$darwin

We checked

Testing was inconclusive: no request could be sent without workspace credit, so difficulty grading, budget transitions and conversation stickiness were not observed.

To settle it Configure a ladder and budget in the app, send requests with varied wording and tools, and inspect the x-darwin-model and x-darwin-rung response headers.

Skeptic: plausible Rule-based request grading, budget thresholds and model fallback are feasible routing behaviors, and the docs describe how they are intended to work. The page does not independently establish routing quality or cost savings.

We used it: inconclusive The app and docs describe the claimed grading, budget-based routing, and keeping a conversation on one model unless it fails or runs out. Those descriptions are not runtime verification; no request was routed during this test.

We checked

Unauthenticated endpoint requests returned 401; successful authenticated format compatibility and MCP behavior were not tested. These responses do not refute compatibility.

To settle it Use a Darwin key to call the documented endpoints with OpenAI- or Anthropic-compatible clients and inspect their responses.

Skeptic: plausible The documented OpenAI- and Anthropic-format endpoints and MCP interface are technically feasible. The documentation is a useful specification, though independent implementation testing is not provided here.

We used it: inconclusive The documentation lists the endpoints and claims OpenAI/Anthropic compatibility, but I did not verify successful authenticated responses. The 401 responses were unauthenticated browser requests, not evidence that the supported formats fail.

We checked

No GPU response or credit deduction was observed; the documented model, limits and prices remain untested.

To settle it Send requests routed to Darwin's GPU and check the reported model, output limits, usage and workspace credit changes.

Skeptic: plausible Running a Qwen model on rented GPUs and charging usage by token is feasible. The stated model deployment, limits and prices are project claims; the pages do not independently verify that the service is operating at those specifications.

We used it: inconclusive The published specification matches the claim, but model identity, output cap, usage accounting, and credit deduction were not verified by an actual GPU response.

We checked

The tester held no DARWIN and performed no burn. Mint identification does not establish memo processing or the claimed credit conversion.

To settle it Review the Solana burn transaction and memo, then compare the resulting workspace credit balance.

Skeptic: plausible A Solana burn transaction with a workspace memo can be read and credited once by an application. The $5-per-million rate is an internal credit rule, not evidence that the operator receives $5 in cash for each million tokens burned.

We used it: inconclusive The UI describes a normal Solana burn transaction with a workspace memo and $5 credit per million tokens, but no burn or corresponding credit was produced. The mint lookup confirmed the provided address is the DARWIN mint; it does not verify the crediting flow. No token burn was performed.

We checked

The rates appeared in the UI, but upgrade activation and credit deduction were not verified; the workspace had zero credit and the attempted interaction was inconclusive.

To settle it Purchase an upgrade in the app and verify its activation, behavior and credit deduction.

Skeptic: plausible Charging credit for model tiers and warm compute is a workable pricing design. Whether the listed rates cover actual GPU and operating costs is not established, and credit creation via burns does not itself fund those costs.

We used it: inconclusive The listed rates match the claim, but an upgrade was not verified as purchased or activated. With no workspace credit, the operational behavior and deduction could not be tested.