Astra Analysis
Success against time per episode
Each point is one reasoning effort over 45 episodes (15 setups × 3 initial states). Bars are 95% Wilson intervals.
What each effort costs
Per episode, averaged over the 45. Usage is the weekly Codex meter, read in whole percent before each batch.
By task family
Successes out of 9 episodes (3 setups × 3 initial states) per family.
Success by call budget
Share of the 45 episodes already solved after a given number of Astra calls (the run stops at 40). Focus the plot and use the arrow keys to step through call counts.
Show values for every call count
How episodes ended
Every episode ends in one of three ways: the task is done, Astra declares done when it is not, or the 40 calls run out.
Every episode
One cell per episode, grouped by effort and initial state. Select a cell to open its replay.