An open-model scorecard built around the task

Apply entry conditions first

List the exact version, permitted environment, access roles and essential functions. A failed condition is a separate finding; a strong average score must not cancel it. Record the actual model, prompt, retrieval material and generation settings.

Separate ordinary, difficult and reserved tasks

Ordinary tasks represent daily work. Difficult cases test missing evidence, conflicts and formatting failures. Reserve separate tasks for acceptance, so repeated tuning does not make the result look better than it is.

Use scores to record judgement

A starting design can score task correctness, source support, correction burden and workflow fit from zero to four. Confirm weights before testing; report factual errors and access-boundary failures separately rather than hiding them in an average.

Compare the full cost of a usable result

Record response time, service or inference costs, equipment, operations, checking and rework. Unknown costs remain unknown. The goal may be a fit-by-task conclusion, not a forced universal winner.

MAKE IT YOURS

Download editable template

Download a UTF-8 text file and edit it in a text editor or word processor. Your completed worksheet stays in your own file; this page does not upload it.

Download editable template

Preview the blank template

# Model evaluation scorecard

Project / evaluation date:
Owner / reviewer:

## A. Entry conditions (pass / fail / unconfirmed)
Exact version and use conditions:
Processing environment and access arrangement:
Essential functions:
Action when a critical condition fails:

## B. Configuration
Model identifier / version / file hash:
Hosted or local runtime:
Prompt version:
Retrieval material version:
Generation settings / concurrency:

## C. Task record (copy for each task)
Task ID / ordinary, difficult or reserved:
Input and expected result:
Run count:
Output location:
Critical errors and scoring reasons:
Task correctness (0–4):
Source support (0–4):
Low correction burden (0–4):
Workflow fit (0–4):
Response time:
Review / correction / rework time:

## D. Weights confirmed before testing
Suggested starting values: 40%, 25%, 20%, 15%; adjust to total 100%.
Total = sum of percentage weight × criterion score ÷ 4.
Failures that an average cannot cancel:

## E. Conclusion
Observed results:
Costs and unconfirmed assumptions:
Suitable / unsuitable tasks:
Further tests needed:
Location of configuration and records: