Specialist-model research desk

Do-it-yourself game plans

Stage 8: hand over, rehearse exit and operate

Frontier Development research deskSection 02 · Page 10 of 15Last updated 26 October 2026

Planning allowance: 1–3 weeks for a bounded rollout, with subsequent ongoing ownership. Deliver the manifest, configuration, artefact hashes, evaluation harness, data dictionary, cost workbook, deployment instructions and rollback procedure. Use a staged rollout appropriate to the risk, with manual fallback where needed.

Run an export/redeployment rehearsal. Test that the buyer can retrieve adapters or weights as contractually permitted, reproduce preprocessing and execute the evaluation elsewhere. Maintain access boundaries for customer data and deletion obligations for suppliers.

Set triggers: a new document family, policy change, critical error, sustained review growth, changed provider pricing or a better practical baseline. A trigger starts diagnosis; it does not automatically authorise retraining. Keep a record of every release and why it passed.

TRACE THE EVIDENCE

Section source library

This page contains no direct external citations. These are the section’s references; they do not independently validate the desk’s proposed scoring or methods.

  1. Demystifying evals for AI agents — Anthropic, 9 January 2026 →www.anthropic.com
  2. Quantization — Hugging Face PEFT documentation →huggingface.co
  3. LoRA Learns Less and Forgets Less — Biderman et al., revision 20 September 2024 →arxiv.org
  4. Pricing — Together AI, snapshot checked 26 October 2026 →www.together.ai
  5. GroupKFold — scikit-learn documentation →scikit-learn.org
  6. A Field Guide to Rapidly Improving AI Products — Hamel Husain →hamel.dev
  7. Guidance on HIPAA & Cloud Computing — HHS Office for Civil Rights →www.hhs.gov
  8. TimeSeriesSplit — scikit-learn documentation →scikit-learn.org
  9. Extend — Extraction Overview →docs.extend.ai
  10. SFT Trainer — Hugging Face TRL documentation →huggingface.co
  11. RouteLLM: Learning to Route LLMs with Preference Data — Ong et al., revision 23 February 2025 →arxiv.org