MODEL FIELDNOTES / LAUNCH EVIDENCE WORKSHEET
By Swapnil Upganlawar / 23 October 2026
Companion: https://modelfieldnotes.com/notes/ai-evaluation-measurement-gap/

This is a proposed review aid, not a certification or a substitute for domain-specific requirements.

DECISION
What action are we deciding whether to delegate?
What observable result counts as correct completion?
Which mistakes require stopping, escalation or rollback?

SYSTEM
Model/provider/version:
Harness/source revision:
Prompt, tool schema and context configuration:
Retrieval/data snapshot and permissions:
Time, token, retry and spending limits:

EVALUATION
Evaluation date and last relevant production change:
Task population, sampling method and known exclusions:
Distinct tasks and repeated trials per task:
Denominator for each metric; incomplete runs included or excluded:
Grader/rubric version; reviewer overlap and calibration:
Baseline and candidate tested on the same frozen cases:
Confidence intervals or other uncertainty estimates:

OUTCOMES
Verified completions:
Incorrect completions:
Timeouts and unfinished work:
Unsafe or duplicate actions:
Human review, intervention and cleanup:
Total cost per acceptable outcome, with pricing assumptions:

NEXT CHECK
Named owner:
Production signal that could contradict the test:
Change or incident that invalidates this evidence:
Re-evaluation date or trigger:

Zero-failure arithmetic applies only under its stated independent-trial assumptions.
Repeated paraphrases, clustered cases and reused answers are not automatically independent evidence.