Compare employee pay against their own baseline, not raw rank; tighten output

Two real problems from live testing:
1. gatherEmployeeAnomalies() ranked by absolute cost, so it always
   surfaced senior/supervisory/longer-shift staff — that's their
   normal rate, not an anomaly. Now compares each employee's £/shift
   this week against their own trailing 4-week average and only
   reports the deviation; the prompt explicitly tells the model not to
   flag high pay in absolute or relative terms, only genuine
   deviations from someone's own baseline. Validated: Jack Evans
   (previously flagged 3 briefings running) is +0.5% vs his own
   baseline — not an anomaly at all — while Joseph Trice-Rolph's +41%
   swing is a genuine standout.
2. Output was hitting the token cap and cutting off mid-sentence
   (confirmed: last generation used exactly 900/900 output tokens).
   Raised cap to 1400, and tightened the system prompt to a hard
   5-bullet-total limit with no per-department/per-employee sections,
   since the model was writing a full structured report instead of a
   short briefing regardless of the token budget.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

backend/src/jobs/ai-insights.js

@@ -6,7 +6,7 @@ import { forecastDayCost } from '../lib/forecast.js'
 
 export const DEFAULT_MODEL = 'claude-haiku-4-5-20251001'
 export const DEFAULT_DAILY_TOKEN_BUDGET = 5000
-const MAX_OUTPUT_TOKENS = 900
+const MAX_OUTPUT_TOKENS = 1400
 export const MANUAL_RATE_LIMIT_MINUTES = 5
 
 function toISODate(d) {
@@ -290,17 +290,23 @@ export async function gatherForecastData(monthProgress) {
   }
 }
 
-// Individual shifts/wages worth a look — top-cost employees over the trailing week.
+// Individual shifts/wages worth a look — over the trailing week, compared against each
-export async function gatherEmployeeAnomalies(days = 7) {
+// employee's OWN trailing 4-week average cost-per-shift (not ranked by raw cost). Ranking by
+// absolute cost or cost-per-shift alone always surfaces senior/supervisory/longer-shift staff,
+// since they're legitimately paid more — that's not an anomaly, it's their normal rate. What's
+// actually worth flagging is a employee costing notably more than THEY usually do.
+export async function gatherEmployeeAnomalies(days = 7, baselineDays = 28) {
   const costCol = await getCostCol()
   const today = new Date()
   const fromStr = toISODate(new Date(today.getTime() - days * 86_400_000))
   const toStr = toISODate(new Date(today.getTime() - 86_400_000))
+  const baselineFromStr = toISODate(new Date(today.getTime() - (days + baselineDays) * 86_400_000))
+  const baselineToStr = toISODate(new Date(today.getTime() - (days + 1) * 86_400_000))
 
   // wage_actuals_detail only stores department_id (a raw Workforce code, e.g. "972312"),
   // not a human-readable name — resolve it from wage_actuals, which has both, so the
   // prompt (and the model) never has to guess which department a code refers to.
-  const [res, deptNamesRes] = await Promise.all([
+  const [res, baselineRes, deptNamesRes] = await Promise.all([
     pool.query(
       `SELECT employee_id, employee_name, department_id,
               SUM(${costCol}) AS cost, SUM(shift_count)::int AS shift_count
@@ -311,22 +317,44 @@ export async function gatherEmployeeAnomalies(days = 7) {
        LIMIT 15`,
       [fromStr, toStr]
     ),
+    pool.query(
+      `SELECT employee_id, SUM(${costCol}) AS cost, SUM(shift_count)::int AS shift_count
+       FROM wage_actuals_detail
+       WHERE date >= $1 AND date <= $2
+       GROUP BY employee_id`,
+      [baselineFromStr, baselineToStr]
+    ),
     pool.query(`SELECT DISTINCT department_id, department_name FROM wage_actuals`),
   ])
 
   const deptNames = Object.fromEntries(deptNamesRes.rows.map(r => [r.department_id, r.department_name]))
+  const baselinePerShift = Object.fromEntries(
+    baselineRes.rows
+      .filter(r => r.shift_count > 0)
+      .map(r => [r.employee_id, parseFloat(r.cost) / r.shift_count])
+  )
 
   return {
     fromStr,
     toStr,
-    employees: res.rows.map(r => ({
+    baselineFromStr,
-      employee_id: r.employee_id,
+    baselineToStr,
-      employee_name: r.employee_name,
+    employees: res.rows.map(r => {
-      department_id: r.department_id,
+      const cost = parseFloat(r.cost)
-      department_name: deptNames[r.department_id] || r.department_id,
+      const costPerShift = r.shift_count > 0 ? cost / r.shift_count : 0
-      cost: parseFloat(r.cost),
+      const baseline = baselinePerShift[r.employee_id] ?? null
-      shift_count: r.shift_count,
+      return {
-    })),
+        employee_id: r.employee_id,
+        employee_name: r.employee_name,
+        department_id: r.department_id,
+        department_name: deptNames[r.department_id] || r.department_id,
+        cost,
+        shift_count: r.shift_count,
+        costPerShift,
+        baselinePerShift: baseline,
+        deviationPct: baseline ? (costPerShift - baseline) / baseline : null,
+      }
+    }),
   }
 }
 
@@ -351,11 +379,12 @@ export async function gatherRevenueCorrelation(monthProgress) {
 export function buildPrompt(monthProgress, priorPeriod, rotaVsActual, forecast, anomalies, revenueCorrelation, recentInsights = [], manualContext = '') {
   const systemMsg =
     "You are an AI assistant for a wage cost controller in a UK hospitality business. Analyze the data below and " +
-    "provide a concise daily briefing (3-5 bullet points). Focus on: how this month's wage cost is tracking " +
+    "produce a SHORT daily briefing: hard limit of 5 bullet points TOTAL across the entire response, each a single " +
-    "against budget, the month-end forecast, notable variance between rota and actual cost by department and " +
+    "sentence (max ~30 words). Do not create a separate section or heading per department or per employee — pick " +
-    "what might be causing it, individual pay anomalies worth a look, and wage cost as a percentage of revenue. " +
+    "only the 4-5 most important things overall (across budget tracking, forecast, rota variance, pay anomalies, " +
-    "Use UK date format (DD/MM/YYYY) and GBP (£) for all monetary values. Be specific with department/employee " +
+    "wage % of revenue) and drop the rest; a department or employee with nothing notable gets no mention at all. " +
-    "names and numbers. Keep it actionable — no fluff or generic advice.\n\n" +
+    "Use UK date format (DD/MM/YYYY) and GBP (£). Be specific with names and numbers in the bullets you do write. " +
+    "No headings, no numbered action-plan section, no closing summary — just the bullets.\n\n" +
     "A 'Manual Context' section may be included below, written by a human who knows things the data can't show " +
     "(e.g. an employee's contract type, a known one-off cause, a planned change). Treat it as ground truth and " +
     "apply it directly — don't flag something as an anomaly or overspend if this context already explains it.\n\n" +
@@ -423,9 +452,23 @@ export function buildPrompt(monthProgress, priorPeriod, rotaVsActual, forecast,
   lines.push('')
 
   lines.push(`## Notable Individual Shifts/Wages (${fmtDateUK(anomalies.fromStr)} - ${fmtDateUK(anomalies.toStr)})`)
-  lines.push('Employee | Department | Cost | Shifts')
+  lines.push(
+    "'vs own avg' compares this week's £/shift to that employee's own trailing 4-week average " +
+    "(" + fmtDateUK(anomalies.baselineFromStr) + " - " + fmtDateUK(anomalies.baselineToStr) + "). Do NOT flag pay as " +
+    "anomalous just because it's high in absolute terms or relative to colleagues — role, seniority, and " +
+    "shift length legitimately vary pay, and a supervisor on a long shift will always cost more than a " +
+    "junior on a short one. Only flag a genuine deviation from that employee's OWN baseline (e.g. +20% or " +
+    "more), a large deviation with no baseline (new/rare worker), or something the Manual Context doesn't " +
+    "already explain — do not list someone here just because they're at the top of the cost column."
+  )
+  lines.push('Employee | Department | Cost | Shifts | £/shift | own avg £/shift | vs own avg')
   for (const e of anomalies.employees) {
-    lines.push(`${e.employee_name} | ${e.department_name} | ${fmtMoney(e.cost)} | ${e.shift_count}`)
+    const baselineStr = e.baselinePerShift != null ? fmtMoney(e.baselinePerShift) : 'no baseline'
+    const devStr = e.deviationPct != null ? `${e.deviationPct >= 0 ? '+' : ''}${(e.deviationPct * 100).toFixed(0)}%` : '-'
+    lines.push(
+      `${e.employee_name} | ${e.department_name} | ${fmtMoney(e.cost)} | ${e.shift_count} | ` +
+      `${fmtMoney(e.costPerShift)} | ${baselineStr} | ${devStr}`
+    )
   }
   lines.push('')