Example proofs

See the proof.

Real ranked proof reports on illustrative synthetic data. One per modality: transcription, summarization, extraction, vision, and reasoning.

All examples use small, clearly-synthetic data. No real client or PHI data.
🎙️
Transcription WER + medical-term F1
Clinical transcription, Whisper v3 vs turbo

Whisper large-v3-turbo wins, 49% cheaper at equal medical accuracy

View proof
📝
Summarization LLM judge
Support-ticket summarization

Qwen3 235B wins, 99% cheaper; the flagship model was both priciest and lowest quality

View proof
📄
Extraction Field F1
Contract field extraction

Qwen3 235B matches accuracy at 54% lower cost (F1 0.98)

View proof
👁️
Vision / OCR Field F1
Document data extraction (vision)

Llama-4 Maverick reads the docs, 85% cheaper (F1 1.00)

View proof
🧠
Reasoning Exact-match + cost-per-correct
MMLU-Pro reasoning

Cheapest-per-correct wins; the most accurate model is not the economic winner (96% cheaper)

View proof