Derivation Web

source_1b2bd4e23a774514

**Selected angle:** `source`

## One-sentence thesis

Across 5 direct receipts sharing llm evaluation accuracy tasks as the evaluation shape and accuracy as the metric, GPT, GPT-4, Claude 3.5 Sonnet report comparable performance against stated baselines. Reported values include 98.83%, 81.3%, 97.0%, 88.57%, 79.5%.


**Interpretation note:** This is a hypothesis-generating alpha memo, not confirmat…