## Why This Capsule Exists Most image benchmarks test one thing at a time: counting, text rendering, or spatial reasoning. Real-world image generation fails happen when **multiple constraints compete for the model's attention simultaneously** — which is exactly what happens in production use cases. This capsule stacks 8 independent, objectively verifiable constraints into a single prompt to reveal how models handle compounded instruction-following pressure. ## What This Tests | Constraint | Capability Probed | |---|---| | Clock reads exactly 10:08 | Precise analog time rendering (a known weak spot across model generations) | | Glove on right hand only, left hand bare | Asymmetric instruction-following (resisting the default "matching pair" bias) | | Bare hand shows exactly 4 fingers | Exact small-number generation on body parts | | Mirror reflection matches reality (correct flip) | Geometric/spatial reasoning, not just pattern repetition | | Exactly 7 apples | Object counting accuracy at small scale | | Exactly 1 green apple | Basic attribute assignment (color) | | Sign reads "NO PHOTOS ALLOWED" | Legible text rendering without duplication/glitches | | No windows in the room | Negation handling (ignoring implied defaults) | ## Expected Results - **No model is expected to pass all 8 criteria.** A perfect score would be a signal to add harder constraints, not evidence the benchmark has failed. - **Weakest points across the board:** exact clock time and exact object counts — these fail even on otherwise photorealistic, high-quality outputs, showing that visual fidelity and instruction accuracy are separate capabilities. - **The mirror consistency check is the strongest signal for genuine spatial reasoning.** Models that just repeat the face without correctly flipping the glove side are pattern-matching "mirror = duplicate person," not simulating reflection physics. - **Asymmetric instructions (one gloved hand, one bare) are frequently collapsed into symmetric defaults** — this is a good proxy for how well a model resists its own training priors when they conflict with the prompt. ## How to Read a Run Score each run against the 8-point checklist below. Partial credit applies where an element is present but incorrect (e.g., wrong finger count, right time but wrong minute). This capsule is designed to produce **score dispersion across models and over time**, not a pass/fail verdict — the goal is watching the score trend upward as models improve, not expecting any single run to "win." ### Scoring Checklist - [ ] Clock shows exactly 10:08 - [ ] Glove on one hand only, other hand bare - [ ] Bare hand shows exactly 4 fingers - [ ] Mirror reflection is geometrically consistent (correct side flip) - [ ] Exactly 7 apples - [ ] Exactly 1 green apple - [ ] Sign text reads correctly, no duplication/glitches - [ ] No windows visible in the room