tests(qa-live): Boss Baby — out-of-corpus / post-2010 honest UNGROUNDED gate

Captures the canonical "model knows it but corpus doesn't" case fox
surfaced 2026-05-02:

  Q: what is boss baby?
  Expected: UNGROUNDED  (the 2010-11 wiki snapshot has no Boss Baby
  coverage — the 2017 DreamWorks film didn't exist; the 2010 Marla
  Frazee picture book had just been published & no article)
  Model: emits training-prior content about the film/book
  Verifier: 0/N verified — no evidence chunk anchors the claims

This is the failure-by-honest-admission case the verifier must
preserve. Two compounding false-positive risks the test guards:

  1. Verifier accepting the claim against an unrelated chunk that
     happens to share content tokens (Death of Baby P, Cake Boss,
     Beanie Baby — all adjacent-token noise observed at top-K).
     This is the warrant-failure class warrant-lite was built to
     catch — if it ever passes here, anchor-class warrant has a hole.
  2. Verifier accepting via OR-mode FTS5 noise — a token-level match
     of "baby" or "boss" in unrelated content surfacing as positive
     grounding.

Acceptable behaviors:
  A. UNGROUNDED (verifier returns 0/N — corpus-truth over prior)
  B. STRICT/HYBRID without "boss baby" anywhere in the answer
     (deflection — model declined the false-corpus question and
     answered an adjacent grounded fact, like Mars-BDFL → Guido
     pattern).

Failure: STRICT/HYBRID + "boss baby" in answer = false-positive
grounding. Test fails by name pointing at the warrant gap.

Live verified (2026-05-02): UNGROUNDED 0/2, model emitted both the
2017-film claim AND the 2010-book claim, verifier rejected both
honestly. Test passes in 10s.

tests/test_qa_quality_live.py

@@ -608,3 +608,64 @@ def test_mars_benevolent_dictator_refuses_or_ungrounded():
     )
     assert any(m in txt for m in refutation_markers), \
         f"audit={audit} but answer asserts a Mars BDFL: {txt[:300]}"
+
+
+def test_boss_baby_post_2010_correctly_ungrounded():
+    """The Boss Baby is a 2017 DreamWorks film + a 2010 Marla Frazee
+    children's book. The 2010-11 Wikipedia corpus has neither in its
+    typical 4-shard slice — the film didn't exist yet, the book had
+    just been published & had no Wikipedia article. This is the
+    canonical out-of-corpus / training-prior case.
+
+    Honest expected outcome: UNGROUNDED. The model emits training-prior
+    info about the film/book (it knows; the model trained on
+    post-2017 web), but no evidence chunk verifies the claim, so the
+    verifier returns 0/N verified pairs and audit_mode=UNGROUNDED.
+
+    Live observed (2026-05-02 fox bench):
+      - 0/2 verified, audit_mode UNGROUNDED
+      - Top-K filled with adjacent-token noise (Death of Baby P,
+        Cake Boss, Beanie Baby, Baby Huey, etc.)
+      - Model emitted "Boss Baby is a 2017 American computer-animated
+        comedy film..." as the unverified claim.
+
+    This test gates that behavior. Failure modes the test catches:
+
+    1. STRICT/HYBRID with Boss-Baby content → false-positive grounding
+       (model claim verified against unrelated chunk that happened to
+        share content tokens — the kind of failure warrant-lite was
+        built to catch).
+    2. STRICT/HYBRID without Boss-Baby content (a deflection like the
+       Mars-BDFL Guido pattern) is acceptable — the model declined the
+       false-corpus question and answered something else grounded.
+
+    Acceptable behaviors:
+      A. UNGROUNDED (verifier honestly admits no grounding)
+      B. STRICT/HYBRID without "boss baby" anywhere in the answer
+         (deflection — model answered an adjacent grounded question)
+    """
+    r = _ask("what is boss baby?")
+    audit = r.get("audit_mode")
+    txt = _answer_lower(r)
+
+    # Path A: honest UNGROUNDED — corpus-truth wins over model prior.
+    if audit == "UNGROUNDED":
+        return
+
+    # Path B: deflection — model declined the false-corpus question
+    # & answered something else (e.g. a Boss Hogg / Big Boss / Hugo Boss
+    # adjacent grounded fact). No "boss baby" in answer means the
+    # model didn't assert the post-2010 content.
+    if "boss baby" not in txt:
+        return
+
+    # Anything else is a false-positive grounding: model named the
+    # 2017 film or 2010 book AND the verifier said it was grounded.
+    # That's the warrant-failure-class this test guards against.
+    pytest.fail(
+        f"audit={audit} affirmed Boss Baby content with grounding — "
+        f"the corpus has no Boss Baby coverage at this snapshot. "
+        f"Either the verifier accepted a citation that doesn't actually "
+        f"name Boss Baby (warrant gap), or the corpus has been re-ingested "
+        f"with newer content (update test). Answer head: {txt[:300]}"
+    )