Specialist-model research desk

Scorecard

What must pass before a score can influence an award?

Frontier Development research deskSection 09 · Page 2 of 13Last updated 26 October 2026

Treat the following as hard gates, tailored to the buyer’s use case:

  • Permitted use: data, teacher outputs, model and code can be used in the proposed way.
  • Independent acceptance: supplier accepts the buyer’s protected test and dated practical baseline.
  • Critical quality: the plan defines non-negotiable failure limits; production award requires measured passage.
  • Deployment and security: required processing, access, availability and incident controls are demonstrably achievable.
  • Ownership: named people own acceptance, operations and handover; mandatory exit rights are resolved.

For a pilot award, distinguish “credible plan to prove the production gate” from “already passed”. Do not demand a production result before feasibility work, but do require a credible way to stop if it fails.

Evaluation design, specific licence terms and risk-management guidance inform these questions. The gates are the desk’s purchasing policy, not a claim that each source mandates the exact same procedure.

TRACE THE EVIDENCE

Sources cited on this page

  1. Evaluation design →www.anthropic.com
  2. specific licence terms →github.com
  3. risk-management guidance →www.nist.gov