Integration workspace

The Business Playground contains Request builder and the available Images, Audio and Video tabs. Model Verification is not shown in this navigation; its existing directly authorized routes remain separate. The builder is for experimenting with a supported model's inputs. Verification is a separate behavioral-reference comparison with a foreground judge and its own evaluation billing.

Request builder

Select a project credential that your role can execute. Inspect the selected protocol and model before sending attachments. Responses file input uses input_file with a complete PDF data URL; Anthropic document input uses its native document block. Image generation in the builder calls the Responses image_generation tool when that capability is supported.

A rejected provider request is not automatically retried as a different protocol. Inspect the displayed diagnostic error and request shape, without exporting secrets. Not every model supports every builder tab.

Verification lifecycle

Choose a model with a ready, versioned reference. The application selects 32 saved facts across sixteen consecutive completed months around the reference's observed RECALL anchor. It does not research a cutoff, generate facts or fill the bank when a company user starts a test. Without sufficient reference coverage the model is unavailable, even when an official cutoff is known.

The tested request and evaluation run in the open browser. Your external endpoint key stays in memory and is never forwarded to the B2B LLM backend. Questions and instructions go to the tested model; expected answers and stored reference observations go only to the judge. The selected reference pins protocol and an exact request contract. New automatic references use minimal-v1 with no optional temperature/reasoning/stream/tool controls; historical references replay their stored legacy options. Changing the route does not silently change protocol or request identity. The company credential used to pay for judging must be executable under the actor's current permissions, group constraints and wallet rules. Your verification key is used only for the requested evaluation under its existing permissions.

Keep the page open until the report is ready. There is no background evaluator or raw user-response recovery store for a closed tab. Retrying grading uses the answer held by that page, not a second tested-model request. A disconnected call can still have incurred a charge. Publication remains opt-in; private normalized reports do not store verbatim per-question answers for browser recovery.

The judge alone evaluates correspondence, including unexpected successes and unexpected failures against the reference distribution. The server does not assign penalties or reinterpret the score. A correct answer can be evidence of differing behavior without becoming a factual error. The assessment is not proof of the provider's identity, a newer model, a cheaper route or training-data provenance. Read its explanations and limitations as well as its score. Older reports and reference series keep their saved historical benchmark.

Return to connection recipes after validating the request you intend to automate.

Unsupported API paths now return 404 endpoint_not_supported with the requested path; a wrong method on a supported path returns 405 method_not_allowed and Allow. Native /v1/balance is unchanged; /v1/user/balance is not implemented. Share the returned request ID, never your API key, when reporting these errors.

Saved Images workspace

The Images tab is for durable reviewed-route generation/editing, unlike the builder's foreground Responses tool. It uses the employee's current project permission and short one-intent grants, with route price confirmation, status/history and private downloads. A reload reads saved work; it never resumes by generating again.