Aesthetic Prompt Governance

# Aesthetic And Prompt Governance

## v2.0 Quality Thesis

The design must be mathematically inspected before it becomes visually spectacular. Prevent dirty, fragmented, overfilled visuals by locking project context, selecting relevant memory cases, scoring candidate directions, measuring residual risk, and exposing route/candidate decisions through `math_trace`.

## Root Causes To Catch

- Fragmented visuals: too many small decorative pieces, random icons, scattered labels, background debris, and no dominant focal anchor.
- Dirty visuals: muddy texture noise, overmixed palettes, low-contrast type, uncontrolled grain, and fake aging effects.
- Weak prompts: generic words such as "高级", "大气", "好看", or "丰富" without concrete composition, material, light, typography, and output rules.
- Text failures: long in-image copy, pseudo-text, misspelling, warped letters, mixed language without hierarchy, and mojibake.
- Layout disorder: no grid, no margin system, no reading path, too many modules, no density ceiling, and overlapping text.
- Redundant mechanisms: multiple agents owning the same decision, repeated QA sections, repeated negative prompts, and unclear handoff owner.
- Context mixing: commercial conversion logic applied to academic competition boards, or research-board density applied to premium packaging/product visuals.

## Mathematical Gates

Before image generation or final delivery, require:

- `route.math`: softmax route probability, entropy, probability margin, and confidence.
- `memory.math_trace`: cosine, jaccard, taxonomy similarity, taxonomy prior, and top case scores.
- `constraints.penalty_vector`: risk load, mitigation strength, residual risk, and constraint satisfaction.
- `candidate_optimization`: Pareto/TOPSIS/weighted-utility ranking.
- `critic_aggregation`: weighted critic score.
- `failure_memory`: relevant failed modes and similarity score.

## Hard Gates

Before image generation or final delivery, require:

- One dominant visual anchor and no more than two secondary supports.
- One explicit project context and 1-3 relevant memory cases, or a statement that no suitable memory case exists.
- Explicit grid, margin, alignment, reading path, and negative space.
- Density ceiling: name what stays empty, quiet, or visually subordinate.
- Exact visible text: spelling, language, hierarchy, max lines, and no pseudo-text.
- UTF-8/no-mojibake check when Chinese or mixed-language text is present.
- Anti-fragmentation negative prompt: no scattered tiny decorations, no dirty texture noise, no random icons, no warped type, no fake logos, no unresolved placeholders.
- One owner each for context routing, case-memory selection, QA, and delivery; remove duplicated mechanisms.

## PromptPacketV2 Contract

PromptPacketV2 must include: