feat(prospector): generalize rationalize.py to the full work-era corpus

Backward CoT distillation over all swept conversations (context + Quinn's actual
reply -> move + gold-anchored trace), full 10-move taxonomy. Produces the
(context -> trace -> move) LoRA training set labeled by what she really did.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

tooling/eval/rationalize.py

@@ -1,65 +1,70 @@
 #!/usr/bin/env python3
-"""Rationalize a labeled cluster into CoT training rows (STaR / backward distillation).
+"""Rationalize labeled conversations into CoT training rows (STaR / backward distill).
 
-Given (client_msg -> Quinn's ACTUAL reply), have the model infer the MOVE she ran
-and a short reasoning trace that leads from the message to that reply. Because the
-gold reply is provided, the rationale is anchored to what she really did (not a
-guess) -- the high-quality way to manufacture CoT training data. Rows that don't
-yield a valid move are dropped.
+Given (conversation context -> Quinn's ACTUAL next reply), infer the MOVE she ran
+and a one-sentence reasoning trace anchored to her real reply (not a forward
+guess). This is the high-quality way to manufacture the LoRA training set for
+move-classification: (context -> trace -> move), labeled by what she actually did.
 
-Output rows: {client_msg, gold_reply, move, trace} -> the (input -> reasoning ->
-move/reply) training set for LoRA-hardening move-classification.
+Input: a JSON list with `gold_reply` and either `context` (sweep) or `client_msg`
+(mined cluster). Default <DATA_DIR>/sweep_labels.json -> the full work-era corpus.
+Output: <DATA_DIR>/traincot_<input-stem>.json.
 
-Env: OSS_URL, DATA_DIR. Arg: cluster name (default bbc) -> reads cluster_<name>.json.
+Env: OSS_URL, DATA_DIR.  Arg: input filename (default sweep_labels.json).
 """
-import json, os, re, sys, urllib.request
+import json, os, sys, urllib.request
+from collections import Counter
 from concurrent.futures import ThreadPoolExecutor, as_completed
 
 OSS_URL = os.environ.get("OSS_URL", "http://localhost:8800/v1/chat/completions")
 DATA = os.environ.get("DATA_DIR", os.path.join(os.path.dirname(__file__), ".data"))
-NAME = sys.argv[1] if len(sys.argv) > 1 else "bbc"
-cluster = json.load(open(os.path.join(DATA, f"cluster_{NAME}.json")))
+INPUT = sys.argv[1] if len(sys.argv) > 1 else "sweep_labels.json"
+stem = os.path.splitext(os.path.basename(INPUT))[0]
+items = json.load(open(os.path.join(DATA, INPUT)))
 
-SYSTEM = """You build training data by analyzing how Quinn (a high-end companion) handled a message. You are GIVEN her actual reply, so infer her real reasoning -- do not invent a different reply.
+MOVES = ["opener", "qualify", "engage", "subhour", "address", "out_of_area", "of", "disengage", "escalate", "skip"]
+SYSTEM = f"""You build training data by analyzing how Quinn (a touring companion, $1000/hr, incall williamsburg NYC, text only, OnlyFans @transquinnftw) handled a conversation. You are GIVEN her actual next reply, so infer her REAL reasoning -- do not invent a different reply.
 
-Classify the MOVE she ran:
-- qualify: engaged a paying prospect / stated her rate / moved toward booking (even if he was crude or bragging, if she gave a rate or booking step this is qualify).
-- engage: answered a flirt/preference question, kept it warm, no rate yet.
-- disengage: brushed off a non-client / lowballer / time-waster / someone offering himself (terse, dismissive, or "can't talk").
-- of: redirected to OnlyFans.
-- address: withheld her address.
+Classify the MOVE she ran (one of: {', '.join(MOVES)}):
+- opener: answered a new hello with her intro.
+- qualify: engaged a paying prospect / gave rate / moved toward booking (even if he was crude, if she gave a rate/booking step it is qualify).
+- engage: answered a flirt/preference question, warm, no rate yet.
+- subhour: gave the <1hr / half-hour rate stance.
+- address: withheld her address when asked before a locked time.
+- out_of_area: told him she's not in his city / named her city.
+- of: redirected to OnlyFans (harvester / free-content / crude / out-of-budget).
+- disengage: brushed off a lowballer / time-waster / someone offering his body / insult.
+- escalate: a collab / photographer / business / personal-decision matter she handled herself.
+- skip: not a prospect (vendor, spam, bot, friend, automated).
 
-Then give a ONE-sentence trace from the client message to her reply: name the subject (his body/offer vs her/booking), his pay-intent, and why her move fits.
+Then give a ONE-sentence trace from the conversation to her reply: the subject (his body/offer vs her/booking), his pay-intent, and why her move fits.
 
-Output ONLY JSON: {"move":"qualify|engage|disengage|of|address","trace":"<one sentence>"}"""
+Output ONLY JSON: {{"move":"<one of the moves>","trace":"<one sentence>"}}"""
 
 SCHEMA = {"type": "object",
-          "properties": {"move": {"type": "string", "enum": ["qualify", "engage", "disengage", "of", "address"]},
-                         "trace": {"type": "string"}},
+          "properties": {"move": {"type": "string", "enum": MOVES}, "trace": {"type": "string"}},
           "required": ["move", "trace"], "additionalProperties": False}
 
-def rationalize(c):
-    user = f"CLIENT: {c['client_msg']}\nQUINN (actual reply): {c['quinn_reply_gold']}"
+def rationalize(it):
+    ctx = it.get("context") or ("CLIENT: " + it.get("client_msg", ""))
+    gold = it.get("gold_reply") or it.get("quinn_reply_gold", "")
+    user = f"{ctx}\nQUINN (actual reply): {gold}"
     body = json.dumps({"model": "quinn-oss",
                        "messages": [{"role": "system", "content": SYSTEM}, {"role": "user", "content": user}],
-                       "temperature": 0.2, "max_tokens": 300,
+                       "temperature": 0.2, "max_tokens": 250,
                        "response_format": {"type": "json_schema", "json_schema": {"name": "r", "schema": SCHEMA, "strict": True}}}).encode()
     req = urllib.request.Request(OSS_URL, data=body, headers={"Content-Type": "application/json"})
     d = json.loads(json.load(urllib.request.urlopen(req, timeout=120))["choices"][0]["message"]["content"])
-    return {"client_msg": c["client_msg"], "gold_reply": c["quinn_reply_gold"],
-            "move": d["move"], "trace": d["trace"]}
+    return {"context": ctx, "gold_reply": gold, "move": d["move"], "trace": d["trace"]}
 
 rows = []
-with ThreadPoolExecutor(max_workers=10) as ex:
-    futs = [ex.submit(rationalize, c) for c in cluster if c.get("quinn_reply_gold")]
+with ThreadPoolExecutor(max_workers=12) as ex:
+    futs = [ex.submit(rationalize, it) for it in items if (it.get("gold_reply") or it.get("quinn_reply_gold"))]
     for f in as_completed(futs):
-        try:
-            rows.append(f.result())
-        except Exception as e:
-            print("ERR", e, flush=True)
+        try: rows.append(f.result())
+        except Exception as e: print("ERR", e, flush=True)
 
-out = os.path.join(DATA, f"traincot_{NAME}.json")
+out = os.path.join(DATA, f"traincot_{stem}.json")
 json.dump(rows, open(out, "w"), ensure_ascii=False)
-import collections
 print(f"rationalized {len(rows)} -> {out}")
-print("move dist:", dict(collections.Counter(r["move"] for r in rows)))
+print("move dist:", dict(Counter(r["move"] for r in rows)))