diff --git a/package.json b/package.json
index 247e04e..0b9d91e 100644
--- a/package.json
+++ b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/run-capsule",
-  "version": "0.3.0",
+  "version": "0.4.0",
   "description": "Turn any agent run's trace into a shareable video — code/terminal/screen/conversation capsule animations, recorded headless and uploaded to a temp link. Secrets are redacted before upload. Consumer of @tangle-network/agent-eval/storyboard.",
   "type": "module",
   "license": "MIT",
diff --git a/src/artifacts.ts b/src/artifacts.ts
index 19ee29f..c2f9310 100644
--- a/src/artifacts.ts
+++ b/src/artifacts.ts
@@ -99,6 +99,26 @@ export interface NarrationResult {
 /** A short voiceover script derived from the run: the ask, the key beats, the
  *  outcome. Kept terse — it's narration, not a transcript. When a verdict is
  *  given, it closes on the gate result so the VO matches the scoreboard shot. */
+/** Make a string safe to SPEAK: strip code fences, `var = value` assignments,
+ *  axis tags like "(X)", markdown punctuation and approximation tildes, then
+ *  keep just the first sentence. Without this the TTS reads raw source — e.g.
+ *  "H_wall = 60" becomes the meaningless "H is 60". Narration is prose, not code. */
+function speakable(s: string, maxLen = 150): string {
+  let t = s
+    .replace(/```[\s\S]*?```/g, ' ') // code blocks
+    .replace(/\b[A-Za-z_]\w*\s*=\s*-?\d[\d.]*\b/g, ' ') // var = number
+    .replace(/\(\s*[XYZ]\s*\)/gi, ' ') // axis tags (X) (Y) (Z)
+    .replace(/[`*_#>{}[\]|]/g, ' ') // markdown / code punctuation
+    .replace(/~/g, '') // approximation tilde
+    .replace(/\s+/g, ' ')
+    .trim()
+  const m = /^(.*?[.!?])(\s|$)/.exec(t) // first sentence only
+  if (m?.[1]) t = m[1]
+  if (t.length > maxLen) t = `${t.slice(0, maxLen).replace(/\s+\S*$/, '')}…`
+  return t.replace(/[.!?]+$/, '').trim() // drop trailing punctuation; templates add their own
+
+}
+
 export function buildNarrationScript(
   spans: readonly Span[],
   title: string,
@@ -106,18 +126,18 @@ export function buildNarrationScript(
 ): string {
   const ev = reduceToSemanticEvents(spans)
   const ask = ev.find((e) => e.kind === 'understood_task')?.summary
-  const reply = [...ev].reverse().find((e) => e.kind === 'agent_reply')?.summary
   const edits = ev.filter((e) => e.kind === 'edited_code').length
   const cmds = ev.filter((e) => e.kind === 'ran_command').length
   const fails = ev.filter((e) => e.kind === 'observed_failure').length
-  const parts: string[] = [`${title}.`]
-  if (ask) parts.push(`The task: ${ask}.`)
+  const parts: string[] = [`${speakable(title)}.`]
+  // The brief, cleaned to its first plain sentence — NOT the raw agent reply
+  // (which is source code and reads as gibberish through TTS).
+  if (ask) parts.push(`The task: ${speakable(ask)}.`)
   const did: string[] = []
   if (edits) did.push(`${edits} code ${edits === 1 ? 'edit' : 'edits'}`)
   if (cmds) did.push(`${cmds} ${cmds === 1 ? 'command' : 'commands'}`)
   if (fails) did.push(`recovering from ${fails} ${fails === 1 ? 'failure' : 'failures'}`)
   if (did.length) parts.push(`The agent worked through ${did.join(', ')}.`)
-  if (reply) parts.push(reply)
   if (result) {
     const total = Object.keys(result.checks).length
     const passed = Object.values(result.checks).filter(Boolean).length
diff --git a/src/renderers.test.ts b/src/renderers.test.ts
index ee9c575..d84fc0d 100644
--- a/src/renderers.test.ts
+++ b/src/renderers.test.ts
@@ -274,6 +274,20 @@ describe('buildNarrationScript (closes on the verdict when scored)', () => {
     ]
     expect(buildNarrationScript(spans, 'House')).not.toMatch(/geometry gate/i)
   })
+
+  it('speaks clean prose — never raw code: no "H_wall = 60", no axis tags, drops the source reply', () => {
+    // Regression for the 3/10 VO: TTS was reading the brief's "(X)/(Y)" + the
+    // agent's raw .scad reply, so it said the meaningless "H is 60".
+    const spans: Span[] = [
+      { spanId: 'b', runId: 'r', kind: 'llm', name: 'brief', model: 'm', messages: [{ role: 'user', content: 'Write OpenSCAD for a two-story house. Requirements: footprint 80 (X) by 60 (Y), H_wall = 60 units.' }], startedAt: 0, endedAt: 1, status: 'ok' } as Span,
+      { spanId: 'a', runId: 'r', kind: 'llm', name: 'reply', model: 'm', messages: [{ role: 'user', content: 'x' }], output: 'W=80; D=60; H_wall=60; difference(){ cube([W,D,H_wall]); }', startedAt: 2, endedAt: 3, status: 'ok' } as Span,
+    ]
+    const vo = buildNarrationScript(spans, 'Agent designs a house', { resolved: true, score: 1, checks: { a: true, b: true } })
+    expect(vo).not.toMatch(/\([XYZ]\)|H_wall|W\s*=\s*80|=\s*60/) // no raw code / axis tags
+    expect(vo).not.toContain('difference(') // the source reply is dropped, not spoken
+    expect(vo).toContain('two-story house') // the cleaned brief survives
+    expect(vo).toContain('2 of 2 checks') // verdict still lands
+  })
 })
 
 describe('redactSpans (P0: never publish a live credential)', () => {