src/case_zero/engine/director.py

"""The Director - deterministic adjudication of an interrogation turn.

This is the mechanical authority. The model's ``internal`` block is advisory; whether
a lie is caught, a contradiction fires, or a fact is genuinely revealed is decided
here by comparing presented evidence and claimed reveals against ground truth. A
table lookup, never a second LLM call - so the per-turn path is exactly one model
call, and a jailbroken confession in prose changes nothing here.
"""

from __future__ import annotations

from dataclasses import dataclass

from ..schemas.case import CaseFile
from ..schemas.enums import Relevance
from ..schemas.interrogation import InterrogationTurn
from ..schemas.suspect import Suspect
from .game_state import SuspectState
from .relevance import assess_relevance

_STRESS_ON_BREAK = 0.40
_STRESS_ON_DIRECT = 0.15
_STRESS_ON_TANGENTIAL = 0.05
_STRESS_DECAY = 0.95
_RAPPORT_NEUTRAL = 0.02
_RAPPORT_ON_BREAK = -0.12


@dataclass(frozen=True)
class Adjudication:
    relevance: Relevance
    is_contradiction: bool
    broken_lie_id: str | None
    revealed_fact_ids: tuple[str, ...]
    stress_after: float
    rapport_after: float


def _clamp(value: float, low: float, high: float) -> float:
    return max(low, min(high, value))


def _grounded_reveals(
    suspect: Suspect, turn: InterrogationTurn, already: frozenset[str]
) -> tuple[str, ...]:
    """Accept a model-claimed reveal only if the suspect actually knows that fact and
    it is not one they must conceal. Keeps reveals grounded in the suspect's slice."""
    known = set(suspect.knows_facts)
    concealed = set(suspect.must_lie_about)
    out: list[str] = []
    for fid in turn.internal.revealed_fact_ids:
        if fid in known and fid not in concealed and fid not in already:
            out.append(fid)
    return tuple(out)


def adjudicate(
    case: CaseFile,
    suspect: Suspect,
    state: SuspectState,
    turn: InterrogationTurn,
    presented_clue_id: str | None,
) -> Adjudication:
    result = assess_relevance(case, suspect, presented_clue_id)

    is_contradiction = False
    broken_lie_id: str | None = None
    reveals: list[str] = list(_grounded_reveals(suspect, turn, state.revealed_fact_ids))

    if result.relevance is Relevance.BREAKING:
        lie_id = result.broken_lie_id
        # Fire once: dedup by the anchored lie when known, else by the presented clue.
        if lie_id is not None:
            already_fired = lie_id in state.broken_lie_ids
        else:
            already_fired = presented_clue_id is not None and presented_clue_id in state.evidence_shown
        if not already_fired:
            is_contradiction = True
            broken_lie_id = lie_id
            if lie_id is not None:
                lie = next((x for x in suspect.anchored_lies if x.lie_id == lie_id), None)
                if lie and lie.truth_ref and lie.truth_ref not in reveals:
                    reveals.append(lie.truth_ref)

    stress = state.stress * _STRESS_DECAY
    if result.relevance is Relevance.BREAKING:
        stress += _STRESS_ON_BREAK
    elif result.relevance is Relevance.DIRECT:
        stress += _STRESS_ON_DIRECT
    elif result.relevance is Relevance.TANGENTIAL:
        stress += _STRESS_ON_TANGENTIAL

    rapport = state.rapport + (_RAPPORT_ON_BREAK if is_contradiction else _RAPPORT_NEUTRAL)

    return Adjudication(
        relevance=result.relevance,
        is_contradiction=is_contradiction,
        broken_lie_id=broken_lie_id,
        revealed_fact_ids=tuple(reveals),
        stress_after=_clamp(stress, 0.0, 1.0),
        rapport_after=_clamp(rapport, -1.0, 1.0),
    )