Protocol
Round 4S throughput cell: this frozen prompt is planned across five shared open-weight checkpoints at 1024x1024 with three unseeded repetitions each. Hosted MiniMax and OpenAI GPT Image 2 Medium calibration arms have run; the five-checkpoint SwarmUI matrix remains blocked and unexecuted.
Frozen prompt template
CONTROL PROMPT R4S-ATTR-001. A wide 16:9 straight-on tabletop photograph of a pure white seamless tabletop against a pure white wall. Exactly three objects and nothing else: on the LEFT half a large matte RED ceramic mug; on the RIGHT half a small glossy BLUE metal thermos bottle; near the BOTTOM CENTER one whole bright YELLOW lemon. The red mug must be clearly left of center. The blue thermos must be clearly right of center. The yellow lemon must sit lower than both vessels and between them. Even soft daylight, no text, no labels, no logos, no extra fruit, no extra vessels, no hands, no patterned surface.Review rubric
- Prompt Adherence
- Attribute Binding
- Spatial Layout
Notes
Frozen prompt sha256 0691aed98a0ca3c6d6fb86cb4e6408ceaeca5b644c2b7107de6abf2c54bceae8 (613 bytes); positive text is byte-identical across all checkpoint arms. Deterministic Pillow signals (ROI colour stats, no OCR): nonwhite means min(R,G,B)<230; left-red-dominance on nonwhite x<0.4W is mean(R-max(G,B)), pass>=25; right-blue-dominance on nonwhite x>=0.6W is mean(B-max(R,G)), pass>=25. bottom-yellow-presence is yellow pixels / all pixels in x=[0.30W,0.70W), y=[0.55H,H), where yellow means R>=150, G>=150, B<=130, min(R,G)-B>=40; pass>=0.05. nonwhite-blob-count (4-connected, area>=0.4%) is diagnostic only because contact shadows can merge objects. Non-claims: N=3 unseeded reps per cell are a launch/screening signal, not a leaderboard; deterministic pixel metrics are objective facts, not quality judgements; rubric-* items require a separate vision review; no OCR or topology facts are produced this round. Failed, moderated, and discarded runs remain in the public data when legally releasable.
