Prompt Review — where a prompt's versions learned the design period

Left: the metric per version in each sample. Right: in sample against out of sample, bubble size = weight, decision or veto rules; hover a bubble for its version. Below the diagonal a version promised more in the design period than it delivered.

The model's reading

What each version of the prompt carries

One row per version. Rules = families present in the version (a rule reworded between versions counts once); hard = weight, decision or veto. Prose = share of sentences without an obligation or a condition. Input = blocks the model reads at the same tick for every version and how many come with a conclusion or a vote. Zeros = share of 0 signals. Click a cell to see what is behind it; ★ marks the version the model would keep.

Weight of each component by version

Read off the text: 0 absent · 0.5 frames the question, no weight · 1 context or path only · 2 adjusts or vetoes · 3 co-sets the direction · 4 sets the direction alone. Click a cell for the role and the sentence that fixes it.

What changed between consecutive versions