Each model sees only an English Wikipedia article title and draws its 120 monthly pageview values from January 2016 through December 2025. It receives no chart, page text, real values or tools. The same 197 articles are scored for every model; this tests recall of real-world quantities and events, not visual chart reading.

Articles come from a seeded sample of an October 2025 pageview dump, keeping complete ten-year series with a median of at least 3,000 monthly views. Eight articles renamed inside the window were removed so their missing earlier-title traffic would not count against a model.

One answer per chart at the provider's default reasoning setting, with one retry if it cannot be parsed. All 197 charts count: a final unparsable answer receives that article's flat-line reference score. OpenRouter calls allow 32,000 output tokens including reasoning. Claude models use the Claude CLI without a recorded output limit. Fable 5.2 and Carbon each reuse 28 answers to identical pilot prompts; Carbon's reused answers used a 16,000-token limit and finished below it.

The score is the mean, across the 197 articles, of the root mean squared difference between predicted and real log10 monthly views. Lower is better; 0 is exact. An error of 0.30 corresponds to a typical multiplicative miss of about 2×. This score measures both the article's popularity level and the curve's shape.

• Average article (baseline): 0.402218
• Flat line (baseline): 0.403034

• The reference predictors know nothing about the article. They use the geometric mean of the other 196 articles: their average monthly curve, or a constant level as a flat line. They are references, not model observations.
• The sample favors articles read around midday UTC in one October 2025 dump. Each model ran once. The study's chart-bootstrap intervals measure variation across articles, not variation from asking a model again.