diff --git a/src/components/ClampText.css b/src/components/ClampText.css
index 1c320be..08de61e 100644
--- a/src/components/ClampText.css
+++ b/src/components/ClampText.css
@@ -57,3 +57,27 @@
 .clamp.is-open .clamp__more svg {
   transform: rotate(180deg);
 }
+
+/* The headline lifted out of an AI paragraph. Slightly heavier than the body
+   and never clamped — it is one sentence by construction, and the point of it
+   is that the finding can be read without opening anything. */
+.clamp__lead {
+  margin: 0;
+  font-size: 0.9rem;
+  line-height: 1.5;
+  color: var(--zk-ink);
+  font-weight: 500;
+}
+.clamp__lead + .clamp__text { margin-top: 8px; }
+
+/* The finding, lifted out of an AI paragraph. Heavier than the body and never
+   clamped — it is one sentence by construction, and the whole point is that it
+   can be read without opening anything. */
+.clamp__lead {
+  margin: 0;
+  font-size: 0.9rem;
+  line-height: 1.5;
+  font-weight: 500;
+  color: var(--zk-ink);
+}
+.clamp__lead + .clamp__text { margin-top: 8px; }
diff --git a/src/components/ClampText.jsx b/src/components/ClampText.jsx
index d0be5c1..4e559ff 100644
--- a/src/components/ClampText.jsx
+++ b/src/components/ClampText.jsx
@@ -1,15 +1,33 @@
 import { useState } from 'react'
 import './ClampText.css'
 
+/**
+ * Splits a block of AI prose into a finding and its working.
+ *
+ * The employees write the conclusion first and the evidence after it —
+ * "POSP-77341 is active and empanelled for motor." then four sentences of
+ * registry detail. So the first sentence already IS the summary, and lifting it
+ * out gives one without inventing text or asking the model for a second field
+ * it would have to be trusted to keep in step with the first.
+ *
+ * Returns null when the split would be useless: no sentence terminator, a first
+ * sentence so long it is the paragraph again, or one so short it is a fragment
+ * rather than a finding. Those fall back to plain folding.
+ */
+function splitLead(value) {
+  const m = value.match(/^(.{40,220}?[.!?])(\s+)(\S[\s\S]*)$/)
+  if (!m) return null
+  return { lead: m[1].trim(), rest: m[3].trim() }
+}
+
 /**
  * Prose, folded.
  *
  * The AI employees write at length — a recommendation rationale runs past 1,500
  * characters — and a lead file that prints all of it is a document, not a
- * screen. Anything long opens as a few lines with the rest one click away, so
- * the page can be scanned and still holds everything for whoever wants it.
+ * screen. Anything long shows its finding, with the reasoning one click away.
  *
- * Whether to fold is decided on length, not on measuring the rendered box: a
+ * Whether to fold is decided on length, not by measuring the rendered box: a
  * ref-and-measure pass would reflow on every resize to answer a question the
  * string itself already answers.
  */
@@ -19,11 +37,30 @@ export default function ClampText({ text, lines = 3, threshold = 150 }) {
 
   if (value.length <= threshold) return <p className="clamp__text clamp__text--short">{value}</p>
 
+  const split = splitLead(value)
+
+  // No sensible headline — fold the whole thing, as before.
+  if (!split) {
+    return (
+      <div className={'clamp' + (open ? ' is-open' : '')}>
+        <p className="clamp__text" style={{ '--clamp-lines': lines }}>{value}</p>
+        <button type="button" className="clamp__more" onClick={() => setOpen((o) => !o)}>
+          {open ? 'Show less' : 'Read more'}
+          <svg viewBox="0 0 12 12" aria-hidden="true">
+            <path d="M2.5 4.5 6 8l3.5-3.5" fill="none" stroke="currentColor" strokeWidth="1.5"
+                  strokeLinecap="round" strokeLinejoin="round" />
+          </svg>
+        </button>
+      </div>
+    )
+  }
+
   return (
     <div className={'clamp' + (open ? ' is-open' : '')}>
-      <p className="clamp__text" style={{ '--clamp-lines': lines }}>{value}</p>
+      <p className="clamp__lead">{split.lead}</p>
+      {open ? <p className="clamp__text clamp__text--short">{split.rest}</p> : null}
       <button type="button" className="clamp__more" onClick={() => setOpen((o) => !o)}>
-        {open ? 'Show less' : 'Read more'}
+        {open ? 'Hide the reasoning' : 'Show the reasoning'}
         <svg viewBox="0 0 12 12" aria-hidden="true">
           <path d="M2.5 4.5 6 8l3.5-3.5" fill="none" stroke="currentColor" strokeWidth="1.5"
                 strokeLinecap="round" strokeLinejoin="round" />
diff --git a/src/components/Timeline.jsx b/src/components/Timeline.jsx
index edbf460..97c2350 100644
--- a/src/components/Timeline.jsx
+++ b/src/components/Timeline.jsx
@@ -84,7 +84,59 @@ export default function Timeline({ instanceId }) {
   // Oldest first: a timeline reads forwards.
   const ordered = [...rows].sort((a, b) => String(a.created_at).localeCompare(String(b.created_at)))
 
-  const items = ordered.map((r, i) => {
+  // ONE SUBMISSION WRITES SEVERAL ROWS. The platform records each stage of a
+  // submission separately, told apart by execution_state:
+  //
+  //   zk-act-qualify   TRIGGER_PERFORMED     the trigger's commit
+  //   zk-act-qualify   TRIGGER_PERFORMED     ...again, on the settle path
+  //   zk-act-qualify   zk-state-qualified    the one that moved the lead
+  //
+  // Rendered literally that is "Qualify Lead" three times in a row, which
+  // reads as the AI having done the same thing three times. Only the row
+  // carrying a real state means anything to somebody reading the file.
+  //
+  // Collapsed on (same activity, within fifteen seconds) rather than on the
+  // execution_state alone, because a genuine repeat has to survive: Collect
+  // Documents really is performed twice on a lead whose first upload was
+  // incomplete, and those are minutes apart, not milliseconds.
+  //
+  // The kept row takes the EARLIEST timestamp — when the operator acted — and
+  // whichever state and payload is actually populated.
+  const isStage = (v) => STAGES.some((s) => s.uid === v)
+  const SAME_SUBMISSION_MS = 15000
+
+  const merged = []
+  for (const r of ordered) {
+    // Look BACK for a match rather than only at the previous entry: the
+    // platform interleaves a DATA_UPDATE row between the two halves of one
+    // submission, so the rows to merge are near each other in time but not
+    // adjacent in the list.
+    let at = -1
+    for (let i = merged.length - 1; i >= 0; i--) {
+      if (Date.parse(r.created_at) - Date.parse(merged[i].created_at) >= SAME_SUBMISSION_MS) break
+      if (merged[i].activity_id === r.activity_id) { at = i; break }
+    }
+
+    if (at >= 0) {
+      const prev = merged[at]
+      merged[at] = {
+        ...prev,
+        // The settle row is the one that names the resulting stage.
+        execution_state: isStage(r.execution_state) ? r.execution_state : prev.execution_state,
+        // Whichever row actually carries the submission and the AI's working.
+        data: (r.data && Object.keys(r.data).length) ? r.data : prev.data,
+        ai_reasoning: r.ai_reasoning || prev.ai_reasoning,
+        ai_confidence: r.ai_confidence ?? prev.ai_confidence,
+        user_name: prev.user_name || r.user_name,
+        user_roles: (prev.user_roles && prev.user_roles.length) ? prev.user_roles : r.user_roles,
+        created_at: prev.created_at,
+      }
+      continue
+    }
+    merged.push(r)
+  }
+
+  const items = merged.map((r, i) => {
     const roles = r.user_roles || []
     const aiRole = roles.find((x) => AI_ROLES[x])
     const isSystem = r.activity_id === 'DATA_UPDATE'