diff --git a/src/agent/mcts.js b/src/agent/mcts.js
index 7de481e..0d2df4b 100644
--- a/src/agent/mcts.js
+++ b/src/agent/mcts.js
@@ -1,3 +1,8 @@
+/* Statistical-memo MCTS: this tree records which (state, action) pairs have been
+   explored and their cumulative rewards, but does NOT re-navigate the browser to
+   match the selected node before action execution. The tree bounds exploration via
+   UCB1 scores but does not enforce state-graph consistency. Full re-navigation
+   is deferred — it increases per-step cost 3-10x. See DECISION_LOG #021. */
 import { selectChild } from './policy.js';
 
 export class MctsNode {
diff --git a/src/index.js b/src/index.js
index 4d81dbd..7013dcd 100644
--- a/src/index.js
+++ b/src/index.js
@@ -306,7 +306,7 @@ async function runArm({ role, page, seed, config, rng, tracer, breadcrumbs, step
     const macroFireProb = config.macros?.fireProbability ?? 0;
     const macroList = config.macros?.list ?? [];
     const stepTimeoutMs = config.run?.stepTimeoutMs ?? 10000;
-    const recentStateIds = [];
+    const recentStateIds = new Set();
     const noveltyDenylist = (config.novelty?.nameDenylist ?? []).map((s) => new RegExp(s, 'i'));
     let prevA11y = null;
     let prevUrl = null;
@@ -321,7 +321,7 @@ async function runArm({ role, page, seed, config, rng, tracer, breadcrumbs, step
       const fileInputs = await getFileInputs(page.raw);
       const forms = await detectFillableForms(page.raw);
       const stateId = clusterId(a11y, config.mcts.abstractionGranularity);
-      recentStateIds.push(stateId);
+      recentStateIds.add(stateId);
       const cands = candidateActions(a11y, {
         weights: config.actions.weights,
         blockedSelectors: config.target.blockedSelectors,
diff --git a/src/llm/prompts.js b/src/llm/prompts.js
index 9b17a77..cc7b6cd 100644
--- a/src/llm/prompts.js
+++ b/src/llm/prompts.js
@@ -23,3 +23,67 @@ export function buildPredictPrompt({ a11yTree, action, recentActions = [] }) {
     .filter(Boolean)
     .join('\n');
 }
+
+export function buildSurprisePrompt({
+  prediction,
+  observed,
+  hardSignals = [],
+  recentActions = [],
+  recentStateIds = new Set(),
+  currentStateId = null,
+}) {
+  const observedJson = JSON.stringify(observed).slice(0, 4000);
+  const recentList = recentActions.length
+    ? recentActions.map((a, i) => `- ${i + 1}. ${a}`).join('\n')
+    : '- (none)';
+  const recentIdsView = recentStateIds.size
+    ? JSON.stringify([...recentStateIds].slice(-8))
+    : '[]';
+  const repeated = Boolean(currentStateId) && recentStateIds.has(currentStateId);
+
+  return [
+    '# QA Surprise Evaluator',
+    '',
+    'You score how surprising a UI action result is. Pick exactly one bucket from the rubric below. Do not blend, average, or interpolate.',
+    '',
+    '## SCORE RUBRIC — pick exactly one of {0.0, 0.2, 0.5, 0.8, 1.0}',
+    '',
+    '| Score | Use when |',
+    '|---|---|',
+    '| 0.0 | Observed state matches (or is a near-duplicate of) a recent state, OR there is no visible change beyond text/timestamps. Refresh / scroll / no-op loops are ALWAYS 0.0. |',
+    '| 0.2 | Predicted change happened but with no novelty: same interactive roles, same URL fragment, only inner text shifted within the same widgets. |',
+    '| 0.5 | New widgets appeared or disappeared — a role/name not present in any recent state. Same screen, different controls. |',
+    '| 0.8 | Distinct new screen reached: URL/route changed, a dialog/modal/drawer opened, or a new section is now interactive. |',
+    '| 1.0 | Page broke, threw, froze, or rendered an unexpected error visible from the tree alone (hard signals are scored separately by the harness). |',
+    '',
+    '## ANTI-LOOP RULES (strict — violating them is a wrong answer)',
+    '',
+    '1. **Repetition is never surprising.** If the current state cluster id appears in the recent state ids list, you MUST return 0.0.',
+    '2. **No hedging.** The only legal scores are 0.0, 0.2, 0.5, 0.8, 1.0. Never invent intermediate values to feel "safe".',
+    '3. **Novelty must be evidenced.** Any score ≥ 0.5 must name, in `reason`, the specific role / accessible name / URL fragment that was not present in recent states.',
+    '4. **Ignore prediction quality.** A vague or stub prediction (e.g. "the page will change") cannot justify surprise. Only the observed tree counts.',
+    '5. **Same-cluster refreshes score 0.0**, even if numeric counters, timestamps, or analytics pixels changed.',
+    '',
+    '## INPUTS',
+    '',
+    `Prediction: ${prediction}`,
+    '',
+    'Recent actions (oldest → newest):',
+    recentList,
+    '',
+    `Recent state cluster ids (most recent last): ${recentIdsView}`,
+    `Current state cluster id: ${currentStateId ?? 'unknown'}`,
+    `Repeated state detected by harness: ${repeated ? 'YES — score MUST be 0.0' : 'no'}`,
+    `Page-level hard signals fired: ${JSON.stringify(hardSignals)}`,
+    '',
+    'Observed accessibility tree (pruned, truncated to 4000 chars):',
+    observedJson,
+    '',
+    '## OUTPUT',
+    '',
+    'Reply with JSON on a single line, no markdown fence, no surrounding prose. Schema:',
+    '{"score": <0.0|0.2|0.5|0.8|1.0>, "reason": "<≤ 14 words; name the new element or write \\"repeat\\">"}',
+    '',
+    'JSON:',
+  ].join('\n');
+}
diff --git a/tests/unit/mcts.test.js b/tests/unit/mcts.test.js
index 1e75127..a820086 100644
--- a/tests/unit/mcts.test.js
+++ b/tests/unit/mcts.test.js
@@ -124,3 +124,22 @@ describe('descend', () => {
     expect(result).toBe(childA);
   });
 });
+
+describe('recentStateIds Set contract', () => {
+  it('does not duplicate on repeated add', () => {
+    const s = new Set();
+    s.add('state-a');
+    s.add('state-a');
+    expect(s.size).toBe(1);
+  });
+  it('.has() returns true for added items', () => {
+    const s = new Set();
+    s.add('state-1');
+    expect(s.has('state-1')).toBe(true);
+  });
+  it('.has() returns false for non-added items', () => {
+    const s = new Set();
+    s.add('state-1');
+    expect(s.has('state-99')).toBe(false);
+  });
+});