# Fiction model test worksheet

A blank worksheet for your own comparison, not a completed benchmark.
Guide: https://novelthread.app/blog/best-llm-for-ai-novel-writing/

## Task and source

- Task: outline / first draft / dialogue / revision / continuity
- Genre and manuscript language:
- Decision you want this test to support:
- Original passage (use text you have permission to send):
- Facts and sentences that must remain unchanged:
- Required scene beats and length target:
- Exact prompt, including any system instructions you control:

## Run log — complete one copy per attempt

- Blind label:
- Date and time (include time zone):
- Exact model identifier / version:
- Service or application and version:
- Fresh conversation:
- Temperature / output limit / reasoning mode (write unknown if unavailable):
- Run number:
- Original output, unchanged:
- Errors, refusals, truncation or retries:
- Observed waiting time:
- Reported input and output tokens:
- Actual charged usage and unit (keep provider currency and app credits separate):

## Read before revealing the model name

- Constraint errors, with supporting passages:
- Dialogue, viewpoint and voice notes, with examples:
- Total output words (or characters; use one method for all candidates):
- Words / characters retained unchanged:
- Retained share = retained count / total output count:
- Minutes spent revising:
- Revised passage:
- Would you use this output for the stated task? Why?

## Decision and publication

- Which candidate worked for this task and manuscript?
- What failed on other attempts?
- What did this test not cover?
- Keep the mapping from blind labels to model identifiers separate until scoring.
- If publishing, retain dates, settings, prompts and the sample-selection method.
- Obtain permission for shared excerpts and remove private information.
- Retest when the model version or writing task changes.