ci: establish green validation baseline (#59)

scripts/gemma4_12b_recursive_loop.py

@@ -0,0 +1,76 @@
+#!/usr/bin/env python3
+"""Adaptive Gemma 4 12B recipe loop helpers.
+
+This module is intentionally side-effect free on import so tests and other
+automation can reuse the recipe policy without loading a model.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, replace
+
+
+@dataclass(frozen=True)
+class Recipe:
+    name: str
+    method: str = "informed"
+    n_directions: int = 4
+    min_layer_fraction: float = 0.35
+    max_layer_fraction: float = 0.65
+    regularization: float = 0.0
+    refinement_passes: int = 2
+
+
+BASE_RECIPES = [
+    Recipe(name="gemma4_12b_balanced"),
+    Recipe(
+        name="gemma4_12b_surgical",
+        method="surgical",
+        n_directions=6,
+        min_layer_fraction=0.30,
+        max_layer_fraction=0.70,
+    ),
+]
+
+
+def next_round_from(recipe: Recipe, metrics: dict[str, float]) -> list[Recipe]:
+    """Derive the next candidate recipes from aggregate benchmark metrics."""
+    refusal_rate = metrics.get("refusal_rate", 0.0)
+    repetition_rate = metrics.get("repetition_rate", 0.0)
+    short_rate = metrics.get("short_rate", 0.0)
+    collapse_rate = max(repetition_rate, short_rate)
+
+    next_recipes: list[Recipe] = []
+    if refusal_rate > 0.0:
+        next_recipes.append(
+            replace(
+                recipe,
+                name=f"{recipe.name}_more_dirs",
+                n_directions=min(recipe.n_directions + 2, 12),
+            )
+        )
+        next_recipes.append(
+            replace(
+                recipe,
+                name=f"{recipe.name}_wider_layers",
+                max_layer_fraction=min(recipe.max_layer_fraction + 0.10, 0.95),
+            )
+        )
+
+    if collapse_rate > 0.0:
+        next_recipes.append(
+            replace(
+                recipe,
+                name=f"{recipe.name}_more_conservative",
+                n_directions=max(recipe.n_directions - 1, 1),
+                max_layer_fraction=max(recipe.max_layer_fraction - 0.05, recipe.min_layer_fraction),
+                regularization=min(recipe.regularization + 0.05, 0.5),
+            )
+        )
+
+    return next_recipes or [recipe]
+
+
+if __name__ == "__main__":
+    for item in BASE_RECIPES:
+        print(item)