Rule Slate Walkthrough

The 10 rules selected for the follow-up self-explanation study. Each row links to its full atlas entry; click 👁 to expand the diagnostic-hands panel and see the five candidate adversarial-hand-selection strategies side by side. Strategy default follows the demo page's toggle.

Serve over HTTP (file:// is blocked): python -m http.server 8765 --directory rule-gallery/analysis/selection then open http://localhost:8765/

Loading…

Legend (click to collapse)
RB/II classification
RB_CLEAR verbalized + classified ✓
MIXED partial articulation
II_CANDIDATE classified OK, can't say why (Ashby II signature)
TOO_HARD no signal — drop
Failure mode (fail_dir)
mostly_overfit lambdas too narrow
mostly_undergen lambdas too broad
balanced mixed direction
near_ceiling rule is easy
low_signal < 5 directional fails
Decile (1=lowest, 10=highest)
D1 D2 D3 D4 D5 D6 D7 D8 D9 D10
Each rule gets one decile per metric (mcc_mean, H_norm, base_rate). Same color scale for all three.
Toggle columns

Distributions

Histogram bars show the count of rules per bin (right y-axis, hidden). Dots overlay each rule at its x-value; vertical jitter on the dots is random, only used to keep dots separable for hovering — it carries no meaning. Hover any dot to identify the rule; click a dot to scroll the table to that row and flash-highlight it. Selected rules render as large star markers so the slate is visible at a glance.

Three dimensions at once

Each rule is one dot in (mcc_mean × H_norm × log10 base_rate) space, colored by RB/II class. Drag to rotate; hover for details.