LLM Evaluation Rubric Builder
Create weighted evaluation criteria, preview manual scores and export a reusable rubric locally.
Build an evaluation framework and enter example scores manually. No prompt or evaluation is executed. Presets are editable examples, not universal standards.
Reusable output
Generic Draft 2020-12 schema checks shape and ranges, not weighted arithmetic or factual correctness. Provider compatibility is not implied. Generated evaluator text requests concise evidence-based explanation, never private chain-of-thought.
Ready. Define your criteria or load an optional preset.
Enable JavaScript to build locally.
Define a reusable evaluation rubric.
Start with custom criteria or edit an optional preset. Presets are examples, not universal standards. All criteria share a configurable scoring range: use 0–1, 1–5, 1–10 or your own minimum and maximum.
Preview deterministic scoring.
Enter manual example scores. Positive weights are divided by their sum without changing your entries. The percentage is 100 × sum(weight × (score − minimum)) / (sum(weights) × (maximum − minimum)). The pass threshold uses the unrounded result. This tool never judges or scores responses using AI.
Export without executing.
Copy a readable rubric, versioned JSON, an inert evaluator prompt or a generic Draft 2020-12 result schema. Criterion IDs follow their order. The schema checks result shape and ranges; it cannot verify weighted arithmetic or evidence. It does not claim provider-specific compatibility or request private chain-of-thought.
Local and bounded.
Up to 12 criteria, names of 64 characters and descriptions of 500 characters. Weights and scores support three decimal places. Content stays in browser memory and clears on reset or leaving the page. Copy writes only to your clipboard.