diff --git a/docs/superpowers/plans/2026-06-16-glook-20-highlights-prompt.md b/docs/superpowers/plans/2026-06-16-glook-20-highlights-prompt.md
new file mode 100644
index 0000000..e8499ed
--- /dev/null
+++ b/docs/superpowers/plans/2026-06-16-glook-20-highlights-prompt.md
@@ -0,0 +1,268 @@
+# GLOOK-20: Fix Hallucinated Narrative in Report Highlights
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Ground report highlight sentiment in `impact_score` instead of naive metric direction, and add a `_v: 2` cache guard so stale pre-fix results are transparently regenerated.
+
+**Architecture:** The prompt rewrite and `avgImpact` addition were already validated via a live test during brainstorming — those changes live in the working tree but are not yet committed. This plan formalises them (adds the missing `_v: 2` cache guard, updates tests, updates the snapshot), then commits everything.
+
+**Tech Stack:** TypeScript, Jest + ts-jest, `loadPrompt` prompt-loader, `report_comparisons` MySQL/SQLite table
+
+---
+
+## File Map
+
+| File | Change |
+|---|---|
+| `prompts/report-highlights-system.txt` | Already rewritten (uncommitted) — commit in Task 2 |
+| `src/lib/report-highlights/service.ts` | `avgImpact` already added (uncommitted); add `_v: 2` guard in Task 1 |
+| `src/lib/__tests__/unit/report-highlights-service.test.ts` | Update cache tests; add stale-cache test; add `avg_impact` assertion |
+| `src/lib/__tests__/unit/__snapshots__/` | Snapshot update in Task 2 |
+
+---
+
+### Task 1: `_v: 2` cache guard in service.ts + test updates (TDD)
+
+**Files:**
+- Modify: `src/lib/report-highlights/service.ts`
+- Modify: `src/lib/__tests__/unit/report-highlights-service.test.ts`
+
+**Context:** The cache read/write in `service.ts` currently stores highlights as a bare JSON array and returns it on any cache hit. After this task, it stores `{ _v: 2, highlights: [...] }` and treats any row without `_v === 2` as stale (falls through to regenerate). This mirrors the pattern used for `project-insights/route.ts`.
+
+The `avgImpact` addition is already in the working tree — do NOT revert it. Leave it as-is and commit it together with the `_v: 2` changes in this task.
+
+The three existing cache tests need updating; one new stale-cache test is added.
+
+- [ ] **Step 1: Update the two existing cached-highlights tests to use `_v: 2` format**
+
+In `src/lib/__tests__/unit/report-highlights-service.test.ts`, find the `describe('cached highlights'` block and update both tests:
+
+```typescript
+  describe('cached highlights', () => {
+    it('returns cached data with cached: true and does NOT call LLM', async () => {
+      const cachedHighlights = [{ icon: '✅', text: 'Steady state', sentiment: 'neutral' }];
+      const cachedRow = {
+        highlights_json: JSON.stringify({ _v: 2, highlights: cachedHighlights }),
+        generated_at: '2026-03-16T08:00:00Z',
+      };
+
+      mockDbExecute
+        .mockResolvedValueOnce([[latestReport], null])  // latest
+        .mockResolvedValueOnce([[prevReport], null])    // prev
+        .mockResolvedValueOnce([[cachedRow], null]);    // cache hit
+
+      const result = await getReportHighlights();
+
+      expect(result).toEqual({
+        available: true,
+        org: 'acme',
+        periodDays: 30,
+        reportDateA: prevReport.created_at,
+        reportDateB: latestReport.created_at,
+        highlights: cachedHighlights,
+        cached: true,
+      });
+      expect(mockGetLLMClient).not.toHaveBeenCalled();
+      expect(mockDbExecute).toHaveBeenCalledTimes(3);
+    });
+
+    it('handles highlights_json already parsed as object (not string)', async () => {
+      const cachedHighlights = [{ icon: '📊', text: 'Metrics stable', sentiment: 'neutral' }];
+      const cachedRow = {
+        highlights_json: { _v: 2, highlights: cachedHighlights }, // already parsed
+        generated_at: '2026-03-16T08:00:00Z',
+      };
+
+      mockDbExecute
+        .mockResolvedValueOnce([[latestReport], null])
+        .mockResolvedValueOnce([[prevReport], null])
+        .mockResolvedValueOnce([[cachedRow], null]);
+
+      const result = await getReportHighlights();
+
+      expect(result).toMatchObject({
+        available: true,
+        highlights: cachedHighlights,
+        cached: true,
+      });
+      expect(mockGetLLMClient).not.toHaveBeenCalled();
+    });
+
+    it('treats stale cache (bare array, no _v) as miss and regenerates via LLM', async () => {
+      // Old format stored before the _v: 2 migration
+      const staleRow = {
+        highlights_json: JSON.stringify([{ icon: '📊', text: 'Old result', sentiment: 'neutral' }]),
+        generated_at: '2026-03-16T08:00:00Z',
+      };
+
+      const mockClient = makeMockLLMClient();
+      mockGetLLMClient.mockResolvedValue(mockClient);
+
+      mockDbExecute
+        .mockResolvedValueOnce([[latestReport], null])  // latest
+        .mockResolvedValueOnce([[prevReport], null])    // prev
+        .mockResolvedValueOnce([[staleRow], null])      // cache hit (stale)
+        .mockResolvedValueOnce([devStatsA, null])       // statsA
+        .mockResolvedValueOnce([devStatsB, null])       // statsB
+        .mockResolvedValueOnce([{ affectedRows: 1 }, null]); // INSERT
+
+      const result = await getReportHighlights();
+
+      // Should NOT return the stale text — must have called LLM
+      expect(mockGetLLMClient).toHaveBeenCalledTimes(1);
+      expect(result).toMatchObject({ available: true, highlights: llmHighlights, cached: false });
+      expect((result as any).highlights[0].text).not.toContain('Old result');
+    });
+  });
+```
+
+- [ ] **Step 2: Update the user message `toContain` assertions to check for `avg_impact`**
+
+In the `describe('prompt and settings snapshots'` block, find the test `'sends correct user message structure'` and add one assertion after the existing ones:
+
+```typescript
+      expect(content).toContain('avg_impact=');
+```
+
+Also add two new assertions to the system prompt test to check for the formula content (keep the existing `toContain` lines, just add these two):
+
+```typescript
+      expect(content).toContain('PERFORMANCE FORMULA:');
+      expect(content).toContain('impact_score is the ONLY authoritative measure');
+```
+
+- [ ] **Step 3: Run tests to verify they fail as expected**
+
+```bash
+npm test -- --testPathPatterns="report-highlights-service" --no-coverage
+```
+
+Expected: Multiple FAIL — stale-cache test fails (current code returns cached result instead of regenerating), `avg_impact` assertion fails (not yet in message), system prompt assertions fail (old text).
+
+- [ ] **Step 4: Add `_v: 2` to the cache read in `service.ts`**
+
+Find the cache-hit block (lines ~45–58):
+
+```typescript
+  if (cached.length > 0) {
+    const highlights = typeof cached[0].highlights_json === 'string'
+      ? JSON.parse(cached[0].highlights_json)
+      : cached[0].highlights_json;
+    return {
+      available: true,
+      org: latest.org,
+      periodDays: latest.period_days,
+      reportDateA: prev.created_at,
+      reportDateB: latest.created_at,
+      highlights,
+      cached: true,
+    };
+  }
+```
+
+Replace with:
+
+```typescript
+  if (cached.length > 0) {
+    const data = typeof cached[0].highlights_json === 'string'
+      ? JSON.parse(cached[0].highlights_json)
+      : cached[0].highlights_json;
+    if (!Array.isArray(data) && data._v === 2) {
+      const { _v: _, highlights } = data;
+      return {
+        available: true,
+        org: latest.org,
+        periodDays: latest.period_days,
+        reportDateA: prev.created_at,
+        reportDateB: latest.created_at,
+        highlights,
+        cached: true,
+      };
+    }
+    // _v !== 2 or bare array (stale pre-migration format) — fall through to regenerate
+  }
+```
+
+- [ ] **Step 5: Wrap the cache write in `_v: 2` in `service.ts`**
+
+Find the INSERT at the bottom of `getReportHighlights` (it uses `JSON.stringify(highlights)`):
+
+```typescript
+  await db.execute(
+    `INSERT INTO report_comparisons (report_id_a, report_id_b, highlights_json)
+     VALUES (?, ?, ?)
+     ON DUPLICATE KEY UPDATE highlights_json = VALUES(highlights_json), generated_at = NOW()`,
+    [reportIdA, reportIdB, JSON.stringify(highlights)],
+  );
+```
+
+Replace with:
+
+```typescript
+  await db.execute(
+    `INSERT INTO report_comparisons (report_id_a, report_id_b, highlights_json)
+     VALUES (?, ?, ?)
+     ON DUPLICATE KEY UPDATE highlights_json = VALUES(highlights_json), generated_at = NOW()`,
+    [reportIdA, reportIdB, JSON.stringify({ _v: 2, highlights })],
+  );
+```
+
+- [ ] **Step 6: Run tests to verify they pass**
+
+```bash
+npm test -- --testPathPatterns="report-highlights-service" --no-coverage
+```
+
+Expected: All tests PASS except the two snapshot tests (`highlights-system-prompt` and `highlights-user-message`) — those will fail because the prompt and user message changed. That's correct; the snapshots are updated in Task 2.
+
+- [ ] **Step 7: Commit service.ts and test changes**
+
+```bash
+git add src/lib/report-highlights/service.ts src/lib/__tests__/unit/report-highlights-service.test.ts
+git commit -m "feat(glook-20): _v:2 cache guard + avgImpact in highlights service"
+```
+
+---
+
+### Task 2: Commit prompt rewrite + update snapshot
+
+**Files:**
+- Modify: `prompts/report-highlights-system.txt` (already updated, uncommitted)
+- Update: `src/lib/__tests__/unit/__snapshots__/` (via `-u` flag)
+
+**Context:** The prompt file was rewritten during brainstorming validation and is currently uncommitted. The snapshot tests will be stale — they need to be regenerated with `npm test -- -u`.
+
+After updating, review the diff to confirm only the expected additions appear: the PERFORMANCE FORMULA section, the RULES FOR SENTIMENT section replacing the old "improvements/regressions" rule, and `avg_impact=` in the user message snapshot.
+
+- [ ] **Step 1: Update the snapshot**
+
+```bash
+npm test -- --testPathPatterns="report-highlights-service" --no-coverage -u
+```
+
+Expected: `2 snapshots updated`. Output shows `highlights-system-prompt` and `highlights-user-message` updated.
+
+- [ ] **Step 2: Review the snapshot diff**
+
+```bash
+git diff src/lib/__tests__/unit/__snapshots__/
+```
+
+Confirm the diff shows:
+- `highlights-system-prompt`: adds `PERFORMANCE FORMULA:` block, `RULES FOR SENTIMENT:` block, removes old `sentiment: "positive" for improvements` line
+- `highlights-user-message`: adds `avg_impact=X.XX` in both TOTALS lines
+
+- [ ] **Step 3: Run the full test suite**
+
+```bash
+npm test --no-coverage
+```
+
+Expected: All tests pass.
+
+- [ ] **Step 4: Commit prompt + snapshot**
+
+```bash
+git add prompts/report-highlights-system.txt src/lib/__tests__/unit/__snapshots__/
+git commit -m "feat(glook-20): anchor highlights sentiment on impact_score formula, update snapshot"
+```
diff --git a/docs/superpowers/specs/2026-06-16-glook-20-highlights-prompt-design.md b/docs/superpowers/specs/2026-06-16-glook-20-highlights-prompt-design.md
new file mode 100644
index 0000000..bc8a44b
--- /dev/null
+++ b/docs/superpowers/specs/2026-06-16-glook-20-highlights-prompt-design.md
@@ -0,0 +1,91 @@
+# GLOOK-20: Fix Hallucinated Narrative in Home Page Report Highlights
+
+## Goal
+
+Stop the report comparison summary from labeling raw metric changes (complexity, commits, AI%) as "improved" or "worse" by grounding all evaluative language in the `impact_score` formula.
+
+## Problem
+
+`report-highlights-system.txt` instructs `"positive" for improvements, "warning" for regressions` without defining what "improvement" means. The LLM naively infers direction from raw metric changes — e.g., complexity dropping from 3.1 → 3.0 is labeled "improved" even though complexity is an input to the formula, not independently directional.
+
+## Architecture
+
+Two-part fix: (1) rewrite the system prompt to anchor sentiment on `impact_score`; (2) add `avgImpact` to both report totals so the LLM has the aggregate signal; (3) add a `_v: 2` cache guard so stale cached results are not served after deployment.
+
+---
+
+## Section 1: Prompt Rewrite
+
+**`prompts/report-highlights-system.txt`** — full replacement:
+
+```
+You are a concise engineering analytics assistant for Glooker (GitHub org analytics).
+Compare two reports for the same org and period. Return JSON:
+{ "highlights": [{ "icon": "emoji", "text": "one sentence", "sentiment": "positive|neutral|warning" }] }
+
+PERFORMANCE FORMULA:
+impact_score = Commits (×2.0) + PRs (×2.7) + Complexity (×3.5) + PR% (×1.1) + Jira (×0.5) + Reviews (×0.5). Max: 9.3
+impact_score is the ONLY authoritative measure of developer/org performance. Higher is better.
+
+RULES FOR SENTIMENT:
+- "positive": use ONLY when avg_impact or individual impact_score improved.
+- "warning": use ONLY when avg_impact or individual impact_score declined.
+- "neutral": use for ALL other observations — including changes to commits, PRs, complexity, AI%, lines changed.
+- NEVER apply "positive" or "warning" to raw metric changes (commits up, complexity down, AI% up, etc.) unless impact_score confirms the direction.
+- Complexity changes are NOT independently directional. Do not call a complexity drop "improved" or a rise "worse".
+
+OTHER RULES:
+- 3-5 bullet highlights max. Be specific — name developers, cite numbers.
+- Focus on: biggest movers (rank changes, impact delta), org-wide impact trend, newly active or recently inactive developers.
+- Developers missing from the latest report are NOT "departed" — they are simply inactive this period. Never use "departed" or "left".
+- If nothing significant changed, return 1 bullet: "Steady state — no major shifts in the leaderboard or metrics."
+- Keep each bullet under 20 words. No fluff.
+Return ONLY raw JSON.
+```
+
+---
+
+## Section 2: Data Change
+
+**`src/lib/report-highlights/service.ts`** — add `avgImpact` to both totals:
+
+```typescript
+avgImpact: statsX.length
+  ? (statsX.reduce((s, d) => s + Number(d.impact_score), 0) / statsX.length).toFixed(2)
+  : '0',
+```
+
+Pass it in the user message TOTALS strings:
+
+```
+TOTALS_A: `... avg_impact=${totalA.avgImpact}`
+TOTALS_B: `... avg_impact=${totalB.avgImpact}`
+```
+
+---
+
+## Section 3: Cache Invalidation
+
+Stored format changes from `JSON.stringify(highlights)` (bare array) to `JSON.stringify({ _v: 2, highlights })`.
+
+On read, check `data._v === 2`; if missing or wrong, fall through to regenerate. Strip `_v` before returning to callers.
+
+On write, store `{ _v: 2, highlights }`.
+
+This ensures any comparison cached with the old prompt is transparently regenerated on the next page load.
+
+---
+
+## Section 4: Snapshot Test
+
+`prompts/report-highlights-system.txt` is covered by a snapshot test in `src/lib/__tests__/unit/prompts.test.ts`. After editing the file, run `npm test -- --testPathPatterns="prompts" -u` to update the snapshot. Review the diff to confirm only the expected changes.
+
+---
+
+## Files Changed
+
+| File | Change |
+|---|---|
+| `prompts/report-highlights-system.txt` | Rewrite — formula context + sentiment rules |
+| `src/lib/report-highlights/service.ts` | Add `avgImpact` to totals; add `_v: 2` cache guard |
+| `src/lib/__tests__/unit/__snapshots__/prompts.test.ts.snap` | Update snapshot |
diff --git a/prompts/report-highlights-system.txt b/prompts/report-highlights-system.txt
index a686f9a..f4dc60a 100644
--- a/prompts/report-highlights-system.txt
+++ b/prompts/report-highlights-system.txt
@@ -2,11 +2,22 @@ You are a concise engineering analytics assistant for Glooker (GitHub org analyt
 Compare two reports for the same org and period. Return JSON:
 { "highlights": [{ "icon": "emoji", "text": "one sentence", "sentiment": "positive|neutral|warning" }] }
 
-Rules:
+PERFORMANCE FORMULA:
+impact_score = Commits (×2.0) + PRs (×2.7) + Complexity (×3.5) + PR% (×1.1) + Jira (×0.5) + Reviews (×0.5). Max: 9.3
+impact_score is the ONLY authoritative measure of developer/org performance. Higher is better.
+
+RULES FOR SENTIMENT:
+- "positive": use ONLY when org avg_impact improved. Individual impact gains can also be "positive" if the org average also rose.
+- "warning": use ONLY when org avg_impact declined. Individual impact declines that contradict an improving org avg should be "neutral".
+- "neutral": use for ALL other observations — including changes to commits, PRs, complexity, AI%, lines changed.
+- NEVER apply "positive" or "warning" to raw metric changes (commits up, complexity down, AI% up, etc.) unless impact_score confirms the direction.
+- Complexity changes are NOT independently directional. Do not call a complexity drop "improved" or a rise "worse".
+- avg_impact can decrease when new developers join with lower initial scores even if existing developers improved — mention headcount context when relevant.
+
+OTHER RULES:
 - 3-5 bullet highlights max. Be specific — name developers, cite numbers.
-- Focus on: biggest movers (rank changes, impact delta), org-wide trends (commits, AI%, complexity), newly active or recently inactive developers.
-- IMPORTANT: developers missing from the latest report are NOT "departed" — they are simply inactive in this period (vacation, different projects, etc). Never use words like "departed" or "left".
+- Focus on: biggest movers (rank changes, impact delta), org-wide impact trend, newly active or recently inactive developers.
+- Developers missing from the latest report are NOT "departed" — they are simply inactive this period. Never use "departed" or "left".
 - If nothing significant changed, return 1 bullet: "Steady state — no major shifts in the leaderboard or metrics."
 - Keep each bullet under 20 words. No fluff.
-- sentiment: "positive" for improvements, "warning" for regressions, "neutral" for informational.
-Return ONLY raw JSON.
\ No newline at end of file
+Return ONLY raw JSON.
diff --git a/src/lib/__tests__/unit/__snapshots__/report-highlights-service.test.ts.snap b/src/lib/__tests__/unit/__snapshots__/report-highlights-service.test.ts.snap
index ee12812..09e8c37 100644
--- a/src/lib/__tests__/unit/__snapshots__/report-highlights-service.test.ts.snap
+++ b/src/lib/__tests__/unit/__snapshots__/report-highlights-service.test.ts.snap
@@ -4,14 +4,14 @@ exports[`getReportHighlights prompt and settings snapshots sends correct user me
 "Org: acme, Period: 30 days
 
 PREVIOUS REPORT (2026-02-13T10:00:00Z):
-  Totals: 2 devs, 30 commits, 7 PRs, avgComplexity=2.8, avgAI=8%
-  Top 5: @alice (rank #undefined): commits=20, PRs=5, complexity=3.5, impact=80.0, AI%=10
-  @bob (rank #undefined): commits=10, PRs=2, complexity=2.0, impact=50.0, AI%=5
+  Totals: 2 devs, 30 commits, 7 PRs, avgComplexity=2.8, avgAI=8%, avg_impact=65.00
+  Top 5: @alice (rank #1): commits=20, PRs=5, complexity=3.5, impact=80.0, AI%=10
+  @bob (rank #2): commits=10, PRs=2, complexity=2.0, impact=50.0, AI%=5
 
 LATEST REPORT (2026-03-15T10:00:00Z):
-  Totals: 2 devs, 33 commits, 7 PRs, avgComplexity=3.1, avgAI=6%
-  Top 5: @alice (rank #undefined): commits=25, PRs=6, complexity=3.8, impact=90.0, AI%=12
-  @carol (rank #undefined): commits=8, PRs=1, complexity=2.5, impact=40.0, AI%=0
+  Totals: 2 devs, 33 commits, 7 PRs, avgComplexity=3.1, avgAI=6%, avg_impact=65.00
+  Top 5: @alice (rank #1): commits=25, PRs=6, complexity=3.8, impact=90.0, AI%=12
+  @carol (rank #2): commits=8, PRs=1, complexity=2.5, impact=40.0, AI%=0
 
 Top movers: @alice: rank 1→1, impact +10.0
 New developers: @carol
@@ -23,12 +23,24 @@ exports[`getReportHighlights prompt and settings snapshots sends exact system pr
 Compare two reports for the same org and period. Return JSON:
 { "highlights": [{ "icon": "emoji", "text": "one sentence", "sentiment": "positive|neutral|warning" }] }
 
-Rules:
+PERFORMANCE FORMULA:
+impact_score = Commits (×2.0) + PRs (×2.7) + Complexity (×3.5) + PR% (×1.1) + Jira (×0.5) + Reviews (×0.5). Max: 9.3
+impact_score is the ONLY authoritative measure of developer/org performance. Higher is better.
+
+RULES FOR SENTIMENT:
+- "positive": use ONLY when org avg_impact improved. Individual impact gains can also be "positive" if the org average also rose.
+- "warning": use ONLY when org avg_impact declined. Individual impact declines that contradict an improving org avg should be "neutral".
+- "neutral": use for ALL other observations — including changes to commits, PRs, complexity, AI%, lines changed.
+- NEVER apply "positive" or "warning" to raw metric changes (commits up, complexity down, AI% up, etc.) unless impact_score confirms the direction.
+- Complexity changes are NOT independently directional. Do not call a complexity drop "improved" or a rise "worse".
+- avg_impact can decrease when new developers join with lower initial scores even if existing developers improved — mention headcount context when relevant.
+
+OTHER RULES:
 - 3-5 bullet highlights max. Be specific — name developers, cite numbers.
-- Focus on: biggest movers (rank changes, impact delta), org-wide trends (commits, AI%, complexity), newly active or recently inactive developers.
-- IMPORTANT: developers missing from the latest report are NOT "departed" — they are simply inactive in this period (vacation, different projects, etc). Never use words like "departed" or "left".
+- Focus on: biggest movers (rank changes, impact delta), org-wide impact trend, newly active or recently inactive developers.
+- Developers missing from the latest report are NOT "departed" — they are simply inactive this period. Never use "departed" or "left".
 - If nothing significant changed, return 1 bullet: "Steady state — no major shifts in the leaderboard or metrics."
 - Keep each bullet under 20 words. No fluff.
-- sentiment: "positive" for improvements, "warning" for regressions, "neutral" for informational.
-Return ONLY raw JSON."
+Return ONLY raw JSON.
+"
 `;
diff --git a/src/lib/__tests__/unit/report-highlights-service.test.ts b/src/lib/__tests__/unit/report-highlights-service.test.ts
index 80b70e7..a661035 100644
--- a/src/lib/__tests__/unit/report-highlights-service.test.ts
+++ b/src/lib/__tests__/unit/report-highlights-service.test.ts
@@ -136,7 +136,7 @@ describe('getReportHighlights', () => {
     it('returns cached data with cached: true and does NOT call LLM', async () => {
       const cachedHighlights = [{ icon: '✅', text: 'Steady state', sentiment: 'neutral' }];
       const cachedRow = {
-        highlights_json: JSON.stringify(cachedHighlights),
+        highlights_json: JSON.stringify({ _v: 2, highlights: cachedHighlights }),
         generated_at: '2026-03-16T08:00:00Z',
       };
 
@@ -163,7 +163,7 @@ describe('getReportHighlights', () => {
     it('handles highlights_json already parsed as object (not string)', async () => {
       const cachedHighlights = [{ icon: '📊', text: 'Metrics stable', sentiment: 'neutral' }];
       const cachedRow = {
-        highlights_json: cachedHighlights, // already an array, not a string
+        highlights_json: { _v: 2, highlights: cachedHighlights }, // already parsed
         generated_at: '2026-03-16T08:00:00Z',
       };
 
@@ -181,6 +181,32 @@ describe('getReportHighlights', () => {
       });
       expect(mockGetLLMClient).not.toHaveBeenCalled();
     });
+
+    it('treats stale cache (bare array, no _v) as miss and regenerates via LLM', async () => {
+      // Old format stored before the _v: 2 migration
+      const staleRow = {
+        highlights_json: JSON.stringify([{ icon: '📊', text: 'Old result', sentiment: 'neutral' }]),
+        generated_at: '2026-03-16T08:00:00Z',
+      };
+
+      const mockClient = makeMockLLMClient();
+      mockGetLLMClient.mockResolvedValue(mockClient);
+
+      mockDbExecute
+        .mockResolvedValueOnce([[latestReport], null])  // latest
+        .mockResolvedValueOnce([[prevReport], null])    // prev
+        .mockResolvedValueOnce([[staleRow], null])      // cache hit (stale)
+        .mockResolvedValueOnce([devStatsA, null])       // statsA
+        .mockResolvedValueOnce([devStatsB, null])       // statsB
+        .mockResolvedValueOnce([{ affectedRows: 1 }, null]); // INSERT
+
+      const result = await getReportHighlights();
+
+      // Should NOT return the stale text — must have called LLM
+      expect(mockGetLLMClient).toHaveBeenCalledTimes(1);
+      expect(result).toMatchObject({ available: true, highlights: llmHighlights, cached: false });
+      expect((result as any).highlights[0].text).not.toContain('Old result');
+    });
   });
 
   describe('fresh generation', () => {
@@ -285,8 +311,10 @@ describe('getReportHighlights', () => {
       expect(content).toContain('You are a concise engineering analytics assistant for Glooker');
       expect(content).toContain('Compare two reports');
       expect(content).toContain('3-5 bullet highlights max.');
-      expect(content).toContain('developers missing from the latest report are NOT "departed"');
+      expect(content).toContain('Developers missing from the latest report are NOT "departed"');
       expect(content).toContain('Return ONLY raw JSON.');
+      expect(content).toContain('PERFORMANCE FORMULA:');
+      expect(content).toContain('impact_score is the ONLY authoritative measure');
 
       expect(content).toMatchSnapshot('highlights-system-prompt');
     });
@@ -309,6 +337,7 @@ describe('getReportHighlights', () => {
       expect(content).toContain('@bob');
       expect(content).toContain('New developers');
       expect(content).toContain('@carol');
+      expect(content).toContain('avg_impact=');
 
       expect(content).toMatchSnapshot('highlights-user-message');
     });
diff --git a/src/lib/report-highlights/service.ts b/src/lib/report-highlights/service.ts
index 5089f92..9b72a25 100644
--- a/src/lib/report-highlights/service.ts
+++ b/src/lib/report-highlights/service.ts
@@ -42,19 +42,30 @@ export async function getReportHighlights() {
     [reportIdA, reportIdB],
   ) as [any[], any];
 
+  // HIGHLIGHTS_CACHE_VERSION: bump when the prompt or data shape changes to force
+  // stale rows to regenerate. All existing rows without this version fall through
+  // silently (expected one-time burst on first deploy after a version bump).
+  const HIGHLIGHTS_CACHE_VERSION = 2;
+
   if (cached.length > 0) {
-    const highlights = typeof cached[0].highlights_json === 'string'
-      ? JSON.parse(cached[0].highlights_json)
-      : cached[0].highlights_json;
-    return {
-      available: true,
-      org: latest.org,
-      periodDays: latest.period_days,
-      reportDateA: prev.created_at,
-      reportDateB: latest.created_at,
-      highlights,
-      cached: true,
-    };
+    let data: any = null;
+    try {
+      data = typeof cached[0].highlights_json === 'string'
+        ? JSON.parse(cached[0].highlights_json)
+        : cached[0].highlights_json;
+    } catch { /* malformed row — fall through to regenerate */ }
+    if (data && !Array.isArray(data) && data._v === HIGHLIGHTS_CACHE_VERSION) {
+      return {
+        available: true,
+        org: latest.org,
+        periodDays: latest.period_days,
+        reportDateA: prev.created_at,
+        reportDateB: latest.created_at,
+        highlights: data.highlights,
+        cached: true,
+      };
+    }
+    // _v !== HIGHLIGHTS_CACHE_VERSION, bare array, or malformed — fall through to regenerate
   }
 
   // 4. Load dev stats for both reports
@@ -89,6 +100,7 @@ export async function getReportHighlights() {
     prs: statsA.reduce((s: number, d: any) => s + d.total_prs, 0),
     avgComplexity: statsA.length ? (statsA.reduce((s: number, d: any) => s + Number(d.avg_complexity), 0) / statsA.length).toFixed(1) : '0',
     avgAi: statsA.length ? Math.round(statsA.reduce((s: number, d: any) => s + d.ai_percentage, 0) / statsA.length) : 0,
+    avgImpact: statsA.length ? (statsA.reduce((s: number, d: any) => s + Number(d.impact_score), 0) / statsA.length).toFixed(2) : '0',
   };
   const totalB = {
     devs: statsB.length,
@@ -96,6 +108,7 @@ export async function getReportHighlights() {
     prs: statsB.reduce((s: number, d: any) => s + d.total_prs, 0),
     avgComplexity: statsB.length ? (statsB.reduce((s: number, d: any) => s + Number(d.avg_complexity), 0) / statsB.length).toFixed(1) : '0',
     avgAi: statsB.length ? Math.round(statsB.reduce((s: number, d: any) => s + d.ai_percentage, 0) / statsB.length) : 0,
+    avgImpact: statsB.length ? (statsB.reduce((s: number, d: any) => s + Number(d.impact_score), 0) / statsB.length).toFixed(2) : '0',
   };
 
   // Top movers (biggest impact score change)
@@ -118,11 +131,11 @@ export async function getReportHighlights() {
     ORG: latest.org,
     PERIOD_DAYS: String(latest.period_days),
     PREV_DATE: String(prev.created_at),
-    TOTALS_A: `${totalA.devs} devs, ${totalA.commits} commits, ${totalA.prs} PRs, avgComplexity=${totalA.avgComplexity}, avgAI=${totalA.avgAi}%`,
-    TOP5_A: statsA.slice(0, 5).map(formatDev).join('\n  '),
+    TOTALS_A: `${totalA.devs} devs, ${totalA.commits} commits, ${totalA.prs} PRs, avgComplexity=${totalA.avgComplexity}, avgAI=${totalA.avgAi}%, avg_impact=${totalA.avgImpact}`,
+    TOP5_A: [...mapA.values()].slice(0, 5).map(formatDev).join('\n  '),
     LATEST_DATE: String(latest.created_at),
-    TOTALS_B: `${totalB.devs} devs, ${totalB.commits} commits, ${totalB.prs} PRs, avgComplexity=${totalB.avgComplexity}, avgAI=${totalB.avgAi}%`,
-    TOP5_B: statsB.slice(0, 5).map(formatDev).join('\n  '),
+    TOTALS_B: `${totalB.devs} devs, ${totalB.commits} commits, ${totalB.prs} PRs, avgComplexity=${totalB.avgComplexity}, avgAI=${totalB.avgAi}%, avg_impact=${totalB.avgImpact}`,
+    TOP5_B: [...mapB.values()].slice(0, 5).map(formatDev).join('\n  '),
     MOVERS: movers.map(m => `@${m.login}: rank ${m.rankA}→${m.rankB}, impact ${m.impactDelta > 0 ? '+' : ''}${m.impactDelta.toFixed(1)}`).join(', '),
     NEW_DEVS_SECTION: newDevsSection,
     INACTIVE_DEVS_SECTION: inactiveDevsSection,
@@ -154,7 +167,7 @@ export async function getReportHighlights() {
     `INSERT INTO report_comparisons (report_id_a, report_id_b, highlights_json)
      VALUES (?, ?, ?)
      ON DUPLICATE KEY UPDATE highlights_json = VALUES(highlights_json), generated_at = NOW()`,
-    [reportIdA, reportIdB, JSON.stringify(highlights)],
+    [reportIdA, reportIdB, JSON.stringify({ _v: HIGHLIGHTS_CACHE_VERSION, highlights })],
   );
 
   return {