JDGMNT / MODEL EVALUATION CARD
Version 1 · 24 October 2026

01 / TASK
Recurring job:
Who uses the result?
Task inputs and approved sources:
Output format and intended destination:
What must remain unchanged?
Actions and tools allowed:
What requires human approval?

02 / ACCEPTANCE
Correctness requirements:
Required evidence or references:
Completeness:
Usability in the destination:
Critical failures that cannot be averaged away:
How repaired outputs will be counted:

03 / COMPARABLE CONDITIONS
Workflow A — provider, model version, settings, tools:
Workflow B — provider, model version, settings, tools:
Shared task pack and source versions:
Time limit and retry policy:
Review method and reviewer:
Fresh examples held back for a later check:

04 / RESULT RECORD — REPEAT FOR EACH TASK
Task ID:
Workflow:
Attempts, including failures and retries:
Accepted unique results:
First-pass acceptance, if tracked:
Service and tool charges:
Review and repair minutes:
Critical failures:
Notes explaining the decision:

05 / COST BOUNDARY
Reviewer hourly cost assumption:
Setup and maintenance costs (record separately or allocate explicitly):
Expected volume if allocating setup:
Total included cost:
Cost per accepted result = included cost / accepted results.
If no result is accepted, the ratio is undefined.

06 / ADOPTION DECISION
Which specific job benefits?
What remains unresolved?
Continue / revise / stop:
Owner and next review date:

A small pilot is not a general model ranking. Compare the full workflow,
not just its first response. Keep serious failures visible beside averages.

You may use and adapt this worksheet, including commercially. Attribution to
JDGMNT is appreciated. It is a planning aid, not a validated benchmark.