Find the smallest model that works
Run evals across multiple LLMs, apply your own judge, and get statistically backed answers on the most efficient model for your task.
Confidence you can trust
Statistically grounded — Wilson intervals & binomial tests, not vibes. Early stopping — stop wasting tokens once a model clearly passes or fails. Judge your way — combine programmatic checks with an LLM judge.
Built for clarity
Every run shows exactly: Which models passed, failed, or stopped early. Confidence intervals vs your target threshold. Cost and latency per successful answer.
Streamline your workflow
Upload datasets & rubrics in JSONL/JSON. Configure runs with p*, α, batch size, and model ladder. Drill into details: per-model metrics, per-item evidence, judge rationales.
Multiple models... just the right one
Compare responses across different models to find the perfect balance of quality, cost, and performance for your specific needs.
Configuration
Run parameters (read-only)
Evaluation Parameters
Target Success Rate
85%
Confidence Level
95%
Sample Size
150
Batch Size
25
Early Stopping
Model Ladder
llama-2-7b
claude-haiku
gpt-5
Number of Models: 3
Run quick sanity checks or comprehensive evaluations.
statistical confidence, and cost optimization.
managing schedules, and growing with data-driven insights.
