diff --git a/tests/CLAUDE.md b/tests/CLAUDE.md
index 36e155ef3..2bfcc9146 100644
--- a/tests/CLAUDE.md
+++ b/tests/CLAUDE.md
@@ -217,6 +217,15 @@ vendored page and use its HTTP API. A render-gate test should call
 helper directly only when the helper itself carries a contract that would otherwise
 be hard to diagnose, such as the traffic wait reaching its deadline.
 
+A reading the layer makes from declarations belongs on a widget that declares them,
+not on whichever shipped entry currently happens to. `serve` takes `layer_registry`
+and `layer_widgets` for that: a project-package entry and its module, written as a
+`render_cases_*` constant and reachable from any test module. Borrowing a shipped tag
+reads the same for as long as the default package carries the declaration and goes red
+the day it stops, with nothing about the reading having changed — which is how the two
+guards over `seat_with_agent` came to rest on `lf-options` and failed when its
+`x-conversation` came off.
+
 Re-vendor before trusting a result that depends on runtime, theme, registry, or
 widget changes. A page directory owns the layer copied into it by `page init`; it
 does not read the checkout's current assets. A green render against a stale page is
diff --git a/tests/render_cases_interaction.py b/tests/render_cases_interaction.py
index c3231c7d2..fa08b7a2d 100644
--- a/tests/render_cases_interaction.py
+++ b/tests/render_cases_interaction.py
@@ -1254,3 +1254,85 @@ def resolve(a, b):
         "</lf-options></lf-decision>",
     },
 ]
+
+# A widget the shipped packages no longer have: one the reader owes an answer on that
+# also seats a conversation of its own. Those two facts together are what produce the
+# layer's one split between the reader's list and the unanswered decisions — a thread
+# standing in the seat while the agent has the next word takes the widget off the list
+# without answering it (`seat_with_agent`, `seatWithAgent`). `lf-options` supplied the
+# pair until 292de9c made the reader's cell an add form and dropped `x-conversation`, and
+# both halves of the split are still implemented, still described, and reachable by any
+# package that declares both. So the guard is declared here rather than borrowed from
+# whichever shipped entry happens to carry it: the reading under test is the layer's, and
+# the tag it runs on is the layer's business rather than the test's.
+SEATED_ASK_TAG = "lf-verdict"
+SEATED_ASK_ENTRY = {
+    "description": (
+        "A proposal the reader settles with one press and may talk over first in a seat "
+        "of its own. `asks` opens the decision; the press settles it."
+    ),
+    "type": "object",
+    "properties": {
+        "id": {"type": "string", "pattern": "^[a-z0-9][a-z0-9-]*$"},
+        "asks": {"type": "boolean"},
+        "restated": {"type": "boolean"},
+    },
+    "required": ["id"],
+    "additionalProperties": False,
+    "x-content": "prose",
+    "x-upgrade": True,
+    "x-state": {
+        "settle": {
+            "detail": {
+                "type": "object",
+                "properties": {"answer": {"type": "string", "enum": ["yes", "no"]}},
+                "required": ["answer"],
+                "additionalProperties": False,
+            },
+            "facet": "verdict",
+            "unit": "widget",
+            # The prerequisite the split is about. Read off the reader's list this
+            # refuses a press on a widget whose seat is mid-conversation; read off the
+            # unanswered decisions, which is what the door owes it, it lets one through.
+            "requires": {"target": "self", "awaiting": True},
+        }
+    },
+    "x-awaits": {"when": {"asks": [True]}, "answers": ["settle"]},
+    "x-conversation": {"when": {"asks": [True]}},
+    "x-example": '<lf-verdict id="verdict-example" asks>Ship it?</lf-verdict>',
+}
+# The press paints before it sends, which is what `lf-options` does with a pick and the
+# reason the browser door matters as much as the POST one: with the wrong list read here
+# the answer is already on the page, so a refusal is not a refusal the reader can see —
+# the control flips, nothing is logged, and the next poll puts it back saying nothing.
+SEATED_ASK_MODULE = """\
+import { conversationBox, offer, once, sendAction } from "/runtime/widget-api.js";
+
+customElements.define(
+  "lf-verdict",
+  class extends HTMLElement {
+    connectedCallback() {
+      if (!once(this)) return;
+      this.press = offer("button", "lf-settle", "Accept");
+      this.press.onclick = () => {
+        this.settled();
+        sendAction(this, "settle", { answer: "yes" });
+      };
+      this.append(this.press);
+      const seat = conversationBox(this, "Say something about this");
+      if (seat) this.append(seat);
+    }
+
+    settled() {
+      this.press.textContent = "Accepted";
+      this.press.setAttribute("aria-pressed", "true");
+    }
+
+    applyAction(action) {
+      if (action === "settle") this.settled();
+    }
+  },
+);
+"""
+SEATED_ASK_LAYER = {SEATED_ASK_TAG: SEATED_ASK_ENTRY}
+SEATED_ASK_WIDGETS = {f"{SEATED_ASK_TAG}.js": SEATED_ASK_MODULE}
diff --git a/tests/render_support.py b/tests/render_support.py
index 5a39520c0..be0c9c9bf 100644
--- a/tests/render_support.py
+++ b/tests/render_support.py
@@ -51,6 +51,8 @@
     ROOMS,
     ROSTER_PAGE,
     SCROLL_SETTLED,
+    SEATED_ASK_LAYER,
+    SEATED_ASK_WIDGETS,
     SEATED_QUESTION_PAGE,
     SETTLED_DECISION_PAGE,
     SHORT_SUGGESTION,
@@ -508,6 +510,8 @@
     "SCROLL_SETTLED",
     "SCROLL_SETTLE_MS",
     "SCROLL_STILL",
+    "SEATED_ASK_LAYER",
+    "SEATED_ASK_WIDGETS",
     "SEATED_QUESTION_PAGE",
     "SENTENCE",
     "SETTLED_DECISION_PAGE",
diff --git a/tests/test_render_controls.py b/tests/test_render_controls.py
index 7cbb2fef3..e8b55deaa 100644
--- a/tests/test_render_controls.py
+++ b/tests/test_render_controls.py
@@ -36,6 +36,8 @@
     SCROLL_SETTLE_MS,
     SCROLL_STILL,
     SCROLLED,
+    SEATED_ASK_LAYER,
+    SEATED_ASK_WIDGETS,
     SUGGESTION_PAGE,
     TOKEN,
     UNBREAKABLE_PAGE,
@@ -1344,35 +1346,34 @@ def test_a_self_eligibility_check_reads_state_before_its_optimistic_gesture(
     page.close()
 
 
-def test_a_seat_conversation_leaves_the_pick_it_is_about_live(
-    browser, serve, tmp_path, monkeypatch
-):
+def test_a_seat_conversation_leaves_the_pick_it_is_about_live(browser, serve):
     """The reader's own remark must not lock the control it is a remark about.
 
-    A conversation standing in the group's seat takes the decision off the reader's
-    list — the banner stops counting it — but answers nothing, so the pick that
-    would answer it is still live. This is the browser half of the split, and the
-    half the reader meets first: the POST door only sees a hand-posted event, while
-    here `actionAvailable` paints the control and `sendAction` guards the press, and
-    `lf-options` has already painted the pick by the time either runs. Reading the
-    reader's list at this door therefore does not refuse the press so much as
-    swallow it — the option flips, nothing is logged, no toast fires, and the next
-    poll puts it back with nothing anywhere saying why."""
-    monkeypatch.chdir(tmp_path)
-    overlay = tmp_path / ".leaf"
-    overlay.mkdir()
-    standard = json.loads((schema_model.DEFAULT_PACKAGE / "registry.json").read_text())
-    options = standard["lf-options"]
-    options["x-state"]["choose"]["requires"] = {"target": "self", "awaiting": True}
-    (overlay / "registry.json").write_text(json.dumps({"lf-options": options}))
+    A conversation standing in the widget's seat takes the decision off the reader's
+    list — the banner stops counting it — but answers nothing, so the press that would
+    answer it is still live. This is the browser half of the split, and the half the
+    reader meets first: the POST door only sees a hand-posted event, while here
+    `actionAvailable` paints the control and `sendAction` guards the press, and the
+    module has already painted the answer by the time either runs. Reading the reader's
+    list at this door therefore does not refuse the press so much as swallow it — the
+    widget flips, nothing is logged, no toast fires, and the next poll puts it back with
+    nothing anywhere saying why.
+
+    The subject is the project widget SEATED_ASK_ENTRY declares rather than an entry out
+    of the default package, because the pair the split needs — a visible ask and a seat
+    of the widget's own — is a pair of declarations and not a tag. No shipped entry has
+    carried both since 292de9c took `x-conversation` off `lf-options`, and the reading
+    under test never asked which widget it was."""
     url = serve(
         leaf_page(
             "seated eligibility",
-            '<h1 id="heading">Choose</h1><lf-decision id="pick-decision"><h2>Which option?</h2>'
-            '<lf-options id="pick" choose>'
-            '<lf-option id="pick-a">A</lf-option>'
-            '<lf-option id="pick-b">B</lf-option></lf-options></lf-decision>',
-        )
+            '<h1 id="heading">Choose</h1><lf-decision id="pick-decision">'
+            "<h2>Cap the retries?</h2>"
+            '<lf-verdict id="pick" asks>Three attempts, then stop.</lf-verdict>'
+            "</lf-decision>",
+        ),
+        layer_registry=SEATED_ASK_LAYER,
+        layer_widgets=SEATED_ASK_WIDGETS,
     )
     events_model.append_event(
         serve.page_dir,
@@ -1388,14 +1389,14 @@ def test_a_seat_conversation_leaves_the_pick_it_is_about_live(
     # Off the reader's list, which is the whole reason the two readings differ here.
     expect(page.locator(".lf-decisions")).to_have_text("Asks (0)")
 
-    page.get_by_role("checkbox", name=re.compile(r"^choose one: A")).click()
+    page.get_by_role("button", name="Accept").click()
     round_trip(page)
 
-    expect(page.locator("#pick-a")).to_have_attribute("chosen", "")
-    # The log is what holds this, and the attribute above cannot: the module paints
-    # the pick before either guard runs, so with the wrong reading at this door the
-    # option wears `chosen` exactly as it does here and the log stays empty.
-    assert [event["action"] for event in actions(serve.page_dir)] == ["choose"]
+    expect(page.get_by_role("button", name="Accepted")).to_have_count(1)
+    # The log is what holds this, and the control above cannot: the module paints the
+    # answer before either guard runs, so with the wrong reading at this door the press
+    # reads exactly as it does here and the log stays empty.
+    assert [event["action"] for event in actions(serve.page_dir)] == ["settle"]
     assert errors == []
     page.close()
 
diff --git a/tests/test_render_navigation.py b/tests/test_render_navigation.py
index f0240da53..9ce457a69 100644
--- a/tests/test_render_navigation.py
+++ b/tests/test_render_navigation.py
@@ -23,6 +23,8 @@
     PANEL_PAGE,
     RENDERED,
     ROOT,
+    SEATED_ASK_LAYER,
+    SEATED_ASK_WIDGETS,
     TARGETS_PAGE,
     TOKEN,
     WHERE_I_STAND_PAGE,
@@ -4528,32 +4530,34 @@ def drop():
 def test_the_ring_holds_on_a_seat_the_agent_has_still_to_answer(browser, serve):
     """Where the reader is standing and what the reader still owes are two facts, and a
     widget mid-conversation with the agent is where they part. Its seat holds the words
-    the reader just wrote, its pick is unmade and its controls are live, and it has left
+    the reader just wrote, its answer is unmade and its controls are live, and it has left
     the banner and the tray because the next word there is the agent's — but the reader
     is standing in it all the same, and it is still the question they are working.
 
     Read off the reader's list, both the ring and `c` went with the count: the moment the
-    remark was sent the ring left from under the reader, and `c` slid from the seat they
-    were writing in down to whichever option their focus rested on. That is a different
-    conversation, not a shorter way into the same one — `{section: "shape"}` is the seat's
-    own anchor and `{section: "sh-steel"}` is not — so the next line of a remark landed
-    somewhere the first line was not. The agent's reply moved both back. Nothing the
-    reader did moved either, which is the whole of the complaint; the reply phase here is
-    what says the ring has stopped tracking the count rather than merely tracking it late.
+    remark was sent the ring left from under the reader, and `c` fell through from the
+    question to whichever item their focus happened to rest in. That is a different
+    conversation, not a shorter way into the same one — a remark on the widget is filed
+    where a remark on the question the widget stands as is not — so the next line of a
+    remark landed somewhere the first line was not. The agent's reply moved both back.
+    Nothing the reader did moved either, which is the whole of the complaint; the reply
+    phase here is what says the ring has stopped tracking the count rather than merely
+    tracking it late.
 
     A picked group is the control on the other side. It is answered, so it is off both
     readings and must stay off: the switch is about a seat the reader is mid-sentence in,
-    not about reopening what a pick has closed."""
+    not about reopening what a pick has closed.
+
+    The seat and the ask are the project widget SEATED_ASK_ENTRY declares, for the reason
+    test_a_seat_conversation_leaves_the_pick_it_is_about_live gives: the split is between
+    two declarations, and since 292de9c no shipped entry carries both."""
     url = serve(
         leaf_page(
             "mid-sentence",
             """
 <h1 id="t">Mid-sentence</h1>
-<lf-decision id="shape-decision"><h2>Which material?</h2>
-<lf-options id="shape" choose>
-  <lf-option id="sh-steel"><strong>Steel</strong> Galvanised, drop-in.</lf-option>
-  <lf-option id="sh-cedar"><strong>Cedar</strong> Cheap; needs sealing.</lf-option>
-</lf-options></lf-decision>
+<lf-decision id="shape-decision"><h2>Galvanised steel for the frame?</h2>
+<lf-verdict id="shape" asks>Drop-in, and it needs no sealing.</lf-verdict></lf-decision>
 <lf-decision id="picked-decision"><h2>Should we keep it?</h2>
 <lf-options id="picked" choose>
   <lf-option id="pk-keep" chosen><strong>Keep it</strong> Settled by a pick.</lf-option>
@@ -4565,7 +4569,9 @@ def test_the_ring_holds_on_a_seat_the_agent_has_still_to_answer(browser, serve):
   <lf-new>Refill when the camera shows it half-empty.</lf-new>
 </lf-suggestion></p>
 """,
-        )
+        ),
+        layer_registry=SEATED_ASK_LAYER,
+        layer_widgets=SEATED_ASK_WIDGETS,
     )
     d = serve.page_dir
     events_model.append_event(
@@ -4583,8 +4589,8 @@ def test_the_ring_holds_on_a_seat_the_agent_has_still_to_answer(browser, serve):
     line = page.locator(".lf-keyline")
     decisions = page.locator(".lf-decisions")
 
-    # The premise, from the reader's list itself: the group has left it, the picked group
-    # was never on it, and the suggestion is what remains to be counted.
+    # The premise, from the reader's list itself: the seated ask has left it, the picked
+    # group was never on it, and the suggestion is what remains to be counted.
     expect(decisions).to_have_text("Asks (1)")
     decisions.click()
     expect(page.locator("button.lf-decisions-row")).to_have_count(1)
@@ -4598,7 +4604,7 @@ def test_the_ring_holds_on_a_seat_the_agent_has_still_to_answer(browser, serve):
     # the scroll that brings a row into view is the tray's own reading. So the decision wears
     # the ring alone, and the tray goes on listing what the reader owes rather than
     # gaining a row for where they happen to be standing.
-    page.locator("#shape .lf-pick").first.focus()
+    page.locator("#shape .lf-settle").focus()
     expect(page.locator("#shape-decision")).to_have_attribute("data-lf-decision", "1")
     expect(page.locator("button.lf-decisions-row")).to_have_count(1)
     assert page.locator(".lf-decisions-row[data-lf-decision]").count() == 0, (
@@ -4608,7 +4614,7 @@ def test_the_ring_holds_on_a_seat_the_agent_has_still_to_answer(browser, serve):
     decisions.click()
 
     # And with it shut, which is every other reading below.
-    page.locator("#shape .lf-pick").first.focus()
+    page.locator("#shape .lf-settle").focus()
     expect(page.locator("#shape-decision")).to_have_attribute("data-lf-decision", "1")
     expect(line).to_contain_text("comment on the decision")
 
@@ -4634,7 +4640,7 @@ def test_the_ring_holds_on_a_seat_the_agent_has_still_to_answer(browser, serve):
         )
     told(page)
     expect(decisions).to_have_text("Asks (2)")
-    expect(page.locator("#shape .lf-pick").first).to_be_focused()
+    expect(page.locator("#shape .lf-settle")).to_be_focused()
     expect(page.locator("#shape-decision")).to_have_attribute("data-lf-decision", "1")
     expect(line).to_contain_text("comment on the decision")
     page.evaluate("() => document.activeElement?.blur()")