Draw the chart from its title

How this was scored

Each model receives only an English Wikipedia article title and the request for monthly pageviews from January 2016 through December 2025. No real values, chart, article text or tools are supplied. This tests recall of quantities and events, rather than visual chart reading.

For each article, convert both series to log10(max(value, 1)), take the squared difference in each of the 120 months, average those differences and take the square root. The overall score is the mean of those 197 article errors, with each article weighted equally. Lower is better; 0 is exact. “Typical miss” is 10 raised to the overall error, so an error of 0.30 is roughly a 2× miss.

The shape score is the mean of the 197 Pearson correlations between the log10 series. Higher is better, from −1 to 1. A constant predicted curve scores 0. It is a second view of the same responses and should not be treated as independent evidence.

Models use the provider's default reasoning setting. OpenRouter requests 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. The numerical download preserves each model's route and settings.

An unparsable answer can be retried once. If it still cannot be parsed, grading uses that article's flat-line reference score. The download preserves the missing model prediction as null; the interactive plot never draws that reference as if the model had answered. The flat reference is a constant geometric mean computed from the other 196 articles, without looking at the article being scored.

The 197 articles came from a seeded sample of one October 2025 hourly pageview dump, keeping complete ten-year series with a median of at least 3,000 monthly views. Eight articles renamed during the window were removed. This favors subjects being read around midday in that October dump. Each model ran once, so asking again may change the results.

Exact shared prompt template

This is a chart-drawing game. The chart shows monthly pageviews of the English Wikipedia article "{title}" from 2016-01 to 2025-12 (human visitors only, desktop and mobile combined, as reported by the Wikimedia pageviews API).

Draw the chart from what you know about the subject: its overall popularity, events that would have sent people to the article, seasonality and long-run drift. Give your best estimate of the absolute number of views for every month.

End your reply with one JSON object mapping each month ("YYYY-MM") to an integer, with all 120 months from 2016-01 to 2025-12 in order.

The complete download includes the exact system prompt, all 197 instantiated user prompts, ground truth, model predictions and per-article reference levels. It contains everything needed to replay both scores.