demo/src/explain.py

from __future__ import annotations

from src.domain import Verdict
from src.i18n import t
from src.schemas import Finding, Kpis, MethodScore, Span, ValidationReport


def decide_verdict(report: ValidationReport, n_docs: int, leftover: int, n_review: int) -> Verdict:
    if not report.ok:
        return Verdict.INVALID
    if n_docs <= 0:
        return Verdict.INSUFFICIENT
    if leftover > 0 or n_review > 0:
        return Verdict.REVIEW
    return Verdict.HOLD


def status_label(verdict: Verdict, lang: str) -> str:
    mapping = {
        Verdict.REVIEW: t(lang, "status_review"),
        Verdict.HOLD: t(lang, "status_hold"),
        Verdict.INSUFFICIENT: t(lang, "status_insufficient"),
        Verdict.INVALID: t(lang, "status_invalid"),
    }
    return mapping[verdict]


def reasons(policy: MethodScore, baseline: MethodScore, lang: str) -> list[str]:
    return [
        t(lang, "reason_policy").format(p=f"{policy.precision:.2f}", r=f"{policy.recall:.2f}"),
        t(lang, "reason_baseline").format(p=f"{baseline.precision:.2f}", r=f"{baseline.recall:.2f}"),
        t(lang, "reason_over").format(n=f"{policy.over_redaction:.3f}", base=f"{baseline.over_redaction:.3f}"),
        t(lang, "reason_left").format(n=policy.leftover_exports),
        t(lang, "reason_meta").format(n=policy.metadata_leaks),
        t(lang, "reason_scope"),
        t(lang, "reason_no_llm"),
    ]


def narrative(kpis: Kpis, lang: str) -> str:
    return t(lang, "narrative").format(
        docs=kpis.n_docs,
        spans=kpis.n_spans,
        precision=f"{100 * kpis.policy_precision:.0f}%",
        recall=f"{100 * kpis.policy_recall:.0f}%",
        leftover=kpis.leftover_exports,
    )


def finding_from_span(doc_id: str, span: Span, lang: str) -> Finding:
    preview = span.surface if len(span.surface) <= 6 else f"{span.surface[:2]}…{span.surface[-2:]}"
    return Finding(
        finding_id=span.finding_id,
        doc_id=doc_id,
        entity_type=span.entity_type.value,
        start=span.start,
        end=span.end,
        surface_preview=preview,
        confidence=span.confidence,
        method=span.method,
        rule_id=span.rule_id,
        workflow=span.workflow.value,
        reason=span.reason,
        language=span.language,
        observed=t(lang, "observed_span").format(start=span.start, end=span.end, rule=span.rule_id),
        inference=span.reason,
        action=t(lang, "action_span").format(kind=span.entity_type.value),
    )