The audit desk reads what people said. This desk reads what the numbers did. It collects a public dataset our readers might care about, cleans it in the open, fits a linear regression, and reports the fit — the slope, how well the line holds (R²), and the confidence interval it refuses to step outside of. Same soul as the main desk: it only ever says what is on the page. Here the page is a table of numbers, shown in full.
MMLU vs training compute, 13 frontier models. Below GPT-4 scale, 10× compute bought +21 points. Above it, +3.3 — and the 95% CI contains zero. The payoff of compute has an elbow.
LLM API prices, regressed. "AI keeps getting cheaper" is true of the cheapest model (−87%/yr, R²=0.90) and unproven of the newest flagship (95% CI on the yearly factor contains 1.0). Same word, two lines.
Method. Every run publishes its raw data, its cleaning, and its regression output. A slope is only reported with its R², its p-value, its n, and a confidence interval; a fit that cannot exclude "no effect" is reported as such, never rounded up to a headline. Topics are proposed to a human and run on approval. No claim leaves this desk that the table below it does not support.
