Research instrument · closed 2026-09-10 · v0.1.0
A public image-generation laboratory where every specimen keeps its frozen prompt, model identity, provenance, review, and visible defects.
Final evidence
oc26r-r5c-source-composition-lock-openai-gpt-image-2-medium-04Reviewed on the labeled refinement contact sheet against the unpublished operator reference. This run is retained as matrix evidence; model ranking is reported at campaign level.
oc26r-r5c-source-composition-lock-openai-gpt-image-2-medium-03Reviewed on the labeled refinement contact sheet against the unpublished operator reference. This run is retained as matrix evidence; model ranking is reported at campaign level.
oc26r-r5c-source-composition-lock-openai-gpt-image-2-medium-02Reviewed on the labeled refinement contact sheet against the unpublished operator reference. This run is retained as matrix evidence; model ranking is reported at campaign level.
oc26r-r5c-source-composition-lock-openai-gpt-image-2-medium-01Reviewed on the labeled refinement contact sheet against the unpublished operator reference. This run is retained as matrix evidence; model ranking is reported at campaign level.
oc26r-r5b-macro-chaotic-order-openai-gpt-image-2-medium-04Reviewed on the labeled refinement contact sheet against the unpublished operator reference. This run is retained as matrix evidence; model ranking is reported at campaign level.
oc26r-r5b-macro-chaotic-order-openai-gpt-image-2-medium-03Reviewed on the labeled refinement contact sheet against the unpublished operator reference. This run is retained as matrix evidence; model ranking is reported at campaign level.
oc26r-r5b-macro-chaotic-order-openai-gpt-image-2-medium-02Reviewed on the labeled refinement contact sheet against the unpublished operator reference. This run is retained as matrix evidence; model ranking is reported at campaign level.
oc26r-r5b-macro-chaotic-order-openai-gpt-image-2-medium-01Reviewed on the labeled refinement contact sheet against the unpublished operator reference. This run is retained as matrix evidence; model ranking is reported at campaign level.
Research queue
How consistently does one model preserve Herman’s identity across repeated unseeded trials?
Which failed constraints still produce useful or aesthetically strong artifacts?
How many simultaneous visual constraints survive in one generated composition?
Can anatomy and tool contact remain plausible around small mechanical parts?
Can the subject remain readable under low-key lighting without crushed detail?
Can brass, enamel, glass, paper, clay, and textiles remain visually distinct?
Bench rule 01
A result enters the public catalog only with a reproducible record and a content hash. Unknown values remain null; they are never invented.
Failures stay visible. Manual-only providers remain manual-only. Provider access rules are data, not an afterthought.