Scope: original public experiment. I built the application workflow and interface.
Use it to explore:
- Explore how a task-specific rubric changes a comparison.
- Inspect disagreements between a judge's numerical scores and its explanation.
- Create rubric
- Generate answers
- Grade
- Compare
What happens inside
The model proposes three weighted criteria. Two models answer the same prompt, and a grader scores each answer per criterion. Validated weights produce totals; ties and unavailable responses are represented explicitly.
Decisions and tradeoffs
Show the rubric
Criteria and weights make the intended comparison inspectable. Automatically generated criteria still need review.
Run independent answers concurrently
Parallel generation keeps the example responsive without implying a quality benefit.
Treat errors as errors
An unavailable model cannot win a comparison. A numerical tie is a tie, rather than an arbitrary preference for the first answer.
Illustrative example
An illustrative weighting calculation
For scores of 4, 3 and 5 with weights 0.5, 0.3 and 0.2, the weighted total is 3.9/5. This arithmetic fixture explains the scoring mechanism; the scores are authored, not a model judgment.
What this does not establish
One prompt is not a model benchmark or leaderboard. LLM judges can be biased, and criteria may not capture the reader's priorities. Repeatable evaluation requires a dataset and independent checks.
Try a live run
Live requests send your prompt to OpenRouter and the selected model provider. Use public, non-sensitive material. Read how demo data is handled.
Checking live availability…
Technical walkthrough reviewed September 2026. Examples do not imply client adoption or production maturity.