"""Deterministic claim-vs-evidence verifier for the numeric spine.

A 7.8M model cannot reliably copy values, so the hard comparison is done by
rules instead of generation. The model's job is confined to what it actually
can do: identifying the claim and the evidence segments. This module:

  1. extracts values from the claim side and the evidence side,
  2. compares them deterministically,
  3. returns a verdict that cannot be hallucinated:
       supports / refutes / not enough information / unclear.

For any case it cannot resolve (no clean numeric pair), it says so -- it never
fabricates an answer.
"""
import re


def _nums(text):
    return re.findall(r"\b\d+(?:,\d{3})*\.?\d*%?\b", text)


def _years(text):
    return re.findall(r"\b(?:19|20)\d{2}\b", text)


def _times(text):
    return re.findall(r"\b\d{1,2}:\d{2}\b", text)


def _clean(v):
    return v.replace(",", "").replace("%", "")


def _vals(text):
    return sorted({_clean(v) for v in (_nums(text) + _years(text) + _times(text))})


def deterministic_verdict(doc):
    """doc: the analyst prompt (claim + evidence). Returns a verdict dict."""
    m = re.split(r"\bEvidence:\s*", doc, flags=re.IGNORECASE)
    claim, evidence = m[0], (m[1] if len(m) > 1 else "")
    cv = _vals(claim)
    ev = _vals(evidence)
    if not cv and not ev:
        return {"verdict": "not enough information", "kind": "no-values",
                "confidence": "LOW", "explain": "no numeric value to compare"}
    if cv and not ev:
        return {"verdict": "not enough information", "kind": "claim-only",
                "confidence": "HIGH", "explain": "evidence has no numeric value to compare"}
    if not cv:
        return {"verdict": "unclear", "kind": "no-claim-value",
                "confidence": "LOW", "explain": "claim has no numeric value"}
    if set(cv) == set(ev):
        return {"verdict": "supports", "kind": "equal", "confidence": "HIGH",
                "explain": f"claim value {sorted(cv)} equals evidence value {sorted(ev)}"}
    if not (set(cv) & set(ev)):
        return {"verdict": "refutes", "kind": "differ", "confidence": "HIGH",
                "explain": f"claim value {sorted(cv)} differs from evidence value {sorted(ev)}"}
    return {"verdict": "unclear", "kind": "partial", "confidence": "LOW",
            "explain": f"claim {sorted(cv)} partially overlaps evidence {sorted(ev)}"}