Document generator reproduction checkpoint

docs/krea2-ab-methodology.md

@@ -45,18 +45,26 @@ Version: `2026-06-30-generated-route-validation-positive-channel-cleanup`
     code path before closing the category. Manual prompt-axis wins are not
     enough; the generated route can still drop the key contact hierarchy or add
     limiting positive-channel wording.
-14. Treat the final prompt that Krea receives and the rendered generator images
+14. Manual prompt-axis exploration can stay hand-written while testing a wide
+    range of new wording. Once a manual atlas prompt shows a promising axis,
+    run a generator-reproduction checkpoint before treating it as route
+    evidence: make the actual SxCP generator emit the closest equivalent final
+    Krea prompt, compare that generated prompt against the manual winner, and
+    record any missing hierarchy, wording order, restored detail, or
+    scene/camera constraint. This avoids building a strong atlas prompt that the
+    generator cannot reproduce.
+15. Treat the final prompt that Krea receives and the rendered generator images
     as the source of truth. Tests are guardrails for known failures, not proof
     that the prompt will render the intended pose. If a test passes while the
     rendered image or prompt audit shows drift, update the test and the method
     rather than trusting the test.
-15. Atlas/catalog prompt wording must use direct visual prompt sentences. Do
+16. Atlas/catalog prompt wording must use direct visual prompt sentences. Do
     not send option-list wording such as `or`, `may`, `optionally`, or `either`
     in pose cues, and do not append meta instructions such as `keep the visible
     partner...` or generic camera-layout prose to atlas routes. Krea2 is not an
     instruction-following LLM in this loop; prompt text should describe the
     image, not explain a policy.
-16. Do not promote generator edits from a few isolated renders unless the final
+17. Do not promote generator edits from a few isolated renders unless the final
     generated prompt is structurally wrong before rendering. For pose wording
     changes, collect prompt/image evidence across multiple women, source cases,
     and seeds when feasible, then patch only the repeated generator-safe
@@ -93,6 +101,20 @@ manual prompt-axis winner, record the failed generated-route image as the
 baseline, tighten the route wording, and validate again before logging the
 candidate as generated-route evidence.
 
+Generator-reproduction checkpoint:
+
+1. Keep the manual winner as the reference prompt and summarize its required
+   visual hierarchy.
+2. Build the closest equivalent prompt through the normal SxCP nodes and Krea2
+   formatter.
+3. Diff the final generated prompt against the manual reference for missing
+   pose hierarchy, changed ordering, restored details, scene anchors, camera
+   framing, and option/meta wording.
+4. If the generated prompt cannot reproduce the manual hierarchy, fix the
+   generator path or record the gap before adding more image evidence.
+5. Only count generated-route evidence after the final generated prompt
+   preserves the manual winner's essential atlas hierarchy.
+
 For location-specific wins, split the implementation:
 
 - the action or role graph owns the pose/contact hierarchy;