Add atlas refine cue seed workflow

ab_batches/atlas_refine_baseline_deck_analysis.md

@@ -0,0 +1,71 @@
+# Atlas Refine Baseline Deck Analysis
+
+Date: 2026-07-01
+
+Folder:
+`/media/unraid/comfyui/output/CodexMCP-Atlas-Refine`
+
+Subject id:
+`atlas_refine_same_woman_001`
+
+## Folder State
+
+The folder contains 16 complete `.png` / `.txt` atlas prompt pairs.
+
+Coverage report:
+
+- 10 clean `baseline_only` entries
+- 6 `needs_prompt_cleanup` entries
+- 0 sidecar prompt variants
+- 0 seedable catalog cue candidates
+- 0 missing image/text pairs
+
+The deck is useful as a controlled same-woman baseline set, not yet as a
+seed/cue system. Reviewed sidecar variants still need to be authored from
+actual prompt/image evidence.
+
+## Prompt Cleanup Queue
+
+These baseline prompts still contain positive-channel negative wording through
+`without` and should be rewritten before visual scoring or seed promotion:
+
+- `pov_cowgirl_alt_low_squat_penetration`
+- `pov_cowgirl_frontal_straddle_penetration`
+- `pov_ejaculation_aftermath_open_thigh_candidate`
+- `pov_handjob_upright_centered`
+- `pov_reverse_cowgirl_alt_upright_back_facing_penetration`
+- `pov_reverse_cowgirl_back_facing_penetration`
+
+## Visual Deck Read
+
+The deck is strong for same-subject comparison:
+
+- the woman identity is stable across poses;
+- the coworking lounge style is consistent;
+- desk rows, laptops, chair wheels, plants, glass partitions, and warm window
+  depth are repeated enough to support workspace-continuity scoring;
+- most poses place the action as the primary foreground subject.
+
+Current limitations:
+
+- all entries are single baseline images, so they do not test seedable
+  alternatives yet;
+- several prompts still use cleanup-needed wording;
+- clothing/control is not represented across most poses;
+- workspace interaction is mostly background context, not varied physical use
+  of the lounge furniture;
+- exact top-view oral remains weak under text-only prompting even after the
+  second MCP batch.
+
+## Next Evidence Priorities
+
+1. Clean the six `without` prompts before using them for cue seeds.
+2. Build prompt-variant sidecars only from tested cues, not invented wording.
+3. For each pose family, collect same-subject fixed-seed variants before
+   touching generator defaults.
+4. Score scene/pose height from background cues: floor plane, desk height,
+   chair wheels, table bases, and whether the viewer/partner appears standing,
+   seated, kneeling, reclined, or supported by furniture.
+5. Keep clothing restore as woman-owned visible detail, and only for garments
+   that the pose crop can actually show.
+