diff --git a/src/autonomy/coordinator.js b/src/autonomy/coordinator.js
index 12f11d5..a173e84 100644
--- a/src/autonomy/coordinator.js
+++ b/src/autonomy/coordinator.js
@@ -1,5 +1,28 @@
+const { EVENT_FAMILY_NAMES, normalizeEventType } = require('./calibration');
+
 const ALLOWED_DIRECTIONS = new Set(['positive', 'negative']);
 const ALLOWED_HORIZONS = new Set([1, 5, 10, 20, 30, 60, 90]);
+// The prompt used to say event_type was a 'stable_enum' and then never listed the
+// enum, so the model invented one label per event and production ended up with 201
+// distinct values. Same closed set the cohort key uses, so a label can never mean
+// one thing in the prompt and another in calibration.
+const ALLOWED_EVENT_TYPES = new Set(EVENT_FAMILY_NAMES);
+
+// An exact family is what we want. If the model ignores the list we try to salvage
+// the label through the same mapper calibration uses, and only give up when it is
+// unplaceable -- an explicit 'other' is a legitimate answer, unplaceable free text
+// is not, and the difference is what stops 'other' quietly becoming the bin again.
+function normalizeProposedEventType(raw, instrument) {
+  const value = String(raw || '').trim().toLowerCase().replace(/[\s-]+/g, '_');
+  if (ALLOWED_EVENT_TYPES.has(value)) return value;
+
+  const salvaged = normalizeEventType(raw);
+  if (salvaged !== 'other') {
+    console.warn(`[coordinator] ${instrument} event_type "${raw}" is not in the enum, mapped to "${salvaged}"`);
+    return salvaged;
+  }
+  throw new Error(`event_type must be one of ${EVENT_FAMILY_NAMES.join(', ')} (got "${raw}")`);
+}
 
 function normalizeProposal(raw, { informationCutoff, model = 'unknown', promptVersion = 'unknown' } = {}) {
   if (!raw || typeof raw !== 'object') throw new Error('coordinator output must be an object');
@@ -18,7 +41,7 @@ function normalizeProposal(raw, { informationCutoff, model = 'unknown', promptVe
     return {
       instrument,
       direction,
-      eventType: String(item.event_type || 'unknown').trim().toLowerCase(),
+      eventType: normalizeProposedEventType(item.event_type, instrument),
       causalChannel: item.causal_channel ? String(item.causal_channel).trim() : null,
       horizonDays,
       evidenceArticleIds: [...new Set(articleIds)],
diff --git a/test/autonomy.test.js b/test/autonomy.test.js
index de2cc8a..4c95002 100644
--- a/test/autonomy.test.js
+++ b/test/autonomy.test.js
@@ -272,3 +272,25 @@ test('order intents require the explicit instrument allowlist', () => {
   assert.equal(intent.side, 'buy');
   assert.equal(db.prepare('SELECT status FROM autonomy_order_intents').get().status, 'shadow');
 });
+
+test('event_type is held to the closed family enum', () => {
+  const base = { instrument: 'NVDA', direction: 'positive', horizon_days: 10, evidence_article_ids: [1] };
+  const typeOf = (eventType) => normalizeProposal({ predictions: [{ ...base, event_type: eventType }] }).predictions[0].eventType;
+
+  // the enum itself, in the shapes a model actually emits
+  assert.equal(typeOf('earnings'), 'earnings');
+  assert.equal(typeOf('Supply Chain'), 'supply_chain');
+  assert.equal(typeOf('M_AND_A'), 'm_and_a');
+
+  // "none of these fit" is a legitimate answer and has to survive
+  assert.equal(typeOf('other'), 'other');
+
+  // off-enum but placeable: salvaged onto the family calibration would have
+  // picked anyway, so we dont throw away a usable prediction over a label
+  assert.equal(typeOf('earnings_beat_q3'), 'earnings');
+  assert.equal(typeOf('ceo resignation'), 'leadership');
+
+  // unplaceable free text is the 201-distinct-values failure, and is rejected
+  assert.throws(() => typeOf('vibes_shifted'), /event_type must be one of/);
+  assert.throws(() => typeOf(''), /event_type must be one of/);
+});
diff --git a/workers/coordinatorWorker.js b/workers/coordinatorWorker.js
index 9f0b647..ef0b254 100644
--- a/workers/coordinatorWorker.js
+++ b/workers/coordinatorWorker.js
@@ -6,6 +6,7 @@ const { initAutonomySchema } = require('../src/autonomy/schema');
 const { leaseNextJob, completeJob, failJob } = require('../src/autonomy/jobs');
 const { callCoordinator } = require('../src/autonomy/llm');
 const { acceptProposal, recordRejectedProposal } = require('../src/autonomy/coordinator');
+const { EVENT_FAMILY_NAMES } = require('../src/autonomy/calibration');
 
 function sleep(ms) { return new Promise((resolve) => setTimeout(resolve, ms)); }
 
@@ -24,10 +25,12 @@ function buildPrompt(event, articles) {
     `[Evidence ${index + 1}] article_id=${article.id}\nTitle: ${article.title}\n${String(article.content || article.description || '').slice(0, 4000)}`
   ).join('\n\n---\n\n');
   return `Event title: ${event.title}\n\n${evidence}\n\nReturn JSON only in this shape:\n${JSON.stringify({ predictions: [{
-    instrument: 'NVDA', direction: 'positive|negative', event_type: 'stable_enum',
-    causal_channel: 'short description', horizon_days: 10,
-    evidence_article_ids: [123], invalidation_condition: 'condition',
-  }] }, null, 2)}\n\nUse only instruments and evidence directly supported by the articles. Return an empty predictions array when there is no clear, tradable hypothesis. Never include probabilities, returns, confidence, position sizes, or actions.`;
+    instrument: '<ticker supported by the articles>', direction: '<positive or negative>',
+    event_type: '<one value from the event_type list below>',
+    causal_channel: '<short description>', horizon_days: '<one of 1, 5, 10, 20, 30, 60, 90>',
+    evidence_article_ids: ['<article_id values from the evidence above>'],
+    invalidation_condition: '<what would falsify this>',
+  }] }, null, 2)}\n\nEvery value in that shape is a placeholder describing the field. Do not copy them. Choose instrument, direction and horizon_days from the evidence in front of you.\n\nevent_type must be exactly one of: ${EVENT_FAMILY_NAMES.join(', ')}. Pick the closest one. Use "other" only when none of them genuinely apply, and never invent a value outside this list.\n\nUse only instruments and evidence directly supported by the articles. Return an empty predictions array when there is no clear, tradable hypothesis. Never include probabilities, returns, confidence, position sizes, or actions.`;
 }
 
 async function runCoordinatorWorker({ archivePath, intelligencePath, workerId = `coordinator-${os.hostname()}-${process.pid}`, pollMs = 1000 } = {}) {
diff --git a/workers/replayWorker.js b/workers/replayWorker.js
index cce73ba..716eada 100644
--- a/workers/replayWorker.js
+++ b/workers/replayWorker.js
@@ -6,6 +6,7 @@ const { initAutonomySchema } = require('../src/autonomy/schema');
 const { enqueueJob, leaseNextJob, completeJob, failJob } = require('../src/autonomy/jobs');
 const { callCoordinator } = require('../src/autonomy/llm');
 const { acceptProposal, recordRejectedProposal } = require('../src/autonomy/coordinator');
+const { EVENT_FAMILY_NAMES } = require('../src/autonomy/calibration');
 
 function sleep(ms) { return new Promise((resolve) => setTimeout(resolve, ms)); }
 
@@ -30,10 +31,14 @@ function replayPrompt(article) {
   return `Historical evidence cutoff: ${article.effective_at}\n\n` +
     `[Evidence 1] article_id=${article.id}\nTitle: ${article.title || ''}\n${String(article.content || article.description || '').slice(0, 6000)}\n\n` +
     `Return JSON only in this shape:\n${JSON.stringify({ predictions: [{
-      instrument: 'NVDA', direction: 'positive|negative', event_type: 'stable_enum',
-      causal_channel: 'short description', horizon_days: 10, evidence_article_ids: [123],
-      invalidation_condition: 'condition',
+      instrument: '<ticker supported by the evidence>', direction: '<positive or negative>',
+      event_type: '<one value from the event_type list below>',
+      causal_channel: '<short description>', horizon_days: '<one of 1, 5, 10, 20, 30, 60, 90>',
+      evidence_article_ids: ['<article_id values from the evidence above>'],
+      invalidation_condition: '<what would falsify this>',
     }] }, null, 2)}\n\n` +
+    'Every value in that shape is a placeholder describing the field. Do not copy them. Choose instrument, direction and horizon_days from the evidence in front of you.\n\n' +
+    `event_type must be exactly one of: ${EVENT_FAMILY_NAMES.join(', ')}. Pick the closest one. Use "other" only when none of them genuinely apply, and never invent a value outside this list.\n\n` +
     'Use only this dated evidence. Return an empty predictions array when there is no clear, tradable hypothesis. Never include probabilities, returns, confidence, position sizes, or actions.';
 }