Do-it-yourself game plans

What should the final decision memo contain?

Frontier Development research deskSection 02 · Page 11 of 15Last updated 26 October 2026

□ One-sentence decision: retain, change prompting/retrieval, pilot adaptation, deploy or stop.
□ Measured results and rejected alternatives, with dates and versions.
□ Error examples and important slices, without exposing protected answers publicly.
□ Whole-life scenarios and unmeasured assumptions.
□ Rights, ownership, support and exit responsibilities.
□ Named approvers, next review trigger and reversal conditions.

A coherent small pilot could occupy several weeks; the stage allowances above are deliberately not a fixed delivery promise. Missing rights, scarce expert reviewers or a difficult deployment environment justify rescoping before more training spend.

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.

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