How do you monitor LLM quality drift in production?

// hourly cron: score a sample of the last hour's traces
const sample = await traces.sample({ lastHours: 1, rate: 0.03, feature: "support" });

for (const t of sample) {
  const score = await judge.score("faithfulness", t.context, t.output);
  await qualityTs.write({ feature: t.feature, model: t.model, score, at: t.at });
}

// alerting: 7-day rolling mean vs trailing 24h, per feature
// fires when delta > 2σ — catches provider-side shifts with no deploy marker
await alerts.evaluate("quality.faithfulness", { window: "24h", baseline: "7d", sigma: 2 });