as Customer Traffic under Section 6.3(a) until the contaminating content is removed. Pasting an example end user message into a criterion does not convert that message into licensed data. This rule applies automatically and does not require notice from either party.
6.5 Limits on the platform-wide licenses. The licenses in Sections 6.3(b) and 6.3(c) are subject to all of the following, each of which is a material term:
(a) Free-Text Fields are gated. Structured Fields are not.
Structured Fields are not subject to this paragraph, on either of two grounds.
Controlled values, under Section 5.6(a), are licensed under Section 6.3(b) with no removal step because a value drawn from a set we control cannot carry your content. Scores, enumerated labels, classifications, confidence values, cluster identifiers and sizes, statistics, telemetry, taxonomy labels, and reason codes are all in this group, whether emitted by an AI judge, assigned by the clustering system, or selected by a human assessor.
Aggregate characterizations, under Section 5.6(b), are licensed with no removal step because they describe what a group of requests has in common rather than the content of any one request, which is the same basis on which aggregated statistics are permitted under Section 6.5(c). We generate them only over clusters meeting a minimum size, and we screen them so they do not reproduce material from any individual item. Should one nonetheless contain your Customer Traffic, Section 6.4A applies and we treat it as Customer Traffic until the content is removed.
Free-Text Fields are excluded from the license in Section 6.3(b) by default. A Free-Text Field becomes eligible only once we have removed from it every excerpt of Customer Traffic, every excerpt of Customer Materials, and all personal data. Unless and until that removal has been performed on a given field, we treat that field as Customer Traffic under Section 6.3(a) and it is not used platform-wide. This is a standing restriction on our use, not a representation that any particular removal process is in place at any given time.
(b) Never-reproduce covenant. We will not use Evaluation Signal in any way that enables the Service, any platform model, or any of our outputs to reproduce your Customer Traffic or Customer Materials, or to identify you or your end users.
(c) No attributed benchmarks. We will not publish, market, or disclose Evaluation Signal in a form that identifies you, your product, or your end users, without your prior written consent. Aggregated and de-identified statistics across many customers are permitted.
(d) No competitive targeting. We will not use your Evaluation Signal to build a feature, dataset, or model for the specific purpose of serving a named competitor of yours at your expense.
(e) Security parity. Evaluation Signal used under Section 6.3(b) remains subject to the security measures in the Data Processing Addendum.
(f) No re-identification. We will not attempt, and will not permit any subprocessor or other recipient to attempt, to re-identify Evaluation Signal that has been de-identified under Section 6.5(a), or to link it back to you, your end users, or any individual. We impose this commitment by contract on any recipient of de-identified Evaluation Signal and maintain internal controls against re-identification. We will not de-identify Evaluation Signal and then treat the result as falling outside these Terms.
6.6 Opt-out. The Service improves by measuring quality across many evaluations. This section states exactly what that means for you, and what you can decline.
(a) What is not optional, and why. Structured Fields and Customer Criteria are licensed to us under Sections 6.3(b) and 6.3(c) as a condition of the Service, and are not subject to opt-out. This is not a claim on your data. Controlled values under Section 5.6(a), such as a score, an enumerated label, a confidence value, a cluster identifier, or a reason code from our own vocabulary, are drawn from sets we control and cannot contain your Customer Traffic, your Customer Materials, or personal data. Aggregate characterizations under Section 5.6(b), meaning the label, title, and description of a failure cluster, describe a pattern across many requests rather than the content of any one of them, and are subject to the safeguards in Section 6.5(a). Customer Criteria describe your standard of quality, not what any of your users said. What we aggregate is our own measurement of quality, not your content.
(b) What you can opt out of. You may opt out of the platform-wide use of Free-Text Fields, meaning judge rationale prose and free-form assessor notes, by written notice to legal@round3.ai or by the setting in your account, if available. On opt-out, those fields are permanently excluded from the license in Section 6.3(b) for you, including after any removal under Section 6.5(a). This is a real restriction: without it, a Free-Text Field becomes eligible for platform-wide use once stripped, and with it, it never does.
(c) Effect of an opt-out. An opt-out operates going forward only. We are not required to retrain or unwind any model, aggregate, or statistic already built before it took effect. An opt-out does not restrict our use of any data to operate, secure, and support the Service for you, and does not affect our ownership or use of Platform Artifacts under Section 6.3(d).
(d) Negotiated variations. An order form may vary this Section 6.6. Where it does, the order form controls, and a broader opt-out may affect pricing.
6.7 Feedback. If you give us feedback or suggestions about the Service, we may use them without restriction or obligation to you.
6.8 Customer models. Any model we build for you under an order form is Customer Materials and is subject to Section 6.3(a). We do not use it or its weights to serve other customers unless your order form says otherwise in writing.
7. Confidentiality
Each party will protect the other's confidential information with at least reasonable care and will not disclose it except to personnel and contractors who need it and are bound by comparable obligations. Customer Traffic and Customer Materials are your confidential information. These obligations do not apply to information that is public through no fault of the receiving party, independently developed, or rightfully received from a third party, and do not prevent disclosure required by law with reasonable prior notice where permitted.
8. Privacy, security, and AI providers
8.1 Data protection. Where we process personal data on your behalf, the Data
Processing Addendum at /dpa applies and is incorporated into these Terms.
8.2 Subprocessors, including model providers. We use subprocessors to deliver
the Service. The current list is at /subprocessors, and you can ask us to notify
you of changes to it.
There are two distinct inference paths, and they differ in whose account pays.
(a) Your evaluations and experiments: your account. When you run an evaluation or experiment against a third-party model, we send Customer Traffic to that provider using the API credentials you supply, against your own account with that provider. Your agreement with that provider governs their use and retention of that traffic, including any no-training terms you have negotiated. We do not substitute our own account for yours. We store your credentials encrypted per organization.
(b) Our own evaluation machinery: our account. Some parts of the Service run
on our own provider account rather than yours, including AI judges we
operate and the labeling of failure clusters. Those calls send content
derived from your Customer Traffic to the providers identified at
/subprocessors as receiving it on our account. For those providers we are
responsible for the contractual terms, and they are our subprocessors under
Section 8 of the Data Processing Addendum.
Specialized models we build and serve for you run on our own infrastructure and are not a third-party provider.
8.3 Security. We maintain the technical and organizational measures described in the Data Processing Addendum. We will notify you without undue delay and within 72 hours of confirming a personal data breach affecting your data.
9. Warranties and disclaimers
9.1 Mutual. Each party warrants it has the authority to enter these Terms.
9.2 Ours. We warrant that the Service will perform materially as described in our documentation during your subscription term, and that we will not materially decrease the security protections we apply to Customer Traffic during that term.
9.3 AI output disclaimer. This section is material to the pricing of the Service. The Service produces evaluation results, judge verdicts, statistical estimates, cluster analyses, and model outputs. These are probabilistic estimates, not guarantees. We do not warrant that any judge verdict, score, experiment result, probability estimate, cluster, or model output is accurate, complete, unbiased, reproducible, fit for any particular purpose, or non-infringing. You are responsible for deciding what to ship, for validating results before relying on them, and for any consequences of decisions you make using the Service. Do not use the Service as the sole basis for a decision that affects legal rights, safety, health, credit, employment, or similar consequential outcomes without independent human review.
9.4 Third-party models. The Service routes requests to third-party model providers. We do not control and do not warrant their availability, output, pricing, or terms. Their outages are not our breach.
9.5 Disclaimer. Except as stated in this Section 9, the Service is provided "as is". We disclaim all other warranties, express or implied, including merchantability, fitness for a particular purpose, accuracy, and non-infringement, to the maximum extent permitted by law.
10. Limitation of liability
10.1 Exclusion. Neither party is liable for indirect, incidental, special, consequential, exemplary, or punitive damages, or for lost profits, lost revenue, lost data, or business interruption, even if advised of the possibility.
10.2 Cap. Each party's total aggregate liability arising out of or related to these Terms is limited to the fees you paid or owed us in the twelve months before the event giving rise to the claim.
10.3 Super cap. For the following claims, the cap in Section 10.2 is replaced by a cap of three times the fees you paid or owed us in the twelve months before the event giving rise to the claim, and Section 10.1 still applies:
(a) our breach of Section 6.3(a), meaning use of Customer Traffic or Customer Materials to improve anything for another customer or the public; (b) our breach of Section 6.5(b), the never-reproduce covenant, or Section 6.5(f), no re-identification; (c) either party's breach of Section 7 (Confidentiality); (d) either party's indemnification obligations under Section 11.
10.4 Unlimited. Neither Section 10.1, 10.2, nor 10.3 limits: your payment obligations; either party's liability for fraud, fraudulent misrepresentation, gross negligence, or willful misconduct; or any liability that cannot be limited under applicable law.
10.5 Data protection claims. Liability for breaches of the Data Processing Addendum is subject to Sections 10.2 and 10.3 unless the order form says otherwise, except where applicable data protection law prohibits that limit.
11. Indemnification
11.1 By us. We will defend you against a third-party claim that the Service, as provided by us and used in accordance with these Terms, infringes that party's intellectual property rights, and will pay resulting damages finally awarded or agreed in settlement. This does not cover claims arising from Customer Traffic, Customer Materials, third-party model outputs, your modifications, or your combination of the Service with anything we did not supply.
11.2 By you. You will defend us against a third-party claim arising from Customer Traffic or Customer Materials, including a claim that we had no lawful basis to process them, and will pay resulting damages finally awarded or agreed in settlement.
11.3 Process. The indemnified party must give prompt notice, let the indemnifying party control the defense, and cooperate reasonably.
12. General
12.1 Governing law and venue. These Terms are governed by the laws of Delaware, without regard to conflict of laws rules. The parties consent to exclusive jurisdiction and venue in the state and federal courts of Delaware.
12.2 Changes to these Terms. We may update these Terms. For material changes, we will give at least thirty days' notice by email or in-product. Changes take effect at your next renewal, or on the stated date for month-to-month subscriptions. If a material change is unacceptable to you, you may terminate before it takes effect and receive a pro-rata refund of prepaid, unused fees.
12.3 Publicity. We will not use your name or logo publicly without your prior written consent. Consent given for a specific case study or quote covers that use only.
12.4 Assignment. Neither party may assign these Terms without the other's consent, except to a successor in a merger, acquisition, or sale of substantially all assets. In any such transfer involving us, the successor is bound by Sections 6.3(a) and 6.5.
12.5 Force majeure. Neither party is liable for delay or failure caused by events beyond its reasonable control, excluding payment obligations.
12.6 Entire agreement. These Terms, the Data Processing Addendum, the subprocessor list, and any order form are the entire agreement and supersede prior discussions. Your purchase order terms do not apply. If any provision is unenforceable, the rest stays in effect.
12.7 Notices. Legal notices to us go to legal@round3.ai and 285 W Wieuca Rd NE STE 67913, Atlanta, GA 30342. Notices to you go to the admin email on your account.
Questions about these Terms: legal@round3.ai