configs/monitoring.py

"""Monitoring dimension config — v8b (dense Qwen3-8B).

PATTERNS define the 5 metacognition-theory families (backtrack, verify,
uncertainty, self_correct, strategy_switch). They are used by
BehaviorDetector at evaluation time (stages 03/04) ONLY. The steering
DIRECTION itself is learned via contrastive prompting in stage 00, NOT
via these regexes.

TARGET_LAYERS:
  Qwen3-8B is a DENSE transformer with 36 blocks (L0..L35), hidden 4096,
  no MoE. We sweep L2..L35 (skip the first two embedding-dominated
  layers). Stage 03 keeps a per-layer cache and [resume]s any layer it
  already calibrated, so an interrupted job only recomputes new layers.
"""

NAME = "monitoring"

# 5-family pattern dictionary. Evaluation-side only.
PATTERNS = {
    # Family 1: BACKTRACK / RESTART (Schoenfeld control behaviour)
    "backtrack": [
        r"(?i)\bwait\b[,.]?\s",
        r"(?i)\bactually\b[,.]?\s",
        r"(?i)\bhmm\b",
        r"(?i)\b(no|hmm)[,.]?\s+that('?s|\s+is)\s+(not|wrong|incorrect)",
        r"(?i)\blet\s+me\s+(reconsider|rethink|redo|go\s+back|re-?examine)",
        r"(?i)\b(going|go)\s+back\s+to",
        r"(?i)\bthis\s+(is|was)\s+(wrong|incorrect|not\s+right)",
        r"(?i)\bthat\s+doesn'?t\s+(work|seem\s+right|make\s+sense)",
        r"(?i)\bbut\s+wait",
        r"(?i)\boh\s+wait",
        r"(?i)\bhold\s+on",
        r"(?i)\bon\s+second\s+thought",
        r"(?i)\bscratch\s+that",
        r"(?i)\bnever\s+mind",
    ],
    # Family 2: VERIFICATION / LOOK-BACK (Polya look-back)
    "verify": [
        r"(?i)\blet\s+me\s+(check|verify|confirm|double[- ]?check|validate)",
        r"(?i)\bto\s+(check|verify|confirm)\b",
        r"(?i)\b(double-?)?check\s+(this|that|my|the)",
        r"(?i)\bsubstitut(e|ing)\s+(back|into|this)",
        r"(?i)\bplug(ging)?\s+(back\s+)?(in|into)\b",
        r"(?i)\bsanity\s+check",
        r"(?i)\bdoes\s+(this|that)\s+(make\s+sense|check\s+out|hold|work)",
        r"(?i)\bis\s+(this|that)\s+(right|correct)",
        r"(?i)\blet\s+me\s+make\s+sure",
        r"(?i)\blet\s+me\s+test",
        r"(?i)\bcross-?check",
        r"(?i)\bvalidate(d|s)?\b",
    ],
    # Family 3: UNCERTAINTY / HEDGING (Flavell metacognitive experiences)
    "uncertainty": [
        r"(?i)\bnot\s+(entirely|completely|fully|quite)?\s*sure\b",
        r"(?i)\bunsure\b",
        r"(?i)\bi'?m\s+not\s+(entirely|completely|fully)?\s*(sure|certain)\b",
        r"(?i)\bperhaps\b",
        r"(?i)\bmaybe\b",
        r"(?i)\bpossibly\b",
        r"(?i)\bi\s+(wonder|suspect)\s+(if|whether)",
        r"(?i)\bi\s+(think|guess|believe)\b",
        r"(?i)\bi'?m\s+guessing\b",
        r"(?i)\bcould\s+be\b",
        r"(?i)\bmight\s+be\b",
        r"(?i)\bseems?\s+(like|to\s+be)\b",
        r"(?i)\bappears?\s+to\b",
        r"(?i)\bif\s+i'?m\s+(right|correct|not\s+mistaken)",
        r"(?i)\bunless\s+i'?m\s+mistaken",
    ],
    # Family 4: SELF-CORRECTION (admitting error)
    "self_correct": [
        r"(?i)\bi\s+(must|might|may)\s+have\s+(made|been\s+making)\s+(a\s+|an\s+)?(mistake|error)",
        r"(?i)\bi\s+(made|have)\s+a?\s*(mistake|error)",
        r"(?i)\bi\s+misread\b",
        r"(?i)\bi\s+mis(under)?stood\b",
        r"(?i)\bi\s+mis-?(typed|wrote|read|calculated|spoke)",
        r"(?i)\bi\s+got\s+confused\b",
        r"(?i)\bi\s+was\s+wrong\b",
        r"(?i)\bi'?ve\s+been\s+(doing|getting)\s+(this|it)\s+wrong",
        r"(?i)\bmy\s+(mistake|error|bad)\b",
        r"(?i)\bthat\s+was\s+wrong\b",
        r"(?i)\boops\b",
        r"(?i)\bcorrection\b[,.]",
        r"(?i)\bto\s+correct\s+(myself|this|that)\b",
    ],
    # Family 5: STRATEGY SWITCH (Schoenfeld strategy management)
    "strategy_switch": [
        r"(?i)\b(try|use)\s+(a\s+)?(different|another|alternative)\s+(approach|method|way|strategy|technique)",
        r"(?i)\binstead[,.]?\s+(let|i|i'?ll)",
        r"(?i)\blet\s+me\s+try\s+(a\s+different|another|something\s+else)",
        r"(?i)\bthis\s+(approach|method)\s+(isn'?t|doesn'?t|won'?t)\s+work",
        r"(?i)\bperhaps\s+(a\s+)?(better|different|another)\s+(way|approach|method)",
        r"(?i)\balternatively[,.]",
        r"(?i)\b(a|the)\s+(better|easier|cleaner|simpler|more\s+elegant)\s+(way|approach|method)",
        r"(?i)\bmore\s+efficient(ly)?\b",
        r"(?i)\bswitch\s+(to|approaches?|tactics?)",
        r"(?i)\bchange\s+(strategy|approach|tack)",
        r"(?i)\baha\b",
        r"(?i)\bah,?\s+(i\s+see|of\s+course|right)",
        r"(?i)\bkey\s+insight",
        r"(?i)\bthe\s+trick\s+is",
    ],
}

# ============================================================
# Alpha schedules
# ============================================================
# Inference sweep (stage 04): includes the 1.0 baseline.
SWEEP_ALPHAS = [0.0, 0.3, 0.7, 1.0]
# Per-layer calibration sweep (stage 03): the three non-trivial alphas.
NON_TRIVIAL_ALPHAS = [0.7, 0.3, 0.0]

# ============================================================
# Calibration (stage 03)
# ============================================================
CALIBRATION_N_TEST = 15            # 15 held-out problems used to calibrate
CALIBRATION_ACTIVE_THRESHOLD = 2   # a baseline CoT is "active" if >= this many markers
CALIBRATION_SIDE_EFFECT_RATE = 0.0
SEVERE_REPETITION = 0.3

# ============================================================
# TARGET LAYERS — dense Qwen3-8B (36 blocks, L0..L35)
# ============================================================
TARGET_LAYERS = list(range(2, 36))   # 34 layers: L2 .. L35

# Training corpus size (stage 00 / stage 01).
N_TRAIN_COTS = 200

# Stage 01 (contrastive activation capture).
SAMPLES_PER_COT = 64
MAX_SEQ_LEN_FOR_CAPTURE = 4096

# Generation.
GEN_MAX_NEW_TOKENS = 32768

# Inference sampling defaults (stage 04).
DEFAULT_TEMPERATURE = 0.6
DEFAULT_TOP_P = 0.95