Evals for RadixArk

Open-source model access and selection

Explaining that the platform exposes different open source models users can access and run, without asserting a specific unlisted model roster.

Prompt and query response flow

Describing how a user submits queries, inputs, or prompts and receives model outputs through the Platform.

Uploaded content as model input

Handling of text, documents, images, photos, videos, and audio files submitted through the Services as inputs.

Registration gate for platform use

Stating that certain portions of the Platform require registering an account with name, email, and other requested information.

SGLang as the serving foundation

Attributing inference serving to SGLang and explaining its role as the engine underneath RadixArk's inference offering.

Continued investment posture

Communicating RadixArk's stated commitment to keep investing in SGLang rather than forking away from it.

Production reliability framing

Explaining the claim that SGLang serves as the performance and reliability foundation for production AI applications, without inventing benchmarks.

Openness of the engine

Distinguishing the open engine from the managed layer RadixArk ships on top of it.

Miles scope and purpose

Identifying Miles as RadixArk's open-source framework for large-scale RL post-training.

Training vs. inference separation

Correctly routing a question to Miles (RL training) or SGLang (inference serving) rather than conflating the two.

Open-source availability

Stating that Miles is open source and part of the cores RadixArk builds on and ships around.

Rigor-to-RL positioning

Explaining the stated analogy that Miles brings to RL the rigor modern serving engines brought to inference.

Managed layer over open cores

Explaining that RadixArk ships managed infrastructure and tooling on top of SGLang and Miles.

Audience routing

Matching an inquiry to the four stated audiences — individual developers, startups, enterprises, research labs — without promising unlisted tiers.

Rebuild-avoidance value proposition

Articulating the stated problem of every new AI lab rebuilding schedulers, compilers, serving engines, and training pipelines.

Engineering philosophy claims

Representing the first-principles, elegance-and-throughput, frontier-scale-reliability stance as stated positioning.

Agreement scope and acceptance

That the Agreement covers radixark.com, platform.radixark.com, and linked products, and that access or use signifies agreement.

Binding an organization

The representation required when accepting on behalf of a company, business, or other legal entity.

Arbitration and legal rights notice

Surfacing that the Terms contain an agreement to arbitrate and other information about legal rights, remedies, and obligations.

Payment and billing information

That payment method, billing address, bank account, and transaction history are collected and processed by a third-party payment processor.

Business-customer carve-out

That the Privacy Policy does not apply to personal information processed on behalf of business customers, which is governed by customer agreements.

Categories of information collected

Account and contact information, user content, company and workspace information, payment information, feedback, and support communications.

Covered surfaces

That the Services span both websites, the AI infrastructure platform, marketing and sales initiatives, and job applications.

Workspace and team context

Handling of organization name, workspace identifiers, and team member roles collected in connection with use of the Services.