What each policy keeps
A rule is planted at one turn of a long history, then the window is compacted to a budget. Each cell counts the seeded histories in which the rule is still there, word for word, in the window the model would see next.
deterministic
How to read it
- Rows
- Where the rule was planted. Turn 1 is the first turn of the history; turn 18 is the last turn before the request.
- Columns
- The window budget in tokens. The full history is about 3,800 tokens, so every column forces a compaction.
- Cells
- Kept histories out of the seeded histories tried, with the 95% Wilson interval in the tooltip. gone ↓ marks a cell where the rule survived in none of them.
- What this is not
- It says what a policy keeps. It does not say what an agent does afterwards, and it does not include the model-written summary policy, which needs a model. Both come from the experiments. The first, EXP-001, is a 10-seed pilot that used only the model-written summary policy; its results are in the repository, not on this site yet.