Vagueness-reactive detectives: branched question banks, prompt steering, stonewall rule

The precinct now has procedures for fog. A deterministic vagueness read on each
answer (hedges, length, checkable-detail heuristics) drives three layers:

1. Branched fallback banks per detective — fq_specific (drilling questions,
{C}-quoting), fq_vague (the fog itself becomes the subject, {A} quotes the
player's own vague answer back at them), fq_streak (2+ vague in a row:
Gimlet goes off the laminate, the Inspector goes glacial). Plus vague_waits
decks (the room reacting to fog) and per-character one-word callouts.
2. Prompt steering — the model is told whether the last answer was vague,
stonewalling (streak), or detailed, and how to play each.
3. A gameplay rule with noir logic: stonewall him 3+ times and the accuse
gate loosens — he stops waiting for evidence and goes with his gut.

Vague flags recorded per turn + behavior stats in the dataset record.
Writers-room material (4 lenses + showrunner): Form 12-B, "{A} is not an
answer, it's weather", "The rain has more content", Bellbird recurrences.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Md6C7wMyiQuBeiDR1M9x25

app.py

@@ -178,15 +178,51 @@ CHARACTERS = {
             "He sits, stands, moves the chair one inch closer, sits again. Ready.",
             "\"Right.\" He looks at your three claims like they owe him money.",
         ],
-        "fallback_questions": [
-            "{C}. Say it again, but slower. Slower than that.",
-            "Who else was there when {C} happened? Names. First names are fine. Actually, all the names.",
-            "What day of the week was {C}? Everyone remembers the day. Liars pick Tuesday.",
-            "Tell me {C} again, but look at the mirror while you do it. No reason.",
-            "{C} — what happened right after? Liars never plan the after.",
-            "One to ten, how proud are you of {C}? There are wrong answers.",
-            "I'm circling {C} right now. On paper. Does that worry you? Why not?",
-            "If {C} went to court, who's your witness? Pets count. Do pets count? Pets count.",
+        "fq_specific": [
+            "You said {C}. What day of the week? Don't look up. People look up when they're building.",
+            "{C} — okay, what time did it end? Not start. Everyone rehearses the start.",
+            "{C}. What's the cross street? Everywhere has a cross street. Except lies.",
+            "Say {C} again, word for word. If it comes out identical, that's worse, actually.",
+            "{C}. What did it smell like? Liars forget smells. It's step six. Or seven. It's laminated.",
+            "You said {C}. Which hand? People always know which hand. Take your time. Not too much time.",
+            "How much did that cost? Round numbers make me nervous. Give me the cents.",
+            "Who can confirm that? A name. Not 'a mate'. Mates have names. That's what makes them mates.",
+            "Name one person who saw it. I'll wait. I'm fast, but I'll wait.",
+            "Was there a receipt? There's always a receipt. The receipt is the truth's shadow.",
+        ],
+        "fq_vague": [
+            "You said {A}. I wrote it down, and now my notes are worse than before you spoke.",
+            "{A} is not an answer, it's weather. Give me furniture. A doorway. Something with edges.",
+            "I've underlined {A} three times and it means less each time.",
+            "{A}. My pencil got all the way to the paper and had nowhere to go. Do you see what you're doing to us.",
+            "Form 12-B has a section called Particulars. It's blank. It's looking at me. Fix it.",
+            "One fact. A name, a time, a number. Any of the three. Pick your favourite.",
+            "Who. Just who. Give me a who and I'll do the what and the when myself.",
+            "Was it before lunch or after lunch? Everything in this world is before lunch or after lunch.",
+            "How many? You shrugged. Numbers don't shrug. How many?",
+            "Can't remember, or won't remember? There are different boxes for those. Neither is the good box.",
+        ],
+        "fq_streak": [
+            "{A}. And before that, also fog. One more fog and I'm starting this interview over. From my name.",
+            "{A}. Again. Do you know what I've written this whole session? A boat. I drew a boat.",
+            "Sixty-one days on the job and nobody has ever put {A} in a report. I won't be first.",
+            "Form 12-B has nine boxes. I have filled in the date. The date was me. I did the date.",
+            "Is this a technique? Are you doing a technique on me? Because it's working and I hate it.",
+            "The Nine Steps don't cover this. I've checked twice. I'm off the card. I'm off the laminate.",
+        ],
+        "vague_waits": [
+            "He whispers 'gotcha,' then, quieter, 'no.'",
+            "Somewhere behind him, a pencil snaps. He is not holding a pencil.",
+            "He draws the boat. He notices he is drawing the boat. He stops. He finishes the boat.",
+            "He pours a very small coffee from a very large thermos and does not drink it.",
+            "He mouths 'okay' four times. The fourth one doesn't make it.",
+            "Behind the glass, Sgt. Wattle writes one word. It is not a long word.",
+            "Gimlet holds his pencil very still, the way you hold something you love.",
+            "He turns the notepad sideways, in case the answer has a landscape orientation.",
+        ],
+        "one_word": [
+            "That's two one-word answers. Hang on, there's a form for this. There is no form for this.",
+            "Fine. I'll ask questions of one word too. Where. When. Who. See? It's awful. Nobody wins.",
         ],
     },
     "marlowe": {
@@ -241,15 +277,51 @@ CHARACTERS = {
             "The sound of a great weight consenting, slowly, to be vertical.",
             "The door opens. He enters with the file already open, as if he was never asleep.",
         ],
-        "fallback_questions": [
-            "Say the part about {C} again. Slower.",
-            "What did {C} cost you?",
-            "The part of {C} you rehearsed. Say the other part.",
-            "{C}. Tell it backwards.",
-            "Describe the weather during {C}. The weather is rarely wrong.",
-            "What happened the day after {C}? A true thing continues.",
-            "Tell me {C} once more. I will not be listening to the words.",
-            "If I visited the place in {C} tomorrow, what would I find?",
+        "fq_specific": [
+            "{C}. Mm. And the second time you told that story — what changed?",
+            "You said {C}. Tell me the part you always leave out.",
+            "{C}. What was the second thing you noticed? Everyone rehearses the first.",
+            "{C}. Which detail are you proudest of? That is the one I will be checking.",
+            "{C}. What did you almost do instead?",
+            "How long did {C} take? People invent events. They rarely invent durations.",
+            "What happened an hour before {C}? Lies begin abruptly. Lives do not.",
+            "Tell it backwards, from the end. The truth survives that. Little else does.",
+            "Name one person who was there. I won't contact them. Knowing I could is enough.",
+            "My watch stopped years ago. Still right twice a day. Your story — once, so far.",
+        ],
+        "fq_vague": [
+            "{A}. Mm. A man told me that in 1991. He was protecting his brother.",
+            "{A}. That isn't memory failing. Memory fails specific. That was built.",
+            "{A}. Say it again, slower. Listen to yourself say it.",
+            "Fog is a choice. Choose a name instead.",
+            "A time. Just a time. The innocent always know roughly when.",
+            "I don't mind fog. I've sat in it since 1987. It clears. Usually around the third question.",
+            "People remember what happened to them. They forget what they invented. Which is this?",
+            "Tell me the time of day. If you cannot, tell me the light. People remember the light.",
+            "One detail. Choose it carefully. I will treat it as though you did.",
+            "A number. Any number that touches your story. Even the wrong number tells me where you keep them.",
+        ],
+        "fq_streak": [
+            "Twice now. Mm. The second fog is never an accident.",
+            "{A}. And before that, the same weather. Mm. I have stopped writing. Notice that I have stopped.",
+            "{A}. Say one true thing. Small. I will know it by the relief.",
+            "{A}. Again. Mm. The rain has more content.",
+            "You've said nothing twice, and beautifully. The BELLBIRD file began this way.",
+            "We can continue this way. I sleep twenty hours a day. I am rested. Are you?",
+        ],
+        "vague_waits": [
+            "The Inspector says nothing. It is a specific nothing. It is aimed.",
+            "Rain on the window. He lets it do the talking for a while.",
+            "He looks at the stopped pocket watch. It is correct.",
+            "He touches the cold tea. He does not drink it. He has never drunk it.",
+            "He turns one page of the slow file. He was not reading it. The gesture lands anyway.",
+            "His eyes almost close. The room holds its breath. They open. Something has been decided.",
+            "He breathes out through his nose. In some precincts this counts as shouting.",
+            "The BELLBIRD file sits under his hand. It does not come up. Its presence comes up.",
+        ],
+        "one_word": [
+            "Two answers. Two words. Mm. Economy. I respect it. I don't believe it.",
+            "The truth is often short. It is never this short twice.",
         ],
     },
 }
@@ -789,12 +861,28 @@ def _detective_prompt(char_key, claims, qa, observer_json, force_accuse):
             pacing = "The clock is ticking. If the picture is clear, accuse."
         task = (f"You have asked {n} of {MAX_QUESTIONS} questions. {pacing}\n\n"
                 + DETECTIVE_JSON_SPEC)
+    # vagueness steering — how the detective plays depends on what he's being given
+    last_vague, streak, vcount = _vague_stats(qa)
+    if streak >= 2:
+        vag = (f"\n\nTHE SUSPECT IS STONEWALLING: {streak} vague answers in a row "
+               f"({vcount} total). Make the vagueness itself the subject. Demand ONE "
+               "concrete fact — a name, a time, a number, a place — and do not accept fog.")
+    elif last_vague:
+        vag = ("\n\nThe suspect's last answer was VAGUE — no checkable detail. Call it "
+               "out in your own dry way, then ask ONE question that pins down a single "
+               "concrete fact: a name, a time, a number, a place.")
+    elif [e for e in qa if e.get("a")]:
+        vag = ("\n\nTheir last answer contained real detail. Test it: cross-reference "
+               "it against the claims and their earlier answers, and ask for the detail "
+               "a liar would not have prepared.")
+    else:
+        vag = ""
     observer = ""
     if observer_json:
         observer = ("\n\nCONFIDENTIAL OBSERVER NOTES from the observation room "
                     "(the suspect cannot hear this):\n" + observer_json)
     system = (
-        f"{c['voice']}\n\n"
+        f"{c['voice']}{vag}\n\n"
         "You are playing Two Truths and a Lie. The suspect filed three claims about "
         "themselves; EXACTLY ONE is false. Interrogate with short, concrete questions "
         "(under 25 words) about places, times, people, what happened right before or "
@@ -987,6 +1075,51 @@ def _norm_letter(v, fallback="B"):
     return _scan_letter(v) or fallback
 
 
+# ── vagueness: the precinct has procedures for fog ───────────────────────────
+# Deterministic read on each answer. Vague answers change how the detective
+# plays: different question banks, prompt steering, and — after a real
+# stonewall — he stops waiting for evidence and goes with his gut.
+
+_HEDGE_WORDS = {"maybe", "probably", "perhaps", "possibly", "dunno", "whatever",
+                "stuff", "things", "somewhere", "someone", "something", "sometime",
+                "somehow", "kinda", "sorta", "ish"}
+_HEDGE_PHRASES = ("i think", "i guess", "kind of", "sort of", "not sure", "no idea",
+                  "don't remember", "dont remember", "can't remember", "cant remember",
+                  "don't know", "dont know", "hard to say", "it depends", "a while ago",
+                  "at some point", "here and there", "now and then", "or something",
+                  "can't recall", "cant recall", "who knows")
+
+
+def _is_vague_answer(text):
+    t = " ".join((text or "").lower().split())
+    words = re.findall(r"[a-z']+", t)
+    if len(words) <= 3:
+        return True
+    hedges = sum(w in _HEDGE_WORDS for w in words) + sum(p in t for p in _HEDGE_PHRASES)
+    has_number = bool(re.search(r"\d", t))
+    content = [w for w in words if w not in _QSTOP and w not in _HEDGE_WORDS and len(w) > 3]
+    if hedges >= 2:
+        return True
+    if hedges >= 1 and len(content) < 4 and not has_number:
+        return True
+    if len(content) < 2 and not has_number:
+        return True
+    return False
+
+
+def _vague_stats(qa):
+    """(last_answer_was_vague, current_streak, total_count) over answered entries."""
+    answered = [e for e in qa if e.get("a")]
+    streak = 0
+    for e in reversed(answered):
+        if e.get("vague"):
+            streak += 1
+        else:
+            break
+    total = sum(1 for e in answered if e.get("vague"))
+    return (bool(answered and answered[-1].get("vague")), streak, total)
+
+
 # ── claims are spoken as WORDS, never letters ────────────────────────────────
 # Players don't remember which claim was "B". The letters exist only as JSON
 # indices and small visual tags; every spoken reference quotes the claim itself.
@@ -1101,15 +1234,20 @@ def parse_analyst(raw):
     return out
 
 
-def _accuse_allowed(n_answered, det_conf, analyst_profile):
-    """Spec pacing: threshold decay + snap-accuse guard."""
+def _accuse_allowed(n_answered, det_conf, analyst_profile, vague_streak=0):
+    """Spec pacing: threshold decay + snap-accuse guard. A real stonewall (3+
+    vague answers in a row) loosens the gate — if the suspect gives him nothing,
+    the detective stops waiting for evidence and goes with his gut."""
     if n_answered < MIN_Q_BEFORE_ACCUSE:
         return False
     pct = CONF_PCT[det_conf]
     thresh = ACCUSE_THRESHOLD.get(n_answered, ACCUSE_FLOOR)
+    stonewalled = vague_streak >= 3
+    if stonewalled:
+        thresh = max(ACCUSE_FLOOR, thresh - 10)
     if pct < thresh:
         return False
-    if n_answered < 8:
+    if n_answered < 8 and not stonewalled:
         susp = max(SUSP_PCT[analyst_profile.get(f"suspicion_{k}", "low")] for k in "ABC") \
             if analyst_profile else 0
         if susp < SNAP_GUARD_MIN_ANALYST:
@@ -1176,7 +1314,10 @@ def build_record(st):
         "claims": claims,
         "labels": {"lie": st["lie_label"], "truths": truths},
         "conversation": [{"turn": e["turn"], "remark": e.get("remark", ""),
-                          "question": e["q"], "answer": e.get("a")} for e in st["qa"]],
+                          "question": e["q"], "answer": e.get("a"),
+                          "vague": bool(e.get("vague"))} for e in st["qa"]],
+        "behavior": {"vague_answers": sum(1 for e in st["qa"] if e.get("vague")),
+                     "final_vague_streak": _vague_stats(st["qa"])[1]},
         "hidden_analysis": st["hidden_analysis"],
         "guess": {"guesser": "model",
                   "selected_lie": acc.get("lie"),
@@ -1267,11 +1408,19 @@ def _stamp_html(n):
             f'<div class="track"><i style="width:{pct}%"></i></div></div>')
 
 
-def _next_wait_line(st, escalated):
-    """No-repeat wait-line deck per character per session; escalation lines gated 15s+."""
+def _next_wait_line(st, escalated, vague=False):
+    """No-repeat wait-line deck per character per session; escalation lines gated
+    15s+. After a vague answer, the room reacts to the fog first."""
     key = st["model_key"]
     char = CHARACTERS[key]
     used = st["session"]["used_waits"][key]
+    if vague and not escalated:
+        vdeck = char.get("vague_waits", [])
+        fresh = [w for w in vdeck if w not in used]
+        if fresh:
+            line = random.choice(fresh)
+            used.append(line)
+            return line
     deck = char["waits_escalation"] if escalated else char["waits"]
     fresh = [w for w in deck if w not in used]
     if not fresh:
@@ -1307,16 +1456,32 @@ def _is_repeat_question(q, qa):
     return False
 
 
+def _fallback_bank_key(st):
+    """Pick the question bank by how forthcoming the suspect is being."""
+    _, streak, _ = _vague_stats(st["qa"])
+    if streak >= 2:
+        return "fq_streak"
+    if streak >= 1:
+        return "fq_vague"
+    return "fq_specific"
+
+
 def _fallback_question(st):
     char = CHARACTERS[st["model_key"]]
-    bank = [q for q in char["fallback_questions"] if q not in st["fallback_used"]]
+    key = _fallback_bank_key(st)
+    bank = [q for q in char[key] if q not in st["fallback_used"]]
+    if not bank:  # this bank exhausted this session — borrow, then reset
+        bank = [q for q in char["fq_specific"] if q not in st["fallback_used"]]
     if not bank:
         st["fallback_used"] = []
-        bank = char["fallback_questions"][:]
+        bank = char[key][:]
     q = random.choice(bank)
     st["fallback_used"].append(q)
     target = _least_questioned_claim(st["qa"], st["claims"])
-    return q.format(C=_claim_snippet(st["claims"][target]))
+    answered = [e for e in st["qa"] if e.get("a")]
+    last_ans = answered[-1]["a"] if answered else ""
+    return q.format(C=_claim_snippet(st["claims"][target]),
+                    A=_claim_snippet(last_ans, limit=36))
 
 
 def _chat_from_state(st, trailing=None):
@@ -1411,6 +1576,7 @@ def add_answer(answer, state, chat):
     if not a:
         raise gr.Error("The room waits. Say something — even liars say something.")
     st["qa"][-1]["a"] = a[:600]
+    st["qa"][-1]["vague"] = _is_vague_answer(a)
     if len(a) < 3:
         st["one_word_count"] += 1
     return "", st, _chat_from_state(st)
@@ -1448,11 +1614,13 @@ def detective_turn(state, chat):
     if first_wait:
         st["session"]["ritual_done"][st["model_key"]] = True
 
+    last_vague, vstreak, _ = _vague_stats(st["qa"])
     wait_start = time.time()
-    shown = ritual.pop(0) if ritual else _next_wait_line(st, escalated=False)
+    shown = ritual.pop(0) if ritual else _next_wait_line(st, escalated=False,
+                                                         vague=last_vague)
     if st["one_word_count"] >= 2 and not st["one_word_called"]:
         st["one_word_called"] = True
-        shown = ONE_WORD_CALLOUT
+        shown = random.choice(char.get("one_word", [ONE_WORD_CALLOUT]))
     outs.chat = _chat_from_state(st, trailing=f"*{shown}*")
     yield outs.tuple()
     next_line_at = wait_start + random.uniform(4, 6)
@@ -1469,7 +1637,8 @@ def detective_turn(state, chat):
                     if ritual:
                         shown = ritual.pop(0)
                     else:
-                        shown = _next_wait_line(st, escalated=(now - wait_start) > 15)
+                        shown = _next_wait_line(st, escalated=(now - wait_start) > 15,
+                                                vague=last_vague)
                     outs.chat = _chat_from_state(st, trailing=f"*{shown}*")
                     next_line_at = now + random.uniform(4, 6)
                     yield outs.tuple()
@@ -1500,7 +1669,8 @@ def detective_turn(state, chat):
     question = _humanize_letters(question, st["claims"])
     why = _humanize_letters(why, st["claims"])
 
-    if action == "accuse" and not force and not _accuse_allowed(n_answered, conf, profile):
+    if action == "accuse" and not force and \
+            not _accuse_allowed(n_answered, conf, profile, vague_streak=vstreak):
         action = "ask"           # blocked: too early / not confident enough — keep pressing
         question = ""
         fb = 3

tests/drive_mock.py

@@ -203,13 +203,53 @@ def extra_checks():
     st["model_key"] = "marlowe"
     st["qa"] = [{"turn": 1, "q": "Who else was there when the dog sighed?", "a": "x"}]
     fq = app._fallback_question(st)
-    check("fallback carries claim words, no {C}", "{C}" not in fq and "“" in fq, fq)
+    check("fallback has no unformatted slots", "{C}" not in fq and "{A}" not in fq, fq)
     check("least-questioned avoids the dog claim",
           app._least_questioned_claim(st["qa"], claims) in ("B", "C"))
 
     check("docket lists all claims (escaped)",
           all(app.html_lib.escape(claims[k]) in app._docket_html(claims) for k in "ABC"))
     check("empty docket has empty state", "No claims on file" in app._docket_html(None))
+
+    print("== vagueness detection ==")
+    check("'kind of' is vague", app._is_vague_answer("kind of"))
+    check("'I guess it was somewhere in town' is vague",
+          app._is_vague_answer("I guess it was somewhere in town"))
+    check("'can't remember, a while ago' is vague",
+          app._is_vague_answer("can't remember, it was a while ago"))
+    check("specific answer with detail is not vague",
+          not app._is_vague_answer("At the takeout place on Fifth Street, a Tuesday in March"))
+    check("specific with number is not vague",
+          not app._is_vague_answer("It took 3 weeks and my friend Sam helped"))
+
+    print("== vagueness-reactive play ==")
+    qa = [{"turn": 1, "q": "q1", "a": "somewhere I think", "vague": True},
+          {"turn": 2, "q": "q2", "a": "kind of", "vague": True}]
+    lv, streak, total = app._vague_stats(qa)
+    check("streak counted", lv and streak == 2 and total == 2, f"{lv},{streak},{total}")
+    stv = app.new_state()
+    stv["claims"] = claims
+    stv["model_key"] = "pip"
+    stv["qa"] = qa
+    check("streak bank selected", app._fallback_bank_key(stv) == "fq_streak")
+    stv["qa"] = qa[:1]
+    check("vague bank selected", app._fallback_bank_key(stv) == "fq_vague")
+    stv["qa"] = [{"turn": 1, "q": "q1", "a": "Fifth Street on a Tuesday", "vague": False}]
+    check("specific bank selected", app._fallback_bank_key(stv) == "fq_specific")
+    stv["qa"] = qa
+    fq2 = app._fallback_question(stv)
+    check("streak fallback formats {A}", "{A}" not in fq2 and "{C}" not in fq2, fq2)
+
+    prof_lo = {f"suspicion_{k}": "low" for k in "ABC"}
+    check("stonewall loosens gate (medium at Q6)",
+          app._accuse_allowed(6, "medium", prof_lo, vague_streak=3))
+    check("no stonewall keeps gate (medium at Q6)",
+          not app._accuse_allowed(6, "medium", prof_lo, vague_streak=2))
+
+    msgs = app._detective_prompt("pip", claims, qa, None, False)
+    check("prompt carries stonewall steering", "STONEWALLING" in msgs[0]["content"])
+    check("vague wait deck exists for both",
+          all(len(app.CHARACTERS[k]["vague_waits"]) >= 8 for k in ("pip", "marlowe")))
     return fails