## Rationale This capsule follows directly from the Terra & Kiln series finding: giving models abstract creative freedom ("invent your own design system") converges toward a shared, consensus-safe aesthetic rather than genuine diversity — measured directly across 7 models on that capsule (near-identical terracotta/cream/sage palettes, quantified via HSL analysis). Two structural changes address this: * **Naming concrete design movements and demanding full commitment** replaces an open-ended creative prompt (which collapses to the safest answer) with a forced choice among (or beyond) named, established visual languages. Which movement each model gravitates to — or invents — becomes a trend signal in itself. * **A color-neutral product brief** (an SF/fantasy bookshop, rather than ceramics/clay) removes the strong lexical color-priming that likely drove convergence last time. A first exploratory pair of runs (v5) validated this: two models produced genuinely distinct, fully-committed outputs (self-named "Terminal-Brutalism / Cyber-Y2K" vs. "Maximalist Editorial — Cosmic Cut-Up"), both scoring 15/16, with no dilution back into generic modern-minimalism. Both runs shared one recurring minor defect — CSS custom properties declared and used but omitted from the visible documentation — which v6 addresses directly. ## Expected spread * **Frontier models:** full, confident commitment to a distinctive movement (from the list or self-named), complete and accurate token documentation, possibly additional documented components beyond the four required categories. * **Mid-tier models:** may pick a movement but water it down with safe modern-minimalist defaults halfway through, or document tokens incompletely. * **Weaker models:** declared movement not visually recognizable in execution, broken responsive behavior, or real book/author leakage. ## Known cheats / risks * Declaring a movement in the documentation that doesn't actually match the executed design (label says "Brutalism," renders as generic soft-shadow modern UI). * Incomplete token documentation — a CSS custom property declared and visibly used somewhere on the page but missing from the swatch/type/spacing inventory (recurring pattern across the last 4 runs tested: Gemini, Grok, and both v5 Nova & Ink runs). * Dead tokens — a custom property declared in `:root` but never actually used anywhere (seen once so far: `--color-black` in the Cosmic Cut-Up run). * Using the JS/animation freedom as a loophole to import an external library via CDN. * Real book title/author leakage despite the explicit instruction — direct fail, IP risk regardless of other scores. * Hardcoded hex/spacing values outside the token system, especially in hover/focus states. ## Grading Truth (GT) | Criterion | Pass | Partial | Fail | | :--- | :--- | :--- | :--- | | **GT1:** Renders with no console errors, zero external network requests | ok | minor non-blocking errors | blocking error or external request | | **GT2:** Native `<details>` present, closed by default, positioned before footer, opens without JS | compliant | position/label slightly off | doc outside `<details>` or requires JS | | **GT3a:** Token documentation completeness — every used token is documented, no dead tokens | complete inventory, no dead tokens | 1–2 used tokens undocumented, or 1 dead token left in | several undocumented tokens, or documentation is decorative only | | **GT3b:** Token fidelity — no hardcoded values betraying the declared system, including hover/focus | zero hardcodes outside the system | 1 isolated hardcode | repeated hardcodes, system ignored in practice | | **GT4:** A named design movement is explicitly stated in the documentation | clearly named | vaguely implied, not explicitly labeled | no movement named at all | | **GT5:** Movement commitment — the page's actual execution genuinely and consistently reflects the named movement's visual vocabulary | strong, consistent execution | partial execution, waters down into generic modern style | movement named but not visually recognizable at all | | **GT6:** Required sections present, in order, no real book/author/franchise names used | all present, fully fictional | 1 section weak/incomplete | 2+ missing, or real IP referenced (automatic fail regardless of other criteria) | | **GT7:** Responsive at 375/768/1440 with no overflow or broken elements | clean at all 3 sizes | 1 size degraded | broken at 2+ sizes, or content visibly lost/truncated (treat as hard Fail regardless of formula, per the Minimax precedent) | | **GT8:** Copy consistent with brand values (wonder, discovery, community) | clearly integrated | vaguely evoked | generic copy | ## Scoring 18 points max (9 criteria × 2, GT3 now split into 3a/3b). Same proportional tiers, rescaled: * **S: 17–18** — flawless commitment and documentation, zero cheats * **A: 14–16** — solid, 1 minor weakness * **B: 10–13** — functional but noticeable gaps in documentation or execution * **C: 5–9** — missing sections or detected cheat * **F: 0–4** — unrenderable, major constraint violation, or real IP leak ## New analytical angle for this capsule Beyond the tier score, tag each run with which movement it chose (from the list, or "invented: [name]"). Across enough runs, this becomes genuinely interesting trend data — does each model have a consistent aesthetic "personality," and does that personality drift over time on the Timeline. This is arguably closer to the series' original "trend observatory" goal than color-palette tracking alone, since movement choice is a much stronger creative signal than which shade of terracotta a model picks.