Given product context, return scored style candidates with reason codes, rejected styles with reasons, secondary pairings, assumptions and confidence. Evidence-provider: the host model makes the final pick. Call before writing UI. The "tone" field (named "tone", not "tones") is an ARRAY of zero or more of: calm, technical, trustworthy, premium, editorial, playful, bold, experimental. Each candidate.score already includes soft penalties (density/motion/tone fit) that are not itemized in scoreBreakdown.

Compare 2-4 directions across product fit, repeated-use suitability, density, trust, distinctiveness, accessibility risk, motion and maintenance. Returns each direction's favorable conditions and likely failure mode — never a single winner.

Select a small set of practical visual-design principles for intent and iteration, semantic hierarchy, adaptive layout and density, typography and localization, tokens and themes, accessible interaction, and complete states and recovery. Filter by design concerns, surface, phase, or explicit ids. Returns placement guidance, application steps, verification checks, cautions, related principles, and deterministic relevance scores.

Select a small, evidence-labeled set of UX principles for a target outcome, surface, design phase, or explicit principle ids. Returns design questions, application steps, verification checks, cautions, references, and relevance scores. Treat the result as decision prompts rather than universal laws.

For a surface (data-table, form, checkout, chat, developer-console), return required / recommended / domain-specific states with triggers, must-show, must-not, accessibility and motion guidance, plus an implementation order. Covers the non-happy-path.

Compile a starting set of role-based design tokens (color, typography, spacing, radius, motion, density) for a style in json / css-variables / tailwind / typescript, with light/dark/both modes, accent override, and WCAG contrast warnings.