[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"capsule:4e6f0094-0ca2-4acc-8e46-7b4921afa14e":3},["Reactive",4],{"id":5,"user_id":6,"title":7,"prompt_content":8,"prompt_type":9,"is_locked":10,"is_public":10,"fork_of":11,"created_at":12,"updated_at":13,"category":14,"brief":15,"attachments":16,"expected_markers":23,"user":25},"4e6f0094-0ca2-4acc-8e46-7b4921afa14e","695c8e6c-9654-4858-b2d3-925adefacc35","Reference Scene 01","Describe everything you see in this image in as much detail as possible.\nInclude: any visible text exactly as written, the exact time shown on any clock, exact counts of any repeated objects broken down by color, the position of any object relative to others (left\u002Fright, above\u002Fbelow), and any visual emphasis such as borders, highlights, or size differences.\nDo not skip small details.","text",true,null,"2026-07-04T14:47:34.60731+00:00","2026-07-04T15:04:06.288+00:00","research","# Reference Scene 01 — Ground Truth & Capsule Brief\n\n## Why This Image (Not an AI-Generated Photo)\n\nUnlike the Mirror Torture Test, this reference image is **constructed as vector graphics with every value fixed by design** — not generated by an image model and hoped to be correct. This matters because a perception\u002Freading capsule needs a ground truth that is 100% certain. Using an AI-generated photo as \"ground truth\" is risky (as proven by the Mirror Torture Test results, where even the reference images had clock\u002Fcount errors). This SVG-based scene removes that risk entirely: every fact below is true by construction, not by inspection.\n\n## Ground Truth (exact, non-negotiable)\n\n| Element | Exact Fact |\n|---|---|\n| Digital clock | Reads **09:15** exactly |\n| Sign text | Reads **\"ZONE 7 \u002F AUTHORIZED ONLY\"** (two lines) |\n| Star row | **8 stars total**, left to right: blue, blue, yellow, blue, red, blue, yellow, blue |\n| Star color counts | **5 blue, 2 yellow, 1 red** |\n| Red star position | **5th star from the left** |\n| Numbered circles | Three circles reading **12, 47, 83** left to right |\n| Highlighted circle | The **middle circle (47)** has a thick gold border; the other two have thin gray borders |\n| Left mini-scene | A **yellow sun** positioned **above** a **gray mountain** |\n| Right mini-scene | A **white cloud** positioned **above** a **dark red house** |\n\n## What This Tests\n\n| Constraint | Capability Probed |\n|---|---|\n| Exact digital time transcription | Baseline reading accuracy (no analog-clock ambiguity, isolates pure OCR) |\n| Exact two-line text transcription | Multi-line text reading without merging or reordering lines |\n| Star color counting (5\u002F2\u002F1 split) | Counting under multiple competing categories at once, not just a single count |\n| Ordinal position (\"5th from left\") | Sequential\u002Fordinal reasoning, not just presence\u002Fabsence |\n| Number transcription (12, 47, 83) | Precise multi-digit numeral reading |\n| Identifying the highlighted element | Noticing a visual emphasis cue (thick gold border) among near-identical elements |\n| Above\u002Fbelow spatial relationships (sun\u002Fmountain, cloud\u002Fhouse) | Basic vertical spatial reasoning |\n| Left\u002Fright scene placement | Horizontal spatial reasoning across the full frame |\n\n## Expected Results\n\n- **Digital clock and sign text should be near-universal successes** — clean, high-contrast, unambiguous glyphs are the easiest case for any vision-capable model. A failure here is a strong negative signal.\n- **The star color breakdown is the most likely failure point** — asking a model to hold three simultaneous counts (5\u002F2\u002F1) and also locate the odd one out ordinally (\"5th from the left\") stacks two capabilities most models handle separately but rarely together.\n- **The highlighted circle (gold border) is a subtlety test** — models tend to report the three numbers correctly but skip noting which one is emphasized, since \"border thickness\u002Fcolor\" is a lower-salience detail compared to the number itself.\n- **Spatial relationships (above\u002Fbelow, left\u002Fright) are expected to be reliable** for well-separated, simple shapes like this — this capsule intentionally keeps spatial reasoning easy so that any failure here is a stronger, more meaningful signal than in a cluttered photo-realistic scene.\n\n## How to Read a Run\n\nGrade the model's description against the ground truth table above — no image inspection needed by the evaluator, since the ground truth is fixed and already documented. Each claim is right, wrong, or missing.\n\n### Scoring Checklist\n- [ ] Clock time stated as exactly \"09:15\"\n- [ ] Sign text transcribed correctly (both lines, correct wording)\n- [ ] Total star count stated as 8\n- [ ] Star color breakdown correct (5 blue, 2 yellow, 1 red)\n- [ ] Red star's ordinal position correctly identified (5th from left)\n- [ ] All three numbers (12, 47, 83) transcribed correctly\n- [ ] Correctly identifies circle \"47\" as the highlighted\u002Femphasized one\n- [ ] Correctly states sun is above the mountain (left side)\n- [ ] Correctly states cloud is above the house (right side)\n- [ ] No hallucinated elements not present in the image",[17],{"url":18,"kind":19,"mime":20,"name":21,"size":22},"https:\u002F\u002Fqoyuqtiafuqqyarrnsxr.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcapsule-files\u002F695c8e6c-9654-4858-b2d3-925adefacc35\u002F2dff39eb-05a6-427a-9be7-954bf85b6b1f.png","image","image\u002Fpng","reference-scene-01.png",55637,{"items":24,"matchAll":10},[],{"username":26,"avatar_url":27},"Ezarwebmaster","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F42354978?v=4"]