Add atlas refine cue seed workflow

ab_batches/blowjob_top_down_vertical_shaft_refine_analysis.md

@@ -0,0 +1,100 @@
+# Blowjob Top-Down Vertical Shaft Refine Pass
+
+Date: 2026-07-01
+
+Pose target: `pov_blowjob_top_down_vertical_shaft`
+
+Reference folder baseline:
+`/media/unraid/comfyui/output/CodexMCP-Atlas-Refine/pov_blowjob_top_down_vertical_shaft_00001_.png`
+
+Prior batch:
+`ab_batches/blowjob_top_down_vertical_shaft_axis_batch.json`
+
+Second-pass batches:
+
+- `ab_batches/blowjob_top_down_vertical_shaft_refine_seed_238365845574312_batch.json`
+- `ab_batches/blowjob_top_down_vertical_shaft_refine_seed_238365845574313_batch.json`
+
+Second-pass results:
+
+- `ab_batches/blowjob_top_down_vertical_shaft_refine_seed_238365845574312_results.json`
+- `ab_batches/blowjob_top_down_vertical_shaft_refine_seed_238365845574313_results.json`
+
+Sampler seeds:
+
+- `238365845574312`
+- `238365845574313`
+
+Total text-only attempts counted for this pose: 51 prompt/seed outcomes
+across the first and second MCP batches.
+
+## Result
+
+The second pass is stable but does not solve the atlas top-view geometry. Most
+variants converge to a front-facing kneeling or seated oral composition: contact
+is often preserved, subject identity stays strong, and the coworking floor grid
+is coherent, but the camera still reads forward/downward rather than true nadir.
+
+This means the exact atlas family shown by `blowjob_top_view` references is a
+weak text-only case for this model under subject-first prompting. Treat it as a
+candidate for stronger image/control guidance if a flatter atlas match is
+required.
+
+## Best Repeatable Partial
+
+The best repeatable partial remains the first-pass direction:
+
+- `axis_mouth_directly_below_torso`
+- `axis_floor_plane_priority`
+
+The refined `floor_plane_mouth_under_torso` prompt preserved contact on both
+sampler seeds and produced stable office-floor framing:
+
+- seed `238365845574312`: `/media/unraid/comfyui/output/agent_bridge/img_b6525cf541d44c34828c7fe19068425b.png`
+- seed `238365845574313`: `/media/unraid/comfyui/output/agent_bridge/img_fe09b3d132004f26ab848f3edb21d8a0.png`
+
+It is better than a generic seated oral prompt because it anchors the partner
+below the viewer, keeps the mouth/contact centered, and uses floor/caster/desk
+anchors. It is not a proven exact top-view atlas reproduction.
+
+## Stable But Insufficient
+
+These cues mostly improved consistency rather than verticality:
+
+- `caster_wheel_floor_grid`
+- `desk_leg_rows_floor_depth`
+- `viewer_feet_gate_head_center`
+- `straight_down_carpet_tiles`
+- `mouth_below_navel_eye_contact`
+- `kneeling_between_shoes_floor_map`
+- `short_column_contact_scale`
+- `desk_edges_side_floor_anchor`
+- `head_crown_forehead_vertical_read`
+- `office_lane_floor_perspective`
+- `low_head_high_floor_ratio`
+
+They tend to preserve the coworking deck and subject identity, but they do not
+force a flat top-view body plane. Several reduce the action to a conventional
+front-facing kneeling oral pose.
+
+## Weak Or Rejected Axes
+
+- `phone_snapshot_abdomen_down` often broke mouth contact and produced a staged
+  kneeling pose.
+- `woman_tank_top_owned_visibility` confirmed owned clothing works, but it did
+  not improve the top-view geometry.
+- More repetitions of `nadir`, `straight-down`, `floor plane`, `viewer feet`,
+  and `desk-leg grid` are unlikely to solve the exact atlas pose alone.
+
+## Prompt Rules From This Pass
+
+- Keep as a provisional prompt-guide partial: `mouth directly below the viewer's
+  torso` plus `floor-plane-priority` office anchors.
+- Keep office-floor anchors concrete: `chair caster wheels`, `desk-leg feet`,
+  `table bases`, `carpet texture`, `floor seams`.
+- Owned clothing remains valid only when phrased as `The woman wears ...`.
+- Do not promote exact top-view success from this pass.
+- For catalog seed variants, this pose needs either a partial seed family
+  labeled as forward/downward POV oral, or stronger control/image guidance for
+  the flatter atlas top-view family.
+