Knowledge cutoff
For most uses of a decision model, when its training data ends hardly matters. Here it is a clue to provenance: models built on the same base model tend to share its cutoff, so a matching cutoff hints at a shared origin.
Each model answered 228 news questions from 2023-01 to 2026-09. A confirmed base model is named in parentheses, and known cutoffs are listed below the chart.
Probability on the right answer, by month
Each cell is a model’s mean probability on the right answer over that month’s news questions, averaged over every run, wording and option order. Every question also offers “don’t know”, so a model past its cutoff can put its weight there instead of guessing. The step marks the split that best separates a higher-scoring early stretch from a lower-scoring late one. The above-floor mark is the last month whose three-month neighbourhood still clearly beats the model’s own score on the last three months of the window. It reaches later than the step for models that know only each month’s biggest events.Faded rows have a weak signal: under 35% before the step, or a drop of less than 20 points, so their estimates say little.
Known cutoffs
Training-data cutoffs for these models, or for the base model they were built on, to check the estimates against. They are not used in the estimates. No cutoff is known for the other models.
- gpt-6-luna-decisions (gpt-6-luna) 2026-05-18 (base model)
- solar-decide (solar-mini4) 2026-02 (base model)
- clef (qwen3.8-27b) 2025-05 (base model, estimated)
- pplx-decider-v1.1-27b (qwen3.8-27b) 2025-05 (base model, estimated)