diff --git a/docs/relaunch/DECISIONS.md b/docs/relaunch/DECISIONS.md
index ff5978d..6f7c01b 100644
--- a/docs/relaunch/DECISIONS.md
+++ b/docs/relaunch/DECISIONS.md
@@ -3,6 +3,12 @@
 Agents append one entry per tie-break decision (01-product.md §0.4). Newest first.
 Format: `YYYY-MM-DD · WP-XX · decision · why · alternatives considered`.
 
+- 2026-10-27 · WP-07 · deterministic voice learn/preview added at the lib/ai boundary (`learnVoiceProfileDeterministic`/`previewVoiceDeterministic` in `lib/ai/voice.ts`): learn tries the §8.1 provider chain first, and only when every provider fails derives the profile extractively from the pasted samples (person from first-person pronouns, sentence length from average words/sentence, emoji from pictograph counts); fields a lexicon-free analyzer cannot honestly infer (tone, formality/energy/humor, signature phrases, ctaStyle) stay at the §7.4 defaults, and both the learn response (`status:"deterministic"`) and the UI ("Basic profile/version — our AI writer was unavailable.") label the result · U3's ≤15 s budget must hold with no usable AI keys; hard rule: never fake a learned/AI profile when no AI ran — labelled basic, not silent default · returning the plain §7.4 default with no label; forking WP-04's learn loop
+- 2026-10-27 · WP-07 · learn/preview provider calls cap the timeout at 12 s (`VOICE_TIMEOUT_MS`) instead of §8.1's 60 s · both are interactive (a user is waiting on step 2 / the preview card) and U3 budgets 15 s end-to-end; 60 s targets background generation · keeping the 60 s default
+- 2026-10-27 · WP-07 · Skip never writes anything: a null `shows.voice_profile` resolves to the §7.4 default at read time via `resolveShowVoiceProfile()` (unit-tested field-for-field) · one source of truth for the default, and nothing false is persisted after Skip · writing the default profile on Skip
+- 2026-10-27 · WP-07 · tone stays exactly 3 chips (§7.4 `tone[3]`): removing a chip disables Continue/Save voice until a replacement is added ("Add a tone word.") · the WP-04 schema fixes the length and silently padding would fabricate a value · relaxing the schema to 1–3
+- 2026-10-27 · WP-07 · learned samples persist to `shows.voice_samples` as raw pasted text; the MK-8 samples list therefore shows a text snippet + word count instead of MK-8's user labels ("Ep 39 show notes") · §7.1 has no label column and adding one is out of MVP scope · adding a labels schema field
+
 - 2026-10-27 · WP-05 · Live ASR smoke test skipped: `.env.relaunch` holds sanitized placeholders (values literally `"[SENSITIVE]"`, no `gsk_` key, ASR_PROVIDER ≠ groq) · the handoff rule says skip the live call when ASR_PROVIDER≠groq; the Groq adapter is covered by fetch-injected unit tests (request shape, error mapping, language gate) instead · running a live call with a placeholder key
 - 2026-10-27 · WP-05 · Chunk manifest (path + offsetMs per ≤10-min chunk) serialized as JSON into `episodes.audio_blob_path` · §7.1 has a single text column and the server must re-run transcription idempotently after a refresh without client state (U4); a plain single-path string still parses as one chunk at 0 ms · new table/column (schema is WP-02's), body-only manifest (lost on refresh)
 - 2026-10-27 · WP-05 · POST /api/episodes accepts `blobPaths[]` + parallel `chunkOffsetsMs[]` beyond the §7.5 body shape · §8.3 requires per-chunk offsets for the overlap dedupe and §7.5 pins `blobPaths[]`; parallel arrays are the smallest coherent extension · nested `chunks[]` objects in create (used only on transcribe), path-embedded offsets
diff --git a/web/app/(app)/app/onboarding/voice/page.tsx b/web/app/(app)/app/onboarding/voice/page.tsx
index 93f92a4..19368fa 100644
--- a/web/app/(app)/app/onboarding/voice/page.tsx
+++ b/web/app/(app)/app/onboarding/voice/page.tsx
@@ -1,26 +1,22 @@
-import Link from "next/link";
+import { redirect } from "next/navigation";
+import { VoiceOnboardingForm } from "@/components/app/onboarding/voice-form";
 import { StepDots } from "@/components/app/step-dots";
+import { requireUser } from "@/lib/auth";
+import { isDbConfigured } from "@/lib/env";
+import { getShowByOwnerSafe } from "@/lib/show/queries-safe";
 
-export const metadata = { title: "Onboarding · EpisodeKit" };
+export const metadata = { title: "Teach EpisodeKit your voice · EpisodeKit" };
 
-/**
- * Step 2 placeholder — the voice-learning UI ships in WP-07 (U3).
- * The route exists so the onboarding flow is complete end to end.
- */
-export default function OnboardingVoicePage() {
+/** Onboarding step 2 (M1): voice samples → profile (U3, MK-2). */
+export default async function OnboardingVoicePage() {
+  const userId = await requireUser();
+  const show = await getShowByOwnerSafe(userId);
+  if (!show && isDbConfigured()) redirect("/app/onboarding/show");
   return (
     <div className="mx-auto flex w-full max-w-xl flex-col items-center gap-6 py-10">
       <StepDots step={2} />
-      <div className="w-full rounded-lg border p-6 text-center" data-testid="voice-placeholder">
-        <h1 className="text-lg font-semibold">Teach EpisodeKit your voice</h1>
-        <p className="mt-2 text-sm text-neutral-500">This page is coming soon.</p>
-        <Link
-          href="/app/onboarding/first"
-          className="mt-4 inline-block text-sm underline"
-          data-testid="skip-voice"
-        >
-          Skip for now
-        </Link>
+      <div className="w-full rounded-lg border p-6">
+        <VoiceOnboardingForm />
       </div>
     </div>
   );
diff --git a/web/app/(app)/app/settings/voice/page.tsx b/web/app/(app)/app/settings/voice/page.tsx
index 2a1166d..5ab0a76 100644
--- a/web/app/(app)/app/settings/voice/page.tsx
+++ b/web/app/(app)/app/settings/voice/page.tsx
@@ -1,11 +1,21 @@
+import { redirect } from "next/navigation";
+import { VoiceSettingsForm } from "@/components/app/settings/voice-form";
+import { requireUser } from "@/lib/auth";
+import { isDbConfigured } from "@/lib/env";
+import { getShowByOwnerSafe } from "@/lib/show/queries-safe";
+import { resolveShowVoiceProfile } from "@/lib/show/voice";
+
 export const metadata = { title: "Settings · EpisodeKit" };
 
-/** Voice settings ship in WP-07 (M9 Voice, MK-8). */
-export default function SettingsVoicePage() {
+/** Settings → Voice (M9, MK-8): §7.4 profile editor, samples, preview. */
+export default async function SettingsVoicePage() {
+  const userId = await requireUser();
+  const show = await getShowByOwnerSafe(userId);
+  if (!show && isDbConfigured()) redirect("/app/onboarding/voice");
   return (
-    <div className="rounded-lg border p-6" data-testid="voice-stub">
-      <h2 className="text-base font-semibold">Voice</h2>
-      <p className="mt-1 text-sm text-neutral-500">This page is coming soon.</p>
-    </div>
+    <VoiceSettingsForm
+      profile={resolveShowVoiceProfile(show)}
+      samples={show?.voiceSamples ?? []}
+    />
   );
 }
diff --git a/web/app/api/show/voice/learn/route.ts b/web/app/api/show/voice/learn/route.ts
new file mode 100644
index 0000000..bfd7c16
--- /dev/null
+++ b/web/app/api/show/voice/learn/route.ts
@@ -0,0 +1,14 @@
+// §7.5 POST /api/show/voice/learn — onboarding step 2 (U3).
+import { getAuthUserId } from "@/lib/auth";
+import { AUTH_REQUIRED_COPY } from "@/lib/copy";
+import { handleVoiceLearn } from "@/lib/show/voice";
+
+export async function POST(req: Request) {
+  const userId = await getAuthUserId();
+  if (!userId) {
+    return Response.json({ error: AUTH_REQUIRED_COPY }, { status: 401 });
+  }
+  const body: unknown = await req.json().catch(() => null);
+  const result = await handleVoiceLearn(userId, body);
+  return Response.json(result.json, { status: result.status });
+}
diff --git a/web/app/api/show/voice/preview/route.ts b/web/app/api/show/voice/preview/route.ts
new file mode 100644
index 0000000..ff5bc39
--- /dev/null
+++ b/web/app/api/show/voice/preview/route.ts
@@ -0,0 +1,14 @@
+// §7.5 POST /api/show/voice/preview — Settings → Voice preview (M9/MK-8).
+import { getAuthUserId } from "@/lib/auth";
+import { AUTH_REQUIRED_COPY } from "@/lib/copy";
+import { handleVoicePreview } from "@/lib/show/voice";
+
+export async function POST(req: Request) {
+  const userId = await getAuthUserId();
+  if (!userId) {
+    return Response.json({ error: AUTH_REQUIRED_COPY }, { status: 401 });
+  }
+  const body: unknown = await req.json().catch(() => null);
+  const result = await handleVoicePreview(userId, body);
+  return Response.json(result.json, { status: result.status });
+}
diff --git a/web/components/app/onboarding/voice-form.tsx b/web/components/app/onboarding/voice-form.tsx
new file mode 100644
index 0000000..bbc1e5a
--- /dev/null
+++ b/web/components/app/onboarding/voice-form.tsx
@@ -0,0 +1,228 @@
+"use client";
+
+// Onboarding step 2 (M1, MK-2, U3): paste 1–3 past show notes (≥150 words
+// each), "Learn my voice" → chips/sliders, or "Skip for now" → the §7.4
+// default stays (null profile). The deterministic no-AI answer is labelled;
+// keyless previews say saving is unavailable instead of faking a save.
+import { useRouter } from "next/navigation";
+import { useMemo, useState } from "react";
+import type { VoiceProfile } from "@/lib/ai/voice";
+import {
+  SAVE_UNAVAILABLE_COPY,
+  SAMPLE_TOO_SHORT_COPY,
+  VOICE_BASIC_PROFILE_COPY,
+} from "@/lib/copy";
+import { MIN_SAMPLE_WORDS, countWords } from "@/lib/words";
+import { Button } from "@/components/ui/button";
+import { Textarea } from "@/components/ui/textarea";
+import { profileComplete, VoiceProfileEditor } from "@/components/app/voice/profile-editor";
+import { VoicePreviewCard } from "@/components/app/voice/preview-card";
+
+type LearnResponse = {
+  status?: "learned" | "deterministic";
+  profile?: VoiceProfile;
+  saved?: boolean;
+  error?: string;
+};
+
+export function VoiceOnboardingForm() {
+  const router = useRouter();
+  const [samples, setSamples] = useState<string[]>([""]);
+  const [phase, setPhase] = useState<"collect" | "learned">("collect");
+  const [profile, setProfile] = useState<VoiceProfile | null>(null);
+  const [learnSaved, setLearnSaved] = useState(false);
+  const [deterministic, setDeterministic] = useState(false);
+  const [learning, setLearning] = useState(false);
+  const [continuing, setContinuing] = useState(false);
+  const [error, setError] = useState<string | null>(null);
+
+  const filled = useMemo(
+    () => samples.map((s) => s.trim()).filter((s) => s.length > 0),
+    [samples],
+  );
+  const canLearn = filled.length > 0 && filled.every((s) => countWords(s) >= MIN_SAMPLE_WORDS);
+
+  function updateSample(index: number, value: string) {
+    setSamples((current) => current.map((s, i) => (i === index ? value : s)));
+  }
+
+  async function learn() {
+    if (learning || !canLearn) return;
+    setError(null);
+    setLearning(true);
+    try {
+      const res = await fetch("/api/show/voice/learn", {
+        method: "POST",
+        headers: { "content-type": "application/json" },
+        body: JSON.stringify({ samples: filled }),
+      });
+      const data = (await res.json().catch(() => ({}))) as LearnResponse;
+      if (!res.ok || !data.profile) {
+        setError(data.error ?? "Something went wrong. Try again.");
+        return;
+      }
+      setProfile(data.profile);
+      setLearnSaved(data.saved === true);
+      setDeterministic(data.status === "deterministic");
+      setPhase("learned");
+    } catch {
+      setError("Something went wrong. Try again.");
+    } finally {
+      setLearning(false);
+    }
+  }
+
+  async function continueToFirstEpisode() {
+    if (!profile || continuing) return;
+    setContinuing(true);
+    try {
+      if (learnSaved) {
+        const res = await fetch("/api/show", {
+          method: "PATCH",
+          headers: { "content-type": "application/json" },
+          body: JSON.stringify({ voiceProfile: profile }),
+        });
+        if (!res.ok) {
+          const data = (await res.json().catch(() => ({}))) as { error?: string };
+          if (res.status === 503) {
+            setError(SAVE_UNAVAILABLE_COPY);
+          } else {
+            setError(data.error ?? "Something went wrong. Try again.");
+          }
+          setContinuing(false);
+          return;
+        }
+      }
+      router.push("/app/onboarding/first");
+    } catch {
+      setError("Something went wrong. Try again.");
+      setContinuing(false);
+    }
+  }
+
+  if (phase === "learned" && profile) {
+    return (
+      <div className="flex flex-col gap-6" data-testid="voice-learned">
+        <div className="flex flex-col gap-1">
+          <h1 className="text-[22px] font-semibold leading-[30px]">Here&apos;s your voice</h1>
+          <p className="text-small text-body">
+            Adjust anything. You can change this later in Settings.
+          </p>
+        </div>
+
+        {deterministic ? (
+          <span className="rounded-md bg-warning-soft px-3 py-2 text-small text-warning" data-testid="voice-note">
+            {VOICE_BASIC_PROFILE_COPY}
+          </span>
+        ) : null}
+        {!learnSaved ? (
+          <span className="rounded-md bg-surface-2 px-3 py-2 text-small text-body" data-testid="voice-saved-note">
+            {SAVE_UNAVAILABLE_COPY}
+          </span>
+        ) : null}
+
+        <VoiceProfileEditor profile={profile} onChange={setProfile} />
+        <VoicePreviewCard profile={profile} autoLoad />
+
+        {error ? (
+          <span role="alert" className="text-small text-danger" data-testid="voice-error">
+            {error}
+          </span>
+        ) : null}
+
+        <div className="flex items-center justify-between">
+          <Button
+            type="button"
+            variant="ghost"
+            size="sm"
+            onClick={() => router.push("/app/onboarding/show")}
+            data-testid="back-to-show"
+          >
+            ← Back
+          </Button>
+          <Button
+            type="button"
+            onClick={continueToFirstEpisode}
+            disabled={!profileComplete(profile) || continuing}
+            data-testid="continue-voice"
+          >
+            {continuing ? "Saving…" : "Continue"}
+          </Button>
+        </div>
+      </div>
+    );
+  }
+
+  return (
+    <div className="flex flex-col gap-6" data-testid="voice-collect">
+      <div className="flex flex-col gap-1">
+        <h1 className="text-[22px] font-semibold leading-[30px]">Teach EpisodeKit your voice</h1>
+        <p className="text-small text-body">
+          Paste 1–3 past show notes or newsletters. We&apos;ll match how you write.
+        </p>
+      </div>
+
+      {samples.map((sample, index) => {
+        const words = countWords(sample);
+        return (
+          <div key={index} className="flex flex-col gap-1">
+            <label htmlFor={`voice-sample-${index}`} className="text-small font-medium">
+              Example {index + 1}
+            </label>
+            <Textarea
+              id={`voice-sample-${index}`}
+              rows={6}
+              value={sample}
+              onChange={(e) => updateSample(index, e.target.value)}
+              data-testid={`voice-sample-${index}`}
+            />
+            <div className="flex items-center justify-between">
+              <span
+                className="font-mono text-xs tabular-nums text-faint"
+                aria-live="polite"
+                data-testid={`sample-counter-${index}`}
+              >
+                {words} words · at least {MIN_SAMPLE_WORDS}
+              </span>
+              {index === samples.length - 1 && samples.length < 3 ? (
+                <button
+                  type="button"
+                  className="text-small underline"
+                  onClick={() => setSamples((current) => [...current, ""])}
+                  data-testid={`add-sample-${index}`}
+                >
+                  + Add another example
+                </button>
+              ) : null}
+            </div>
+          </div>
+        );
+      })}
+
+      {error ? (
+        <span role="alert" className="text-small text-danger" data-testid="voice-error">
+          {error}
+        </span>
+      ) : null}
+      {filled.length > 0 && !canLearn ? (
+        <span role="status" className="text-small text-warning" data-testid="voice-min-hint">
+          {SAMPLE_TOO_SHORT_COPY}
+        </span>
+      ) : null}
+
+      <div className="flex items-center justify-between">
+        <Button type="button" onClick={learn} disabled={!canLearn || learning} data-testid="learn-voice">
+          {learning ? "Learning your voice…" : "Learn my voice"}
+        </Button>
+        <button
+          type="button"
+          className="text-small underline"
+          onClick={() => router.push("/app/onboarding/first")}
+          data-testid="skip-voice"
+        >
+          Skip for now
+        </button>
+      </div>
+    </div>
+  );
+}
diff --git a/web/components/app/settings/voice-form.tsx b/web/components/app/settings/voice-form.tsx
new file mode 100644
index 0000000..2c63433
--- /dev/null
+++ b/web/components/app/settings/voice-form.tsx
@@ -0,0 +1,186 @@
+"use client";
+
+// Settings → Voice (M9, MK-8): left — profile editor + writing samples +
+// [Save voice]; right — preview card with [Update preview]. Everything saves
+// through PATCH /api/show (voiceProfile + voiceSamples), validated by the
+// shared §7.4 schema server-side.
+import { useState } from "react";
+import type { VoiceProfile } from "@/lib/ai/voice";
+import { SAVE_UNAVAILABLE_COPY } from "@/lib/copy";
+import { MIN_SAMPLE_WORDS, countWords } from "@/lib/words";
+import { Button } from "@/components/ui/button";
+import { Textarea } from "@/components/ui/textarea";
+import { profileComplete, VoiceProfileEditor } from "@/components/app/voice/profile-editor";
+import { VoicePreviewCard } from "@/components/app/voice/preview-card";
+
+type SaveResponse = { show?: unknown; error?: string };
+
+function snippet(text: string): string {
+  const firstLine = text.trim().split("\n")[0] ?? "";
+  return firstLine.length > 44 ? `${firstLine.slice(0, 44).trimEnd()}…` : firstLine;
+}
+
+export function VoiceSettingsForm({
+  profile: initialProfile,
+  samples: initialSamples,
+}: {
+  profile: VoiceProfile;
+  samples: string[];
+}) {
+  const [profile, setProfile] = useState(initialProfile);
+  const [samples, setSamples] = useState(initialSamples);
+  const [adding, setAdding] = useState(false);
+  const [draft, setDraft] = useState("");
+  const [saving, setSaving] = useState(false);
+  const [saved, setSaved] = useState(false);
+  const [error, setError] = useState<string | null>(null);
+
+  const draftWords = countWords(draft);
+  const draftValid = draftWords >= MIN_SAMPLE_WORDS;
+
+  function addDraft() {
+    if (!draftValid) return;
+    setSamples((current) => [...current, draft.trim()]);
+    setDraft("");
+    setAdding(false);
+    setSaved(false);
+  }
+
+  function removeSample(index: number) {
+    setSamples((current) => current.filter((_, i) => i !== index));
+    setSaved(false);
+  }
+
+  async function save() {
+    if (saving) return;
+    setError(null);
+    setSaved(false);
+    setSaving(true);
+    try {
+      const res = await fetch("/api/show", {
+        method: "PATCH",
+        headers: { "content-type": "application/json" },
+        body: JSON.stringify({ voiceProfile: profile, voiceSamples: samples }),
+      });
+      const data = (await res.json().catch(() => ({}))) as SaveResponse;
+      if (res.status === 503) {
+        setError(SAVE_UNAVAILABLE_COPY);
+        return;
+      }
+      if (!res.ok) {
+        setError(data.error ?? "Something went wrong. Try again.");
+        return;
+      }
+      setSaved(true);
+    } catch {
+      setError("Something went wrong. Try again.");
+    } finally {
+      setSaving(false);
+    }
+  }
+
+  return (
+    <div className="grid items-start gap-6 md:grid-cols-[minmax(0,1fr)_320px]">
+      <div className="flex flex-col gap-6 rounded-lg border p-5" data-testid="voice-settings-editor">
+        <VoiceProfileEditor profile={profile} onChange={setProfile} showAlwaysInclude />
+
+        <div className="flex flex-col gap-2" data-testid="voice-samples">
+          <span className="text-small font-medium">Writing samples</span>
+          {samples.map((sample, index) => (
+            <div
+              key={`${index}-${sample.slice(0, 12)}`}
+              className="flex items-center justify-between gap-3 rounded-md border px-3 py-2 text-small"
+              data-testid={`voice-sample-row-${index}`}
+            >
+              <span className="min-w-0 truncate text-body">
+                {snippet(sample)} · {countWords(sample)} words
+              </span>
+              <button
+                type="button"
+                className="shrink-0 text-small underline"
+                onClick={() => removeSample(index)}
+                data-testid={`remove-sample-${index}`}
+              >
+                Remove
+              </button>
+            </div>
+          ))}
+          {adding ? (
+            <div className="flex flex-col gap-1">
+              <label htmlFor="new-voice-sample" className="text-small font-medium">
+                New sample
+              </label>
+              <Textarea
+                id="new-voice-sample"
+                rows={6}
+                value={draft}
+                onChange={(e) => setDraft(e.target.value)}
+                data-testid="new-voice-sample"
+              />
+              <span
+                className="font-mono text-xs tabular-nums text-faint"
+                aria-live="polite"
+                data-testid="new-sample-counter"
+              >
+                {draftWords} words · at least {MIN_SAMPLE_WORDS}
+              </span>
+              <div className="flex gap-2">
+                <Button type="button" size="sm" onClick={addDraft} disabled={!draftValid} data-testid="confirm-add-sample">
+                  Add sample
+                </Button>
+                <Button
+                  type="button"
+                  variant="ghost"
+                  size="sm"
+                  onClick={() => {
+                    setDraft("");
+                    setAdding(false);
+                  }}
+                  data-testid="cancel-add-sample"
+                >
+                  Cancel
+                </Button>
+              </div>
+            </div>
+          ) : (
+            <Button
+              type="button"
+              variant="secondary"
+              size="sm"
+              className="self-start"
+              onClick={() => setAdding(true)}
+              disabled={samples.length >= 3}
+              data-testid="add-sample"
+            >
+              + Add sample
+            </Button>
+          )}
+        </div>
+
+        <div className="flex flex-wrap items-center gap-3">
+          <Button
+            type="button"
+            onClick={save}
+            disabled={saving || !profileComplete(profile)}
+            data-testid="save-voice"
+          >
+            {saving ? "Saving…" : "Save voice"}
+          </Button>
+          <span className="text-small text-faint">Used for every new episode.</span>
+          {saved ? (
+            <span role="status" className="text-small text-success" data-testid="voice-saved">
+              Saved
+            </span>
+          ) : null}
+          {error ? (
+            <span role="alert" className="text-small text-danger" data-testid="voice-error">
+              {error}
+            </span>
+          ) : null}
+        </div>
+      </div>
+
+      <VoicePreviewCard profile={profile} autoLoad />
+    </div>
+  );
+}
diff --git a/web/components/app/voice/chip-row.tsx b/web/components/app/voice/chip-row.tsx
new file mode 100644
index 0000000..65497e6
--- /dev/null
+++ b/web/components/app/voice/chip-row.tsx
@@ -0,0 +1,147 @@
+"use client";
+
+// MK-2/MK-8 chip + segmented primitives (mock.css .chip/.seg, §5.2 tokens).
+import { useState, type KeyboardEvent } from "react";
+import { PlusIcon, XIcon } from "lucide-react";
+import { Button } from "@/components/ui/button";
+import { Input } from "@/components/ui/input";
+
+/** Removable chip list with an inline "+ Add" flow (tone, Never say, …). */
+export function ChipList({
+  label,
+  items,
+  testIdPrefix,
+  onRemove,
+  onAdd,
+  max,
+  maxedCopy,
+}: {
+  label: string;
+  items: string[];
+  testIdPrefix: string;
+  onRemove: (item: string) => void;
+  onAdd: (item: string) => void;
+  max: number;
+  maxedCopy?: string;
+}) {
+  const [adding, setAdding] = useState(false);
+  const [draft, setDraft] = useState("");
+  const full = items.length >= max;
+
+  function commit() {
+    const value = draft.trim();
+    setDraft("");
+    setAdding(false);
+    if (!value) return;
+    if (items.some((i) => i.toLowerCase() === value.toLowerCase())) return;
+    onAdd(value);
+  }
+
+  function onKeyDown(event: KeyboardEvent<HTMLInputElement>) {
+    if (event.key === "Enter") {
+      event.preventDefault();
+      commit();
+    } else if (event.key === "Escape") {
+      setDraft("");
+      setAdding(false);
+    }
+  }
+
+  return (
+    <div className="flex flex-col gap-1.5">
+      <span className="text-small font-medium">{label}</span>
+      <div className="flex flex-wrap items-center gap-2" data-testid={`${testIdPrefix}-chips`}>
+        {items.map((item) => (
+          <span
+            key={item}
+            data-testid={`${testIdPrefix}-chip-${slug(item)}`}
+            className="inline-flex h-7 items-center gap-1.5 rounded-4xl border border-border-strong bg-surface px-2.5 text-small"
+          >
+            {item}
+            <button
+              type="button"
+              aria-label={`Remove ${item}`}
+              data-testid={`${testIdPrefix}-remove-${slug(item)}`}
+              onClick={() => onRemove(item)}
+              className="text-faint transition-colors hover:text-danger"
+            >
+              <XIcon className="size-3.5" aria-hidden />
+            </button>
+          </span>
+        ))}
+        {adding && !full ? (
+          <Input
+            autoFocus
+            value={draft}
+            onChange={(e) => setDraft(e.target.value)}
+            onBlur={commit}
+            onKeyDown={onKeyDown}
+            aria-label={`New ${label.toLowerCase()} term`}
+            className="h-7 w-36 py-0 text-small"
+            maxLength={60}
+          />
+        ) : (
+          <Button
+            type="button"
+            variant="ghost"
+            size="sm"
+            data-testid={`${testIdPrefix}-add`}
+            disabled={full}
+            title={full ? maxedCopy : undefined}
+            onClick={() => setAdding(true)}
+          >
+            <PlusIcon data-icon="inline-start" aria-hidden /> Add
+          </Button>
+        )}
+      </div>
+    </div>
+  );
+}
+
+/** §MK-2/MK-8 segmented control (Emoji, Write as, Sentence length). */
+export function Segmented<T extends string>({
+  label,
+  value,
+  options,
+  testIdPrefix,
+  onChange,
+}: {
+  label: string;
+  value: T;
+  options: Array<{ value: T; label: string }>;
+  testIdPrefix: string;
+  onChange: (value: T) => void;
+}) {
+  return (
+    <div className="flex items-center gap-3">
+      <span className="text-small font-medium">{label}</span>
+      <div
+        role="radiogroup"
+        aria-label={label}
+        className="inline-flex overflow-hidden rounded-md border border-border-strong"
+        data-testid={testIdPrefix}
+      >
+        {options.map((option) => (
+          <button
+            key={option.value}
+            type="button"
+            role="radio"
+            aria-checked={value === option.value}
+            data-testid={`${testIdPrefix}-${option.value}`}
+            onClick={() => onChange(option.value)}
+            className={`px-3 py-1.5 text-small transition-colors ${
+              value === option.value ? "bg-accent-soft font-medium text-accent" : "text-body hover:bg-surface-2"
+            }`}
+          >
+            {option.label}
+          </button>
+        ))}
+      </div>
+    </div>
+  );
+}
+
+/** Stable DOM/testid slug for a chip value. */
+export function slug(value: string): string {
+  return value.toLowerCase().replace(/[^a-z0-9]+/g, "-").replace(/^-+|-+$/g, "") || "term";
+}
diff --git a/web/components/app/voice/preview-card.tsx b/web/components/app/voice/preview-card.tsx
new file mode 100644
index 0000000..ece0413
--- /dev/null
+++ b/web/components/app/voice/preview-card.tsx
@@ -0,0 +1,111 @@
+"use client";
+
+// MK-2/MK-8 preview card: a short sample paragraph written in the saved (or
+// current) voice. [Update preview] posts the live editor state to
+// /api/show/voice/preview; a deterministic (no-AI) answer is labelled with
+// the §6 fallback copy instead of passing as model text.
+import { useCallback, useEffect, useRef, useState } from "react";
+import { Button } from "@/components/ui/button";
+import { VOICE_BASIC_PREVIEW_COPY } from "@/lib/copy";
+
+type PreviewRequestProfile = {
+  tone: string[];
+  formality: number;
+  energy: number;
+  humor: number;
+  person: string;
+  sentenceLength: string;
+  emoji: string;
+  neverSay: string[];
+};
+
+type PreviewResponse = {
+  status?: "ai" | "deterministic";
+  text?: string;
+  error?: string;
+};
+
+export function VoicePreviewCard({
+  profile,
+  autoLoad = false,
+}: {
+  profile: PreviewRequestProfile;
+  autoLoad?: boolean;
+}) {
+  const [text, setText] = useState<string | null>(null);
+  const [basic, setBasic] = useState(false);
+  const [loading, setLoading] = useState(false);
+  const [failed, setFailed] = useState(false);
+  const autoLoaded = useRef(false);
+
+  const update = useCallback(async () => {
+    setLoading(true);
+    setFailed(false);
+    try {
+      const res = await fetch("/api/show/voice/preview", {
+        method: "POST",
+        headers: { "content-type": "application/json" },
+        body: JSON.stringify({ profile }),
+      });
+      const data = (await res.json().catch(() => ({}))) as PreviewResponse;
+      if (!res.ok || !data.text) {
+        setFailed(true);
+        return;
+      }
+      setText(data.text);
+      setBasic(data.status === "deterministic");
+    } catch {
+      setFailed(true);
+    } finally {
+      setLoading(false);
+    }
+  }, [profile]);
+
+  useEffect(() => {
+    if (autoLoaded.current || !autoLoad) return;
+    autoLoaded.current = true;
+    void update();
+  }, [autoLoad, update]);
+
+  return (
+    <div className="rounded-lg border p-4" data-testid="voice-preview-card">
+      <span className="text-micro text-faint">PREVIEW</span>
+      {text !== null ? (
+        <p className="mt-2 text-small leading-[22px]" data-testid="preview-text">
+          {text}
+        </p>
+      ) : (
+        <div className="mt-2 flex flex-col gap-2" aria-hidden data-testid="preview-skeleton">
+          <div className="h-3.5 w-full rounded bg-surface-3" />
+          <div className="h-3.5 w-4/5 rounded bg-surface-3" />
+          <div className="h-3.5 w-3/5 rounded bg-surface-3" />
+        </div>
+      )}
+      <span aria-live="polite" className="sr-only" data-testid="preview-status">
+        {loading ? "Writing a preview." : text ? "Preview ready." : ""}
+      </span>
+      {basic ? (
+        <span className="mt-2 block text-micro text-faint" data-testid="preview-note">
+          {VOICE_BASIC_PREVIEW_COPY}
+        </span>
+      ) : null}
+      {failed ? (
+        <span role="alert" className="mt-2 block text-small text-danger" data-testid="preview-error">
+          Something went wrong. Try again.
+        </span>
+      ) : null}
+      <div className="mt-3">
+        <Button
+          type="button"
+          variant="secondary"
+          size="sm"
+          onClick={update}
+          disabled={loading}
+          data-testid="preview-update"
+        >
+          {loading ? "Writing…" : "Update preview"}
+        </Button>
+      </div>
+    </div>
+  );
+}
diff --git a/web/components/app/voice/profile-editor.tsx b/web/components/app/voice/profile-editor.tsx
new file mode 100644
index 0000000..c76b28a
--- /dev/null
+++ b/web/components/app/voice/profile-editor.tsx
@@ -0,0 +1,129 @@
+"use client";
+
+// §7.4 voice profile editor shared by onboarding step 2 (MK-2 after-learning)
+// and Settings → Voice (MK-8): tone chips, 1–5 sliders, segmented options,
+// Never say / Always include chips. Pure controlled component — persistence
+// belongs to the forms that render it.
+import type { VoiceProfile } from "@/lib/ai/voice";
+import { VOICE_TONE_HINT_COPY } from "@/lib/copy";
+import { Slider } from "@/components/ui/slider";
+import { ChipList, Segmented } from "./chip-row";
+
+const SLIDERS = [
+  { key: "formality", label: "Formality" },
+  { key: "energy", label: "Energy" },
+  { key: "humor", label: "Humor" },
+] as const;
+
+export function VoiceProfileEditor({
+  profile,
+  onChange,
+  showAlwaysInclude = false,
+}: {
+  profile: VoiceProfile;
+  onChange: (next: VoiceProfile) => void;
+  showAlwaysInclude?: boolean;
+}) {
+  const set = <K extends keyof VoiceProfile>(key: K, value: VoiceProfile[K]) =>
+    onChange({ ...profile, [key]: value });
+
+  return (
+    <div className="flex flex-col gap-6">
+      <ChipList
+        label="Tone"
+        testIdPrefix="tone"
+        items={profile.tone}
+        max={3}
+        onRemove={(item) => set("tone", profile.tone.filter((t) => t !== item))}
+        onAdd={(item) => set("tone", [...profile.tone, item])}
+      />
+
+      <div className="flex flex-col gap-4">
+        {SLIDERS.map(({ key, label }) => (
+          <div key={key} className="grid grid-cols-[90px_1fr_24px] items-center gap-3">
+            <span className="text-small font-medium">{label}</span>
+            <Slider
+              min={1}
+              max={5}
+              step={1}
+              value={[profile[key]]}
+              onValueChange={([v]) => set(key, v)}
+              ariaLabel={label}
+              data-testid={`slider-${key}`}
+            />
+            <span className="text-right font-mono text-xs tabular-nums" aria-hidden>
+              {profile[key]}
+            </span>
+          </div>
+        ))}
+      </div>
+
+      <div className="flex flex-wrap items-center gap-x-6 gap-y-3">
+        <Segmented
+          label="Emoji"
+          testIdPrefix="seg-emoji"
+          value={profile.emoji}
+          onChange={(v) => set("emoji", v)}
+          options={[
+            { value: "none", label: "None" },
+            { value: "light", label: "Light" },
+            { value: "heavy", label: "Heavy" },
+          ]}
+        />
+        <Segmented
+          label="Write as"
+          testIdPrefix="seg-person"
+          value={profile.person}
+          onChange={(v) => set("person", v)}
+          options={[
+            { value: "I", label: "I" },
+            { value: "we", label: "We" },
+            { value: "they", label: "They" },
+          ]}
+        />
+        <Segmented
+          label="Sentence length"
+          testIdPrefix="seg-sentencelength"
+          value={profile.sentenceLength}
+          onChange={(v) => set("sentenceLength", v)}
+          options={[
+            { value: "short", label: "Short" },
+            { value: "medium", label: "Medium" },
+            { value: "long", label: "Long" },
+          ]}
+        />
+      </div>
+
+      <ChipList
+        label="Never say"
+        testIdPrefix="never-say"
+        items={profile.neverSay}
+        max={30}
+        onRemove={(item) => set("neverSay", profile.neverSay.filter((t) => t !== item))}
+        onAdd={(item) => set("neverSay", [...profile.neverSay, item])}
+      />
+
+      {showAlwaysInclude ? (
+        <ChipList
+          label="Always include"
+          testIdPrefix="always-include"
+          items={profile.alwaysInclude}
+          max={5}
+          onRemove={(item) => set("alwaysInclude", profile.alwaysInclude.filter((t) => t !== item))}
+          onAdd={(item) => set("alwaysInclude", [...profile.alwaysInclude, item])}
+        />
+      ) : null}
+
+      {profile.tone.length < 3 ? (
+        <span role="status" className="text-small text-warning" data-testid="tone-hint">
+          {VOICE_TONE_HINT_COPY}
+        </span>
+      ) : null}
+    </div>
+  );
+}
+
+/** §7.4 fixes tone at exactly 3 — forms gate Save/Continue on this. */
+export function profileComplete(profile: VoiceProfile): boolean {
+  return profile.tone.length === 3;
+}
diff --git a/web/db/queries.ts b/web/db/queries.ts
index 3d893e9..5acebf1 100644
--- a/web/db/queries.ts
+++ b/web/db/queries.ts
@@ -142,6 +142,9 @@ export async function updateShow(
     listenUrl: string | null;
     defaultCta: string | null;
     vocabulary: string[];
+    /** §7.4 profile JSON — validated with VoiceProfileSchema before it gets here. */
+    voiceProfile: unknown;
+    voiceSamples: string[];
   }>,
 ): Promise<ShowRow> {
   const db = getDb();
diff --git a/web/lib/ai/voice.ts b/web/lib/ai/voice.ts
index 8a14daa..578052d 100644
--- a/web/lib/ai/voice.ts
+++ b/web/lib/ai/voice.ts
@@ -158,3 +158,89 @@ export async function previewVoice(args: {
   }
   return { ok: false, errorCode: "voice_preview_failed" };
 }
+
+// ---- deterministic no-AI mode (§8.1 last resort, added in WP-07) ------------
+// U3 gives "Learn my voice" ≤15 s; with no usable provider key the chain above
+// fails in seconds and onboarding must still produce a profile. Both functions
+// below derive ONLY from the user's own input (extractive, never invented);
+// fields a lexicon-free analyzer cannot honestly infer stay at the §7.4
+// defaults. Callers must label the result as the basic version — it is never
+// presented as a learned AI profile.
+
+/** First-person pronoun counts decide person (word-boundary, contractions included). */
+const PRONOUN_RE = {
+  I: /\bI\b|\bI'm\b|\bI've\b|\bI'll\b|\bI'd\b|\bme\b|\bmy\b|\bmine\b|\bmyself\b/gi,
+  we: /\bwe\b|\bwe're\b|\bwe've\b|\bwe'll\b|\bwe'd\b|\bus\b|\bour\b|\bours\b|\bourselves\b/gi,
+  they: /\bthey\b|\bthey're\b|\bthey've\b|\bthey'll\b|\bthem\b|\btheir\b|\btheirs\b/gi,
+} as const;
+
+/** Sentence splitter without the engine's 40-char floor: short-sentence
+ * authors would otherwise look "long" because their sentences get dropped. */
+function splitSentences(text: string): string[] {
+  return text
+    .split(/(?<=[.!?])\s+/)
+    .map((s) => s.trim())
+    .filter(Boolean);
+}
+
+export function learnVoiceProfileDeterministic(samples: string[]): VoiceProfile {
+  const text = samples.join(" ").replace(/\s+/g, " ").trim();
+  const count = (re: RegExp) => (text.match(re)?.length ?? 0);
+  const iCount = count(PRONOUN_RE.I);
+  const weCount = count(PRONOUN_RE.we);
+  const theyCount = count(PRONOUN_RE.they);
+  const max = Math.max(iCount, weCount, theyCount);
+  const person: VoiceProfile["person"] =
+    max === 0 ? DEFAULT_VOICE_PROFILE.person : iCount === max ? "I" : weCount === max ? "we" : "they";
+
+  const sents = splitSentences(text);
+  const words = text.split(" ").filter(Boolean);
+  const avgWords = sents.length ? words.length / sents.length : 0;
+  const sentenceLength: VoiceProfile["sentenceLength"] =
+    sents.length === 0 || (avgWords >= 12 && avgWords <= 22)
+      ? DEFAULT_VOICE_PROFILE.sentenceLength
+      : avgWords < 12
+        ? "short"
+        : "long";
+
+  const emojiCount = (text.match(/\p{Extended_Pictographic}/gu) ?? []).length;
+  const emoji: VoiceProfile["emoji"] =
+    emojiCount === 0 ? DEFAULT_VOICE_PROFILE.emoji : emojiCount <= 5 ? "light" : "heavy";
+
+  return { ...DEFAULT_VOICE_PROFILE, person, sentenceLength, emoji };
+}
+
+/**
+ * A ~60-word sample paragraph rendered from the profile's own parameters
+ * (person, tone, sentence length, emoji). Pure function of its inputs so the
+ * preview card always reflects the editor state; contains no never-say term
+ * because the template vocabulary is fixed.
+ */
+export function previewVoiceDeterministic(voice: VoiceProfile, show?: { name?: string }): string {
+  const name = show?.name?.trim() || "the show";
+  const subject = voice.person === "I" ? "I" : voice.person === "we" ? "We" : "They";
+  const poss = voice.person === "I" ? "my" : voice.person === "we" ? "our" : "their";
+  const casual = voice.formality <= 2;
+
+  const opener = `This week on ${name}: a ${voice.tone.join(", ")} look at the habits behind a calmer week.`;
+  const core =
+    voice.sentenceLength === "short"
+      ? `${subject} ${casual ? "get into" : "discuss"} ${poss} favorite routine.`
+      : `${subject} ${casual ? "get into" : "discuss"} why small routines beat big plans, and what ${poss} listeners can borrow from them.`;
+  const extra =
+    voice.sentenceLength === "short"
+      ? ""
+      : `And a quick answer to a listener question about keeping ${poss} workflow simple.`;
+  const closer =
+    voice.energy >= 4
+      ? "Bring some energy — this one moves."
+      : voice.humor >= 4
+        ? casual
+          ? "There's at least one bad joke ahead."
+          : "There is at least one bad joke ahead."
+        : casual
+          ? "Have a listen."
+          : "Enjoy the episode.";
+  const marks = voice.emoji === "heavy" ? " 🎧✨" : voice.emoji === "light" ? " 🎧" : "";
+  return `${[opener, core, extra, closer].filter(Boolean).join(" ")}${marks}`;
+}
diff --git a/web/lib/copy.ts b/web/lib/copy.ts
index c1e9c62..1c88e17 100644
--- a/web/lib/copy.ts
+++ b/web/lib/copy.ts
@@ -16,6 +16,22 @@ export const SHOW_NAME_REQUIRED_COPY = "Enter your show name.";
 
 export const URL_INVALID_COPY = "Enter a valid URL.";
 
+// WP-07 — voice onboarding + settings (§7.4, MK-2, MK-8). Where the §6 deck is
+// silent these are plain second person, ≤ 12 words (01-product.md §0.4 rule 4).
+export const SAMPLE_TOO_SHORT_COPY = "Each example needs at least 150 words.";
+
+export const SAMPLE_LIMIT_COPY = "Up to three examples.";
+
+export const VOICE_PROFILE_INVALID_COPY = "That voice profile isn't valid.";
+
+/** Deterministic no-AI results are labelled, never passed off as AI (U3/U10). */
+export const VOICE_BASIC_PROFILE_COPY = "Basic profile — our AI writer was unavailable.";
+
+export const VOICE_BASIC_PREVIEW_COPY = "Basic version — our AI writer was unavailable.";
+
+/** §7.4 fixes tone at exactly 3 words; Save/Continue waits for a replacement. */
+export const VOICE_TONE_HINT_COPY = "Add a tone word.";
+
 // --- WP-05: intake + ASR (§6 copy deck unless noted) ---
 
 /** U19 — verbatim. */
diff --git a/web/lib/show/input.ts b/web/lib/show/input.ts
index fd0f8ab..b54e875 100644
--- a/web/lib/show/input.ts
+++ b/web/lib/show/input.ts
@@ -1,5 +1,14 @@
 import { z } from "zod";
-import { SHOW_NAME_REQUIRED_COPY, URL_INVALID_COPY } from "@/lib/copy";
+import {
+  SAMPLE_LIMIT_COPY,
+  SAMPLE_TOO_SHORT_COPY,
+  SHOW_NAME_REQUIRED_COPY,
+  URL_INVALID_COPY,
+} from "@/lib/copy";
+import { VoiceProfileSchema } from "@/lib/ai/voice";
+import { MIN_SAMPLE_WORDS, countWords } from "@/lib/words";
+
+export { MIN_SAMPLE_WORDS, countWords };
 
 const emptyStringToUndefined = (v: unknown) =>
   typeof v === "string" && v.trim() === "" ? undefined : v;
@@ -37,8 +46,43 @@ export const showCreateSchema = z.object({
   vocabulary: z.array(requiredText(60).refine((v) => v.length > 0)).max(50).optional(),
 });
 
-/** PATCH accepts any subset of the General settings fields (M9). */
-export const showPatchSchema = showCreateSchema.partial();
+// ---- voice (M1 step 2, §7.5) -------------------------------------------------
+
+/** M1: past show notes of at least 150 words, up to 3, each ≤50k chars. */
+export const MAX_SAMPLE_CHARS = 50_000;
+
+export const voiceSampleSchema = z
+  .string()
+  .trim()
+  .min(1, SAMPLE_TOO_SHORT_COPY)
+  .max(MAX_SAMPLE_CHARS, SAMPLE_TOO_SHORT_COPY)
+  .refine((v) => countWords(v) >= MIN_SAMPLE_WORDS, SAMPLE_TOO_SHORT_COPY);
+
+function voiceSampleArray() {
+  return z
+    .array(voiceSampleSchema)
+    .min(1, SAMPLE_TOO_SHORT_COPY)
+    .max(3, SAMPLE_LIMIT_COPY);
+}
+
+/** POST /api/show/voice/learn — 1–3 past show notes (M1/U3). */
+export const voiceLearnSchema = z.object({
+  samples: voiceSampleArray(),
+});
+
+/** POST /api/show/voice/preview — profile to preview (stored one when absent). */
+export const voicePreviewSchema = z.object({
+  profile: z.unknown().optional(),
+  topic: z.string().trim().max(200).optional(),
+});
+
+/** PATCH accepts any subset of the General settings fields (M9), plus the
+ * §7.4 voice profile / samples the Voice settings page and onboarding step 2
+ * save (profile validated by the shared WP-04 schema). */
+export const showPatchSchema = showCreateSchema.partial().extend({
+  voiceProfile: VoiceProfileSchema.optional(),
+  voiceSamples: voiceSampleArray().optional(),
+});
 
 export const rssPreviewSchema = z.object({
   url: z.string().trim().min(1).max(2048),
diff --git a/web/lib/show/service.ts b/web/lib/show/service.ts
index 1f2f767..b25a4ee 100644
--- a/web/lib/show/service.ts
+++ b/web/lib/show/service.ts
@@ -77,6 +77,8 @@ export async function handleShowPatch(userId: string, raw: unknown): Promise<Api
   if (patch.listenUrl !== undefined) updates.listenUrl = patch.listenUrl ?? null;
   if (patch.defaultCta !== undefined) updates.defaultCta = patch.defaultCta ?? null;
   if (patch.vocabulary !== undefined) updates.vocabulary = patch.vocabulary;
+  if (patch.voiceProfile !== undefined) updates.voiceProfile = patch.voiceProfile;
+  if (patch.voiceSamples !== undefined) updates.voiceSamples = patch.voiceSamples;
   const show = await queries.updateShow(existing.id, updates);
   return { status: 200, json: { show: serializeShow(show) } };
 }
diff --git a/web/lib/show/voice.ts b/web/lib/show/voice.ts
new file mode 100644
index 0000000..4ba8992
--- /dev/null
+++ b/web/lib/show/voice.ts
@@ -0,0 +1,162 @@
+// §7.4/§7.5 — voice profile resolution and the POST /api/show/voice/learn and
+// /api/show/voice/preview handlers (WP-07). AI runs ONLY through lib/ai
+// (learnVoiceProfile/previewVoice); when every provider is unavailable the
+// deterministic mode at the lib/ai boundary takes over and the response is
+// labelled "deterministic" — a failed AI call never masquerades as a learn.
+import type { ApiResult } from "@/lib/rss/preview";
+import { NO_SHOW_COPY, VOICE_PROFILE_INVALID_COPY } from "@/lib/copy";
+import { isDbConfigured } from "@/lib/env";
+import * as queries from "@/db/queries";
+import {
+  learnVoiceProfile,
+  learnVoiceProfileDeterministic,
+  parseVoiceProfile,
+  previewVoice,
+  previewVoiceDeterministic,
+  DEFAULT_VOICE_PROFILE,
+  type VoiceProfile,
+} from "@/lib/ai/voice";
+import { effectiveAiSelection } from "@/lib/ai/settings";
+import { resolveAppSettingsStore } from "@/lib/ai/settings-server";
+import type { LlmProvider } from "@/lib/ai/types";
+import { firstErrorMessage, voiceLearnSchema, voicePreviewSchema } from "./input";
+
+export interface ShowVoiceDeps {
+  /** provider chain override (tests inject fakes; the route uses the registry) */
+  chain?: LlmProvider[];
+}
+
+/**
+ * U3 gives "Learn my voice" ≤ 15 s. §8.1's 60 s timeout targets background
+ * generation; an interactive onboarding step caps the provider timeout lower
+ * so the deterministic answer still arrives inside the budget (DECISIONS.md).
+ */
+const VOICE_TIMEOUT_MS = 12_000;
+
+/**
+ * The show's effective voice profile: a stored §7.4 JSON when present and
+ * valid, otherwise the spec default. Skip never writes a profile — null IS
+ * the default — so this is the single read-time resolution point (U3).
+ */
+export function resolveShowVoiceProfile(
+  show: { voiceProfile: unknown } | null | undefined,
+): VoiceProfile {
+  return (show ? parseVoiceProfile(show.voiceProfile) : null) ?? DEFAULT_VOICE_PROFILE;
+}
+
+/** Registry selection (app_settings.ai → env → defaults) for the handlers. */
+async function resolveSelection() {
+  return (await effectiveAiSelection(resolveAppSettingsStore())).selection;
+}
+
+/**
+ * POST /api/show/voice/learn (U3): validate 1–3 samples (≥150 words each),
+ * learn through the §8.1 provider chain, and persist profile + samples. When
+ * every provider fails the deterministic profile is derived from the samples
+ * and labelled; keyless previews return saved:false instead of faking a save.
+ */
+export async function handleVoiceLearn(
+  userId: string,
+  raw: unknown,
+  deps: ShowVoiceDeps = {},
+): Promise<ApiResult> {
+  const parsed = voiceLearnSchema.safeParse(raw);
+  if (!parsed.success) {
+    return { status: 400, json: { error: firstErrorMessage(parsed.error) } };
+  }
+  const samples = parsed.data.samples;
+
+  const db = isDbConfigured();
+  const show = db ? await queries.getShowByOwner(userId) : null;
+  if (db && !show) {
+    return { status: 404, json: { error: NO_SHOW_COPY } };
+  }
+
+  const result = await learnVoiceProfile({
+    samples,
+    deps: deps.chain?.length
+      ? { chain: deps.chain }
+      : { selection: await resolveSelection(), timeoutMs: VOICE_TIMEOUT_MS },
+  });
+
+  let profile: VoiceProfile;
+  let learnStatus: "learned" | "deterministic";
+  let provider: string | null;
+  let model: string | null;
+  if (result.ok) {
+    profile = result.profile;
+    learnStatus = "learned";
+    provider = result.provider;
+    model = result.model;
+  } else {
+    profile = learnVoiceProfileDeterministic(samples);
+    learnStatus = "deterministic";
+    provider = null;
+    model = null;
+  }
+
+  let saved = false;
+  if (show) {
+    await queries.updateShow(show.id, { voiceProfile: profile, voiceSamples: samples });
+    saved = true;
+  }
+
+  return { status: 200, json: { status: learnStatus, profile, provider, model, saved } };
+}
+
+/**
+ * POST /api/show/voice/preview (M9/MK-8): write one short sample paragraph in
+ * the voice. The profile may be posted (live editor state) or falls back to
+ * the stored one. Deterministic text is labelled so the UI can say the AI
+ * writer was unavailable instead of faking a model answer.
+ */
+export async function handleVoicePreview(
+  userId: string,
+  raw: unknown,
+  deps: ShowVoiceDeps = {},
+): Promise<ApiResult> {
+  const parsed = voicePreviewSchema.safeParse(raw);
+  if (!parsed.success) {
+    return { status: 400, json: { error: firstErrorMessage(parsed.error) } };
+  }
+
+  const db = isDbConfigured();
+  const show = db ? await queries.getShowByOwner(userId) : null;
+
+  let voice: VoiceProfile;
+  if (parsed.data.profile !== undefined) {
+    const fromRequest = parseVoiceProfile(parsed.data.profile);
+    if (!fromRequest) {
+      return { status: 400, json: { error: VOICE_PROFILE_INVALID_COPY } };
+    }
+    voice = fromRequest;
+  } else {
+    voice = resolveShowVoiceProfile(show);
+  }
+
+  const showContext = { name: show?.name ?? "the show" };
+  const result = await previewVoice({
+    voice,
+    show: showContext,
+    topic: parsed.data.topic,
+    deps: deps.chain?.length
+      ? { chain: deps.chain }
+      : { selection: await resolveSelection(), timeoutMs: VOICE_TIMEOUT_MS },
+  });
+
+  if (result.ok) {
+    return {
+      status: 200,
+      json: { status: "ai", text: result.text, provider: result.provider, model: result.model },
+    };
+  }
+  return {
+    status: 200,
+    json: {
+      status: "deterministic",
+      text: previewVoiceDeterministic(voice, showContext),
+      provider: null,
+      model: null,
+    },
+  };
+}
diff --git a/web/lib/words.ts b/web/lib/words.ts
new file mode 100644
index 0000000..02240e0
--- /dev/null
+++ b/web/lib/words.ts
@@ -0,0 +1,10 @@
+// Client-safe word helpers (no lib/ai imports — countWords is used in browser
+// bundles by the voice forms; the zod schemas in lib/show/input reuse it).
+
+/** Whitespace-split word count, matching the onboarding counter. */
+export function countWords(text: string): number {
+  return text.trim().split(/\s+/).filter(Boolean).length;
+}
+
+/** M1: past show notes have at least 150 words (1–3 samples, §7.5). */
+export const MIN_SAMPLE_WORDS = 150;
diff --git a/web/tests/e2e/onboarding.spec.ts b/web/tests/e2e/onboarding.spec.ts
index eeb32a6..29ebe6e 100644
--- a/web/tests/e2e/onboarding.spec.ts
+++ b/web/tests/e2e/onboarding.spec.ts
@@ -54,9 +54,11 @@ test.describe("WP-02 keyless shells", () => {
     });
   });
 
-  test("step 2 placeholder offers skip to step 3", async ({ page }) => {
+  test("step 2 renders the voice form; skip lands on step 3 (WP-07 shell)", async ({ page }) => {
     await page.goto("/app/onboarding/voice");
-    await expect(page.getByText("This page is coming soon.")).toBeVisible();
+    await expect(page.getByRole("heading", { name: "Teach EpisodeKit your voice" })).toBeVisible();
+    await expect(page.getByTestId("sample-counter-0")).toHaveText("0 words · at least 150");
+    await expect(page.getByTestId("learn-voice")).toBeDisabled();
     await page.getByTestId("skip-voice").click();
     await expect(page.getByRole("heading", { name: "Add your first episode" })).toBeVisible();
   });
diff --git a/web/tests/e2e/settings.spec.ts b/web/tests/e2e/settings.spec.ts
index 265940d..c2f74a3 100644
--- a/web/tests/e2e/settings.spec.ts
+++ b/web/tests/e2e/settings.spec.ts
@@ -35,9 +35,7 @@ test.describe("WP-02 settings and app shells", () => {
     );
   });
 
-  test("settings tabs exist; Voice/Account are stubs and Billing is real (WP-08)", async ({ page }) => {
-    await page.goto("/app/settings/voice");
-    await expect(page.getByTestId("voice-stub")).toBeVisible();
+  test("settings tabs exist; Voice and Billing are real; Account is a stub (WP-07/WP-08)", async ({ page }) => {
     await page.goto("/app/settings/billing");
     await expect(page.getByTestId("billing-plan-card")).toBeVisible();
     await page.goto("/app/settings/account");
diff --git a/web/tests/e2e/voice.spec.ts b/web/tests/e2e/voice.spec.ts
new file mode 100644
index 0000000..80b025a
--- /dev/null
+++ b/web/tests/e2e/voice.spec.ts
@@ -0,0 +1,138 @@
+import { readFileSync } from "node:fs";
+import { join } from "node:path";
+import { expect, test } from "@playwright/test";
+
+/**
+ * WP-07 e2e — voice onboarding step 2 (U3, MK-2) and Settings → Voice
+ * (M9, MK-8), keyless: no AI keys and no DATABASE_URL, so "Learn my voice"
+ * runs the deterministic lib/ai path and answers are labelled; saving reports
+ * the preview copy instead of faking a write.
+ */
+// Playwright runs from web/ (playwright.config.ts location).
+const voiceSample = readFileSync(join(process.cwd(), "tests/fixtures/voice-sample.txt"), "utf8");
+
+test.describe("WP-07 onboarding step 2 (U3)", () => {
+  test("learn flow: paste sample → Learn → chips/sliders → Continue → step 3", async ({ page }) => {
+    await page.goto("/app/onboarding/voice");
+    await page.getByLabel("Example 1").fill(voiceSample);
+    await expect(page.getByTestId("sample-counter-0")).toContainText("170 words · at least 150");
+    await expect(page.getByTestId("learn-voice")).toBeEnabled();
+
+    await page.getByTestId("learn-voice").click();
+
+    // U3: chips + sliders appear within 15 s (deterministic path is instant here).
+    await expect(page.getByRole("heading", { name: "Here's your voice" })).toBeVisible({
+      timeout: 15_000,
+    });
+    await expect(page.getByTestId("tone-chips")).toBeVisible();
+    // Deterministic learn keeps the §7.4 tone trio; person is derived from the fixture ("I").
+    await expect(page.getByTestId("tone-chip-clear")).toBeVisible();
+    await expect(page.getByTestId("tone-chip-warm")).toBeVisible();
+    await expect(page.getByTestId("tone-chip-curious")).toBeVisible();
+    await expect(page.getByTestId("slider-formality")).toBeVisible();
+    await expect(page.getByTestId("slider-energy")).toBeVisible();
+    await expect(page.getByTestId("slider-humor")).toBeVisible();
+    await expect(page.getByTestId("seg-person-I")).toHaveAttribute("aria-checked", "true");
+    // Honest labelling: no AI keys exist here, so this was NOT a learned AI profile.
+    await expect(page.getByTestId("voice-note")).toContainText("Basic profile");
+    await expect(page.getByTestId("voice-saved-note")).toContainText(
+      "Saving is unavailable in this preview.",
+    );
+    // MK-2 shows a preview in the after-learning card, rendered from the profile.
+    await expect(page.getByTestId("preview-text")).toBeVisible({ timeout: 15_000 });
+
+    await page.getByTestId("continue-voice").click();
+    await expect(page.getByRole("heading", { name: "Add your first episode" })).toBeVisible();
+  });
+
+  test("an under-150-word sample keeps Learn disabled with the hint", async ({ page }) => {
+    await page.goto("/app/onboarding/voice");
+    await page.getByLabel("Example 1").fill("Too short by far.");
+    await expect(page.getByTestId("sample-counter-0")).toContainText("4 words · at least 150");
+    await expect(page.getByTestId("learn-voice")).toBeDisabled();
+    await expect(page.getByTestId("voice-min-hint")).toHaveText(
+      "Each example needs at least 150 words.",
+    );
+  });
+
+  test("a second example can be added up to three", async ({ page }) => {
+    await page.goto("/app/onboarding/voice");
+    await expect(page.getByLabel("Example 1")).toBeVisible();
+    await page.getByTestId("add-sample-0").click();
+    await expect(page.getByLabel("Example 2")).toBeVisible();
+    await page.getByTestId("add-sample-1").click();
+    await expect(page.getByLabel("Example 3")).toBeVisible();
+    await expect(page.getByTestId("add-sample-2")).toHaveCount(0);
+  });
+
+  test("tone chips are removable and the editor gates Continue on three tones", async ({ page }) => {
+    await page.goto("/app/onboarding/voice");
+    await page.getByLabel("Example 1").fill(voiceSample);
+    await page.getByTestId("learn-voice").click();
+    await expect(page.getByTestId("tone-chips")).toBeVisible({ timeout: 15_000 });
+
+    await page.getByTestId("tone-remove-clear").click();
+    await expect(page.getByTestId("tone-hint")).toHaveText("Add a tone word.");
+    await expect(page.getByTestId("continue-voice")).toBeDisabled();
+    await page.getByTestId("tone-add").click();
+    await page.getByLabel("New tone term").fill("wry");
+    await page.getByLabel("New tone term").press("Enter");
+    await expect(page.getByTestId("tone-chip-wry")).toBeVisible();
+    await expect(page.getByTestId("tone-hint")).toHaveCount(0);
+    await expect(page.getByTestId("continue-voice")).toBeEnabled();
+  });
+
+  test("skip for now lands on step 3 without learning", async ({ page }) => {
+    await page.goto("/app/onboarding/voice");
+    await page.getByTestId("skip-voice").click();
+    await expect(page.getByRole("heading", { name: "Add your first episode" })).toBeVisible();
+  });
+});
+
+test.describe("WP-07 settings → Voice (M9, MK-8)", () => {
+  test("default §7.4 profile renders keyless with preview and samples block", async ({ page }) => {
+    await page.goto("/app/settings/voice");
+    await expect(page.getByTestId("tone-chip-clear")).toBeVisible();
+    await expect(page.getByTestId("tone-chip-warm")).toBeVisible();
+    await expect(page.getByTestId("tone-chip-curious")).toBeVisible();
+    await expect(page.getByTestId("slider-humor")).toBeVisible();
+    await expect(page.getByTestId("seg-person-we")).toHaveAttribute("aria-checked", "true");
+    await expect(page.getByTestId("seg-emoji-none")).toHaveAttribute("aria-checked", "true");
+    await expect(page.getByTestId("seg-sentencelength-medium")).toHaveAttribute(
+      "aria-checked",
+      "true",
+    );
+    await expect(page.getByTestId("never-say-chip-delve")).toBeVisible();
+    await expect(page.getByTestId("always-include-chips")).toBeVisible();
+    await expect(page.getByTestId("voice-samples")).toBeVisible();
+
+    // Preview renders from the profile (deterministic keyless) and is labelled.
+    await expect(page.getByTestId("preview-text")).toBeVisible({ timeout: 15_000 });
+    await expect(page.getByTestId("preview-note")).toContainText("Basic version");
+  });
+
+  test("Update preview re-renders; Save keyless shows the preview copy", async ({ page }) => {
+    await page.goto("/app/settings/voice");
+    await expect(page.getByTestId("preview-text")).toBeVisible({ timeout: 15_000 });
+    await page.getByTestId("preview-update").click();
+    await expect(page.getByTestId("preview-text")).toBeVisible();
+
+    await page.getByTestId("save-voice").click();
+    await expect(page.getByTestId("voice-error")).toHaveText(
+      "Saving is unavailable in this preview.",
+    );
+  });
+
+  test("adding a sample requires 150 words; remove works", async ({ page }) => {
+    await page.goto("/app/settings/voice");
+    await page.getByTestId("add-sample").click();
+    await page.getByTestId("new-voice-sample").fill("Not nearly long enough yet.");
+    await expect(page.getByTestId("confirm-add-sample")).toBeDisabled();
+    await page.getByTestId("new-voice-sample").fill(voiceSample);
+    await expect(page.getByTestId("confirm-add-sample")).toBeEnabled();
+    await page.getByTestId("confirm-add-sample").click();
+    await expect(page.getByTestId("voice-sample-row-0")).toContainText("170 words");
+    await page.getByTestId("remove-sample-0").click();
+    await expect(page.getByTestId("voice-sample-row-0")).toHaveCount(0);
+  });
+});
diff --git a/web/tests/fixtures/voice-sample.txt b/web/tests/fixtures/voice-sample.txt
new file mode 100644
index 0000000..ef22d90
--- /dev/null
+++ b/web/tests/fixtures/voice-sample.txt
@@ -0,0 +1,5 @@
+This week I look at the small habits that keep a recording week calm. I talk about why I plan episodes in batches, and how I turn one rough idea into a clear outline before I ever press record. I also share the checklist I run the night before a session, from checking levels to closing every tab that wants my attention.
+
+Later in the episode I walk through the way I review an outline, cut the sections that do not help the listener, and leave room for the parts that do. I keep the gear list short on purpose. The tools matter less than the routine, and the routine fits on one page. If your week feels loud, this episode is a quiet place to start.
+
+I close with one question to ask yourself before your next recording, and a simple way to keep the answer visible all week. It is the same question I ask myself every Sunday evening, and it takes less than a minute to answer.
diff --git a/web/tests/unit/show-voice.test.ts b/web/tests/unit/show-voice.test.ts
new file mode 100644
index 0000000..8fff1da
--- /dev/null
+++ b/web/tests/unit/show-voice.test.ts
@@ -0,0 +1,348 @@
+// WP-07 — voice onboarding + settings: §7.4 default exactness, deterministic
+// no-key learn path, sample word-count validation, and route boundary
+// behaviour (DB and providers mocked at the boundary, never faked inline).
+import { afterEach, describe, expect, it, vi } from "vitest";
+import { readFileSync } from "node:fs";
+import {
+  DEFAULT_VOICE_PROFILE,
+  VoiceProfileSchema,
+  learnVoiceProfileDeterministic,
+  previewVoiceDeterministic,
+  type VoiceProfile,
+} from "@/lib/ai/voice";
+import { handleVoiceLearn, handleVoicePreview, resolveShowVoiceProfile } from "@/lib/show/voice";
+import { showPatchSchema, voiceLearnSchema } from "@/lib/show/input";
+import { SAMPLE_TOO_SHORT_COPY, VOICE_PROFILE_INVALID_COPY } from "@/lib/copy";
+import { fakeProvider, failingProvider } from "./helpers/ai";
+
+vi.mock("@/db/queries", () => ({
+  getShowByOwner: vi.fn(),
+  updateShow: vi.fn(),
+}));
+
+import * as queries from "@/db/queries";
+
+const mocked = vi.mocked(queries);
+
+const voiceSample = readFileSync(new URL("../fixtures/voice-sample.txt", import.meta.url), "utf8");
+
+const showRow = {
+  id: "show-1",
+  workspaceId: "ws-1",
+  name: "The Focus Hour",
+  rssUrl: null,
+  artworkUrl: null,
+  listenUrl: null,
+  defaultCta: null,
+  vocabulary: [],
+  voiceProfile: null,
+  voiceSamples: [],
+  createdAt: new Date(),
+};
+
+/** A learned profile that differs from the default in every derived field. */
+const LEARNED: VoiceProfile = {
+  tone: ["dry", "direct", "curious"],
+  formality: 2,
+  energy: 4,
+  humor: 3,
+  person: "I",
+  sentenceLength: "short",
+  emoji: "light",
+  signaturePhrases: ["here's the thing"],
+  neverSay: ["synergy"],
+  alwaysInclude: ["follow the show"],
+  ctaStyle: "Ask for one reply.",
+};
+
+afterEach(() => {
+  delete process.env.DATABASE_URL;
+  vi.clearAllMocks();
+});
+
+describe("§7.4 default profile — byte-for-byte (Skip path)", () => {
+  it("the default constant equals the spec literal field for field", () => {
+    expect(DEFAULT_VOICE_PROFILE).toEqual({
+      tone: ["clear", "warm", "curious"],
+      formality: 3,
+      energy: 3,
+      humor: 2,
+      person: "we",
+      sentenceLength: "medium",
+      emoji: "none",
+      signaturePhrases: [],
+      neverSay: ["delve", "game-changer", "unlock", "dive deep", "in today's fast-paced world"],
+      alwaysInclude: [],
+      ctaStyle: "Invite listeners to follow the show.",
+    });
+  });
+
+  it("a null show row and a null stored profile both resolve to the exact default", () => {
+    expect(resolveShowVoiceProfile(null)).toEqual(DEFAULT_VOICE_PROFILE);
+    expect(resolveShowVoiceProfile(showRow)).toEqual(DEFAULT_VOICE_PROFILE);
+    expect(resolveShowVoiceProfile({ voiceProfile: null })).toEqual(DEFAULT_VOICE_PROFILE);
+  });
+
+  it("a corrupt stored profile falls back to the exact default (never crashes)", () => {
+    expect(resolveShowVoiceProfile({ voiceProfile: { tone: ["nope"], formality: 99 } })).toEqual(
+      DEFAULT_VOICE_PROFILE,
+    );
+  });
+
+  it("a stored valid profile is returned as-is", () => {
+    expect(resolveShowVoiceProfile({ voiceProfile: LEARNED })).toEqual(LEARNED);
+  });
+
+  it("Skip persists nothing (no updateShow call happens on the skip path)", async () => {
+    // Skip is a client-side navigation; the default is resolved at read time.
+    // This pins the contract the UI relies on: resolve(null) === default.
+    expect(mocked.updateShow).not.toHaveBeenCalled();
+    expect(resolveShowVoiceProfile(null).tone).toEqual(["clear", "warm", "curious"]);
+  });
+});
+
+describe("deterministic learn (no-key path)", () => {
+  it("keeps every underivable field at the §7.4 default and stays schema-valid", () => {
+    const profile = learnVoiceProfileDeterministic([voiceSample]);
+    expect(VoiceProfileSchema.safeParse(profile).success).toBe(true);
+    expect(profile.tone).toEqual(DEFAULT_VOICE_PROFILE.tone);
+    expect(profile.formality).toBe(DEFAULT_VOICE_PROFILE.formality);
+    expect(profile.energy).toBe(DEFAULT_VOICE_PROFILE.energy);
+    expect(profile.humor).toBe(DEFAULT_VOICE_PROFILE.humor);
+    expect(profile.signaturePhrases).toEqual(DEFAULT_VOICE_PROFILE.signaturePhrases);
+    expect(profile.neverSay).toEqual(DEFAULT_VOICE_PROFILE.neverSay);
+    expect(profile.alwaysInclude).toEqual(DEFAULT_VOICE_PROFILE.alwaysInclude);
+    expect(profile.ctaStyle).toBe(DEFAULT_VOICE_PROFILE.ctaStyle);
+  });
+
+  it("derives person from first-person pronouns in the samples", () => {
+    const first = learnVoiceProfileDeterministic([
+      "I plan my week on Sunday. My notes are short. I record alone and I edit the same day.",
+    ]);
+    expect(first.person).toBe("I");
+
+    const together = learnVoiceProfileDeterministic([
+      "We plan our week together. Our notes keep us honest. We record on Mondays and we ship the same day.",
+    ]);
+    expect(together.person).toBe("we");
+
+    const third = learnVoiceProfileDeterministic([
+      "They plan their week early. Their notes stay short. They record on Mondays and they ship the same day.",
+    ]);
+    expect(third.person).toBe("they");
+  });
+
+  it("derives sentence length from the average sentence", () => {
+    const short = learnVoiceProfileDeterministic([
+      "Short sentences. Very short. They land. Every time. No filler. Cut it. Then cut again. Ship it now.",
+    ]);
+    expect(short.sentenceLength).toBe("short");
+
+    const long = learnVoiceProfileDeterministic([
+      "When you sit down on a Sunday evening and look at the week ahead, the plan you wrote on Friday morning can feel like it was left there by a stranger who had far more energy and far fewer meetings than the person reading it now, and that gap is exactly where most plans quietly die.",
+    ]);
+    expect(long.sentenceLength).toBe("long");
+  });
+
+  it("derives emoji usage from the sample text", () => {
+    const plain = learnVoiceProfileDeterministic([voiceSample]);
+    expect(plain.emoji).toBe("none");
+    const light = learnVoiceProfileDeterministic([
+      `${voiceSample} Great episode today 🎧 big thanks for listening ✨`.repeat(1),
+    ]);
+    expect(light.emoji).toBe("light");
+  });
+
+  it("is deterministic: the same samples give the same profile", () => {
+    expect(learnVoiceProfileDeterministic([voiceSample])).toEqual(
+      learnVoiceProfileDeterministic([voiceSample]),
+    );
+  });
+});
+
+describe("sample validation (§M1: 1–3 past show notes, ≥150 words each)", () => {
+  const words = (n: number) => Array.from({ length: n }, (_, i) => `word${i}`).join(" ");
+
+  it("accepts three samples of 150+ words", () => {
+    const parsed = voiceLearnSchema.safeParse({ samples: [words(150), words(160), words(200)] });
+    expect(parsed.success).toBe(true);
+  });
+
+  it("rejects a sample under 150 words with the friendly copy", () => {
+    const parsed = voiceLearnSchema.safeParse({ samples: [words(149)] });
+    expect(parsed.success).toBe(false);
+    if (!parsed.success) expect(parsed.error.issues[0]?.message).toBe(SAMPLE_TOO_SHORT_COPY);
+  });
+
+  it("rejects more than three samples and an empty body", () => {
+    expect(voiceLearnSchema.safeParse({ samples: [words(150), words(150), words(150), words(150)] }).success).toBe(false);
+    expect(voiceLearnSchema.safeParse({ samples: [] }).success).toBe(false);
+    expect(voiceLearnSchema.safeParse({}).success).toBe(false);
+  });
+});
+
+describe("handleVoiceLearn (POST /api/show/voice/learn)", () => {
+  it("400 with friendly copy when a sample is too short", async () => {
+    const result = await handleVoiceLearn("u1", { samples: ["too short"] });
+    expect(result.status).toBe(400);
+    expect(result.json).toEqual({ error: SAMPLE_TOO_SHORT_COPY });
+    expect(mocked.getShowByOwner).not.toHaveBeenCalled();
+  });
+
+  it("provider success persists the learned profile + samples (saved: true)", async () => {
+    process.env.DATABASE_URL = "postgres://test";
+    mocked.getShowByOwner.mockResolvedValue(showRow);
+    mocked.updateShow.mockResolvedValue(showRow);
+    const provider = fakeProvider("groq", [JSON.stringify(LEARNED)]);
+
+    const result = await handleVoiceLearn(
+      "u1",
+      { samples: [voiceSample] },
+      { chain: [provider] },
+    );
+
+    expect(result.status).toBe(200);
+    expect(result.json).toMatchObject({ status: "learned", profile: LEARNED, saved: true });
+    expect(mocked.updateShow).toHaveBeenCalledWith("show-1", {
+      voiceProfile: LEARNED,
+      voiceSamples: [voiceSample.trim()],
+    });
+  });
+
+  it("every provider failing yields the deterministic profile, honestly labelled (U3, no fake learn)", async () => {
+    process.env.DATABASE_URL = "postgres://test";
+    mocked.getShowByOwner.mockResolvedValue(showRow);
+    mocked.updateShow.mockResolvedValue(showRow);
+
+    const result = await handleVoiceLearn(
+      "u1",
+      { samples: [voiceSample] },
+      { chain: [failingProvider("glm-coding"), failingProvider("groq")] },
+    );
+
+    expect(result.status).toBe(200);
+    expect(result.json).toMatchObject({
+      status: "deterministic",
+      provider: null,
+      model: null,
+      saved: true,
+    });
+    const profile = (result.json as { profile: VoiceProfile }).profile;
+    expect(VoiceProfileSchema.safeParse(profile).success).toBe(true);
+    expect(profile.person).toBe("I"); // derived from the fixture, not defaulted silently
+    expect(mocked.updateShow).toHaveBeenCalled();
+  });
+
+  it("404 when the DB is configured but no show exists yet", async () => {
+    process.env.DATABASE_URL = "postgres://test";
+    mocked.getShowByOwner.mockResolvedValue(null);
+    const result = await handleVoiceLearn(
+      "u1",
+      { samples: [voiceSample] },
+      { chain: [failingProvider("groq")] },
+    );
+    expect(result.status).toBe(404);
+    expect(mocked.updateShow).not.toHaveBeenCalled();
+  });
+
+  it("keyless preview: learn still runs (deterministic) but reports saved: false and writes nothing", async () => {
+    const result = await handleVoiceLearn(
+      "u1",
+      { samples: [voiceSample] },
+      { chain: [failingProvider("glm-coding")] },
+    );
+    expect(result.status).toBe(200);
+    expect(result.json).toMatchObject({ status: "deterministic", saved: false });
+    expect(mocked.getShowByOwner).not.toHaveBeenCalled();
+    expect(mocked.updateShow).not.toHaveBeenCalled();
+  });
+});
+
+describe("handleVoicePreview (POST /api/show/voice/preview)", () => {
+  it("returns the AI text when a provider answers", async () => {
+    const provider = fakeProvider("groq", ["Here is your sample paragraph in your voice."]);
+    const result = await handleVoicePreview("u1", { profile: LEARNED }, { chain: [provider] });
+    expect(result.status).toBe(200);
+    expect(result.json).toMatchObject({ status: "ai", provider: "groq" });
+    expect(result.json.text).toContain("sample paragraph");
+  });
+
+  it("falls back to the deterministic preview text (labelled) when providers fail", async () => {
+    const result = await handleVoicePreview(
+      "u1",
+      { profile: LEARNED },
+      { chain: [failingProvider("glm-coding")] },
+    );
+    expect(result.status).toBe(200);
+    expect(result.json).toMatchObject({ status: "deterministic", provider: null });
+    expect(typeof result.json.text).toBe("string");
+    expect(result.json.text).not.toBe("");
+  });
+
+  it("uses the stored profile when none is sent", async () => {
+    process.env.DATABASE_URL = "postgres://test";
+    mocked.getShowByOwner.mockResolvedValue({ ...showRow, voiceProfile: LEARNED });
+    const provider = fakeProvider("groq", ["A fresh preview."]);
+    const result = await handleVoicePreview("u1", {}, { chain: [provider] });
+    expect(result.json).toMatchObject({ status: "ai", text: "A fresh preview." });
+  });
+
+  it("rejects an invalid profile with friendly copy", async () => {
+    const result = await handleVoicePreview("u1", { profile: { formality: 42 } });
+    expect(result.status).toBe(400);
+    expect(result.json).toEqual({ error: VOICE_PROFILE_INVALID_COPY });
+  });
+});
+
+describe("deterministic preview (renders from the profile)", () => {
+  it("is a plain-text sample whose content tracks the profile", () => {
+    const a = previewVoiceDeterministic(LEARNED, { name: "The Focus Hour" });
+    expect(a).not.toMatch(/[{}]/); // plain text, not JSON
+    expect(a.split(/\s+/).filter(Boolean).length).toBeGreaterThan(25);
+    expect(a.split(/\s+/).filter(Boolean).length).toBeLessThan(90);
+    expect(a).toContain("The Focus Hour");
+    expect(a).not.toMatch(/\s{2,}/); // no double spaces from template joins
+    // person pronoun follows the profile
+    expect(a).toContain("I ");
+  });
+
+  it("changes when the profile changes", () => {
+    const a = previewVoiceDeterministic(LEARNED, { name: "The Focus Hour" });
+    const b = previewVoiceDeterministic(
+      { ...LEARNED, person: "we", emoji: "heavy", sentenceLength: "long" },
+      { name: "The Focus Hour" },
+    );
+    expect(a).not.toBe(b);
+  });
+
+  it("never contains a never-say word", () => {
+    const banned = LEARNED.neverSay[0];
+    const text = previewVoiceDeterministic({ ...LEARNED, neverSay: [banned] }, { name: "X" });
+    expect(text.toLowerCase()).not.toContain(banned.toLowerCase());
+  });
+
+  it("is stable for the same inputs", () => {
+    expect(previewVoiceDeterministic(LEARNED, { name: "X" })).toBe(
+      previewVoiceDeterministic(LEARNED, { name: "X" }),
+    );
+  });
+});
+
+describe("PATCH /api/show voice fields (settings Save voice, step-2 Continue)", () => {
+  it("accepts a valid voiceProfile and voiceSamples", () => {
+    const parsed = showPatchSchema.safeParse({
+      voiceProfile: LEARNED,
+      voiceSamples: [voiceSample],
+    });
+    expect(parsed.success).toBe(true);
+    if (parsed.success) {
+      expect(parsed.data.voiceProfile).toEqual(LEARNED);
+      expect(parsed.data.voiceSamples).toEqual([voiceSample.trim()]);
+    }
+  });
+
+  it("rejects an invalid profile and an under-length sample", () => {
+    expect(showPatchSchema.safeParse({ voiceProfile: { formality: 9 } }).success).toBe(false);
+    expect(showPatchSchema.safeParse({ voiceSamples: ["too short"] }).success).toBe(false);
+  });
+});