AI 辅助技术美术 · source map

AI × TA 工作台：候选不是交付：代码定位

这里显示课程实际引用的代码片段。先看每段在说人话，再回到实验验证它对画面的影响。

候选输入约束

候选先对齐平台、预算和成功指标，再谈模型信心。

src/lib/lab/production-case/case-source.ts · section frame

当前起点查看这段精确实现

// SECTION:frame
function frameDecision(brief) {
  return {
    platform: brief.platform,
    budgetMs: brief.targetMs,
    deadline: brief.deadlineWeeks,
    successMetric: brief.successMetric,
  };
}

人工 Gate

QA 不通过或人工未批准时，状态必须是 blocked。

src/lib/lab/production-case/case-source.ts · section ai-gate

按需展开查看这段精确实现

// SECTION:ai-gate
function publishCandidate(candidate, report, humanReview) {
  if (!report.passed || !humanReview.approved) return { status: 'blocked' };
  return { status: 'ready', provenance: candidate.source, rollbackToken: report.rollbackToken };
}

交付证据

先做 baseline 和单变量证据，再决定是否写回。

src/lib/lab/production-case/case-source.ts · section evidence

按需展开查看这段精确实现

// SECTION:evidence
function buildEvidencePlan(strategy) {
  return [
    'baseline on representative content',
    'change one variable and capture debug view',
    'measure target device and record failure boundary',
  ];
}

风险取舍

AI 建议也要放进质量、性能、交付和风险的同一张表。

src/lib/lab/production-case/case-source.ts · section tradeoffs

按需展开查看这段精确实现

// SECTION:tradeoffs
function scoreStrategy(strategy, weights, riskBudget) {
  const quality = strategy.quality * weights.quality;
  const performance = strategy.performance * weights.performance;
  const delivery = strategy.delivery * weights.delivery;
  const riskPenalty = Math.max(0, strategy.risk - riskBudget) * 0.38;
  return quality + performance + delivery - riskPenalty;
}