@dzhechkov/p-replicator 1.5.17 → 1.5.18

package/templates/.claude/skills/goap-research-ed25519/scripts/population_match.py

@@ -0,0 +1,591 @@
+#!/usr/bin/env python3
+"""
+Population applicability — *was this number obtained in people like this patient?*
+
+A signed, correctly-sourced fact can still mislead the one reader who matters. Four
+real reversals motivated this module (ADR-001 §"Context"): a testosterone/weight-loss
+finding measured in men with obesity applied to a patient at BMI 25; an
+erectile-dysfunction RCT that enrolled only BMI >= 30; TRAVERSE's safety reassurance
+established in high-cardiovascular-risk men; a +44.5% LDL effect observed at baseline
+triglycerides >= 800 mg/dL quoted at 236 mg/dL.
+
+Two criterion KINDS, because the four reversals are two different failures:
+
+    eligibility  — the patient would NOT have been enrolled          → verdict 'none'
+    baseline     — enrollable, but his starting value is outside the
+                   range the effect was measured FROM                → verdict 'partial'
+
+An inclusion-criteria-only model reproduces two of the four and calls it done.
+
+WHAT THIS MODULE DOES NOT CLAIM
+    * that `verbatim` was transcribed truthfully from the source — nothing here can
+      check that; it is printed next to every discrepancy so a human can;
+    * that the criterion the source used is the criterion that matters clinically;
+    * that a `full` verdict means the claim is true. It means the patient is inside
+      the population the number came from. Nothing more.
+
+Stdlib only. No network, no global state. THE EVALUATION PATH IS I/O-FREE —
+`parse_study_population`, `evaluate`, `derive_verdict`, `match`, `match_from_fact` and
+`render_population_match` never touch the filesystem. The single exception is
+`load_field_cases()`, which reads the committed fixture so the tests and the README
+example replay the SAME artifact; it is not on any evaluation path.
+
+This paragraph used to promise blanket purity while `load_field_cases()` opened a file
+two screens below (QE G8). The header is narrowed to the guarantee the code actually
+keeps, rather than the function being deleted to fit the header — and
+`test_population_match.py` asserts the exception stays the only one.
+"""
+
+from __future__ import annotations
+
+import json
+from dataclasses import dataclass, field as dc_field
+from typing import Any, Dict, Iterable, List, Mapping, Optional, Sequence, Tuple
+
+# --- Grammar (a new study never forces an edit here) --------------------------
+CRITERION_OPS = (">=", "<=", ">", "<", "==", "in", "range")
+CRITERION_KINDS = ("eligibility", "baseline")
+
+DISCREPANCY_KINDS = (
+    "eligibility-excluded",    # the patient would NOT have been enrolled      → 'none'
+    "baseline-out-of-range",   # enrollable, but his starting value is outside → 'partial'
+    "criterion-unstated",      # the paper does not state this axis            → 'unknown'
+    "patient-value-missing",   # the profile does not carry this axis          → 'unknown'
+)
+
+DISCREPANCY_DIRECTIONS = ("below", "above", "outside-set", "absent")
+
+POPULATION_MATCH_VERDICTS = ("full", "partial", "none", "unknown")
+
+# The `field` a wholly-unstated population reports. It is a SENTINEL, not an axis:
+# there is no criterion to name because the source named none. Exported so the report
+# gate can recognise it instead of demanding this literal token appear in the prose —
+# the bug that punished the sanctioned `StudyPopulation.unstated(reason)` path (QE G5).
+UNSTATED_POPULATION_FIELD = "(study population)"
+
+# ALLOWLIST OVER EMITTED STRUCTURE (ADR-001 §1, D-16's counterpart for output).
+# The renderer iterates this tuple, so a field added to Discrepancy without being
+# added here is simply NOT PRINTED — the safe direction, and the opposite of a
+# blocklist that must anticipate every bad word.
+DISCREPANCY_KEYS = (
+    "field", "kind", "patient_value", "study_requirement", "verbatim", "locator", "direction",
+)
+
+# --- Content wearing grammar's clothes (ADR-001 §1 accepted this knowingly) ----
+# A closed list of clinical axes. A thirteenth axis is a reviewed code change.
+# The cost is made LOUD, never silent: an unlisted field RAISES at construction
+# (D-16). A silently-dropped criterion is exactly how a patient gets told a number
+# applies to him — the one outcome this module exists to prevent.
+CRITERION_FIELDS = (
+    "age", "sex", "bmi", "weight", "baseline_condition", "cv_risk",
+    "triglycerides", "ldl", "hba1c", "egfr", "testosterone", "smoking_status",
+)
+
+# Unit spellings a source (or a patient profile) may append to a field name.
+# `triglycerides_mg_dl_min` and `triglycerides` are the same AXIS; the unit is not
+# a new axis. Stripping is exact-suffix only — no fuzzy matching, so an unknown
+# spelling still reaches D-16's refusal rather than being guessed at.
+UNIT_SUFFIXES = ("mg_dl", "mmol_l", "ng_dl", "nmol_l", "pct", "percent", "kg", "years", "ml_min")
+
+# Comparator spellings a flat criteria dict may use as a key suffix.
+_KEY_OP_SUFFIXES = (("_min", ">="), ("_max", "<="), ("_gt", ">"), ("_lt", "<"),
+                    ("_in", "in"), ("_range", "range"))
+
+_ORDERED_OPS = (">=", "<=", ">", "<")
+
+
+class PopulationError(ValueError):
+    """Every refusal in this module. A subclass of ValueError so callers that
+    already handle malformed populations keep working."""
+
+
+def normalize_field(name: str) -> str:
+    """Map a wire key onto a CRITERION_FIELDS axis, or RAISE (D-16).
+
+    `bmi_min` → `bmi`; `triglycerides_mg_dl_min` → `triglycerides`.
+    An unlisted axis is never dropped, never coerced, never ignored.
+    """
+    if not isinstance(name, str) or not name.strip():
+        raise PopulationError("criterion key must be a non-empty string")
+    key = name.strip().lower()
+    for suffix, _op in _KEY_OP_SUFFIXES:
+        if key.endswith(suffix) and len(key) > len(suffix):
+            key = key[: -len(suffix)]
+            break
+    if key in CRITERION_FIELDS:
+        return key
+    for unit in UNIT_SUFFIXES:
+        tail = "_" + unit
+        if key.endswith(tail) and key[: -len(tail)] in CRITERION_FIELDS:
+            return key[: -len(tail)]
+    raise PopulationError(
+        f"criterion field {name!r} normalises to {key!r}, which is not in CRITERION_FIELDS "
+        f"{CRITERION_FIELDS}. Growing the vocabulary is a reviewed code change; a criterion "
+        f"the matcher cannot name must NOT be silently dropped (D-16)."
+    )
+
+
+def _op_for_key(name: str) -> str:
+    key = name.strip().lower()
+    for suffix, op in _KEY_OP_SUFFIXES:
+        if key.endswith(suffix) and len(key) > len(suffix):
+            return op
+    return "=="
+
+
+@dataclass(frozen=True)
+class Criterion:
+    """One comparable statement drawn from the source's own population description.
+
+    `verbatim` is MANDATORY and non-blank: a criterion nobody transcribed is a
+    criterion nobody can check. `unstated_reason` marks an axis the paper does not
+    state — the honest third state, never a guessed value.
+    """
+
+    field: str
+    op: str
+    value: Any
+    kind: str
+    verbatim: str
+    locator: Optional[str] = None
+    unstated_reason: Optional[str] = None
+
+    def __post_init__(self) -> None:
+        object.__setattr__(self, "field", normalize_field(self.field))
+        if self.kind not in CRITERION_KINDS:
+            raise PopulationError(
+                f"criterion kind {self.kind!r} must be one of {CRITERION_KINDS}. The kind decides "
+                f"whether a mismatch is an EXCLUSION (verdict 'none') or an out-of-range BASELINE "
+                f"(verdict 'partial'); defaulting it would silently pick the verdict."
+            )
+        if not isinstance(self.verbatim, str) or not self.verbatim.strip():
+            raise PopulationError(
+                f"criterion {self.field!r} needs a non-empty `verbatim` — the source's own words, "
+                f"so a reader can check and overrule the machine"
+            )
+        if self.unstated_reason is not None:
+            if not str(self.unstated_reason).strip():
+                raise PopulationError(
+                    f"criterion {self.field!r}: unstated_reason must say WHY the axis is unstated"
+                )
+            return  # an unstated axis carries no op/value to validate
+        if self.op not in CRITERION_OPS:
+            raise PopulationError(f"criterion op {self.op!r} must be one of {CRITERION_OPS}")
+        if self.value is None:
+            raise PopulationError(
+                f"criterion {self.field!r} has no value and no unstated_reason — one of the two is "
+                f"required; an absent value is not the same as an axis the paper never stated"
+            )
+        if self.op == "in" and not isinstance(self.value, (list, tuple)):
+            raise PopulationError(f"criterion {self.field!r} with op 'in' needs a list of values")
+        if self.op == "range" and (not isinstance(self.value, (list, tuple)) or len(self.value) != 2):
+            raise PopulationError(f"criterion {self.field!r} with op 'range' needs [low, high]")
+
+    def requirement_text(self) -> str:
+        if self.unstated_reason is not None:
+            return f"{self.field}: not stated by the source ({self.unstated_reason})"
+        if self.op == "in":
+            return f"{self.field} in {list(self.value)}"
+        if self.op == "range":
+            return f"{self.field} between {self.value[0]} and {self.value[1]}"
+        return f"{self.field} {self.op} {self.value}"
+
+
+@dataclass(frozen=True)
+class StudyPopulation:
+    """Who the finding was obtained IN, as the source states it.
+
+    Valid only when `description` is non-blank AND (`criteria` non-empty OR
+    `unstated_reason` non-empty). `StudyPopulation(description="adults", criteria=())`
+    is NOT a shape — that sentinel is what stops `study_population` from becoming
+    the next field that is present and meaningless.
+    """
+
+    description: str
+    criteria: Tuple[Criterion, ...] = ()
+    unstated_reason: Optional[str] = None
+    locator: Optional[str] = None
+
+    def __post_init__(self) -> None:
+        if not isinstance(self.description, str) or not self.description.strip():
+            raise PopulationError("StudyPopulation.description must be a non-empty string")
+        object.__setattr__(self, "criteria", tuple(self.criteria or ()))
+        for criterion in self.criteria:
+            if not isinstance(criterion, Criterion):
+                raise PopulationError("StudyPopulation.criteria must contain Criterion objects")
+        blank_reason = self.unstated_reason is None or not str(self.unstated_reason).strip()
+        if not self.criteria and blank_reason:
+            raise PopulationError(
+                "StudyPopulation needs at least one criterion, or an explicit unstated_reason. "
+                "An empty criteria list with no stated reason is indistinguishable from a bug "
+                "(the same discipline as create_listing_fact(reason=…))."
+            )
+
+    @classmethod
+    def unstated(cls, reason: str, description: str = "study population not stated by the source",
+                 locator: Optional[str] = None) -> "StudyPopulation":
+        """The single constructor for 'the paper does not say'. `reason` is MANDATORY
+        and stored verbatim — mirroring `create_listing_fact(reason=…)`."""
+        if not isinstance(reason, str) or not reason.strip():
+            raise PopulationError("StudyPopulation.unstated() requires a non-empty reason")
+        return cls(description=description, criteria=(), unstated_reason=reason.strip(), locator=locator)
+
+    def to_dict(self) -> Dict[str, Any]:
+        """The opaque JSON value that crosses into `ed25519_verifier` (never a type)."""
+        out: Dict[str, Any] = {"description": self.description}
+        if self.locator:
+            out["locator"] = self.locator
+        if self.unstated_reason:
+            out["unstated_reason"] = self.unstated_reason
+        criteria: Dict[str, Any] = {}
+        for criterion in self.criteria:
+            spec: Dict[str, Any] = {"kind": criterion.kind, "verbatim": criterion.verbatim}
+            if criterion.unstated_reason is not None:
+                spec["unstated_reason"] = criterion.unstated_reason
+            else:
+                spec["op"] = criterion.op
+                spec["value"] = list(criterion.value) if isinstance(criterion.value, tuple) else criterion.value
+            if criterion.locator:
+                spec["locator"] = criterion.locator
+            criteria[criterion.field] = spec
+        if criteria:
+            out["criteria"] = criteria
+        return out
+
+
+@dataclass(frozen=True)
+class PatientProfile:
+    """The one person the number is being applied to. A query key, not a record."""
+
+    values: Mapping[str, Any] = dc_field(default_factory=dict)
+
+    def __post_init__(self) -> None:
+        if not isinstance(self.values, Mapping):
+            raise PopulationError("PatientProfile.values must be a mapping")
+        normalized: Dict[str, Any] = {}
+        for key, value in self.values.items():
+            # ASYMMETRY, ON PURPOSE. A *criterion* naming an unlisted axis RAISES
+            # (D-16) — it is a claim the matcher cannot check. A *patient* value
+            # naming one is kept under its raw key instead: refusing a whole profile
+            # because it carries one extra lab would be a loud failure in a
+            # patient-facing flow, and the unsafe direction is closed anyway — an
+            # unrecognised patient key can never satisfy a criterion, so the axis
+            # reads `patient-value-missing` → verdict `unknown`, never `full`.
+            try:
+                name = normalize_field(key)
+            except PopulationError:
+                name = str(key).strip().lower()
+            if name in normalized:
+                raise PopulationError(
+                    f"patient profile has two keys that mean the same axis {name!r} — refusing "
+                    f"rather than silently keeping one of them"
+                )
+            normalized[name] = value
+        object.__setattr__(self, "values", normalized)
+
+    def has(self, field: str) -> bool:
+        return field in self.values and self.values[field] is not None
+
+    def get(self, field: str) -> Any:
+        return self.values.get(field)
+
+
+@dataclass(frozen=True)
+class Discrepancy:
+    """A NAMED difference. `«не просто вердикт, а перечисление того, чем именно
+    пациент отличается»` — enforced as a construction precondition, not a rendering habit."""
+
+    field: str
+    kind: str
+    patient_value: str
+    study_requirement: str
+    verbatim: str
+    locator: Optional[str] = None
+    direction: str = "outside-set"
+
+    def __post_init__(self) -> None:
+        if self.kind not in DISCREPANCY_KINDS:
+            raise PopulationError(f"discrepancy kind {self.kind!r} must be one of {DISCREPANCY_KINDS}")
+        if self.direction not in DISCREPANCY_DIRECTIONS:
+            raise PopulationError(
+                f"discrepancy direction {self.direction!r} must be one of {DISCREPANCY_DIRECTIONS}"
+            )
+
+
+def derive_verdict(discrepancies: Sequence[Discrepancy]) -> str:
+    """THE ONLY producer of a verdict (D-9), imported by the runtime and the tests alike.
+
+    Load-bearing invariant (D-10): `verdict == "full"` IFF `discrepancies == []`.
+    `unknown` outranks `partial` but not `none` — an established exclusion is
+    knowledge; an unestablished criterion is not, and must never be reported as a
+    milder kind of match.
+    """
+    kinds = {d.kind for d in discrepancies}
+    if "eligibility-excluded" in kinds:
+        return "none"
+    if kinds & {"criterion-unstated", "patient-value-missing"}:
+        return "unknown"
+    if discrepancies:
+        return "partial"
+    return "full"
+
+
+@dataclass(frozen=True)
+class PopulationMatch:
+    """Derived, never asserted. There is no settable verdict."""
+
+    verdict: str
+    discrepancies: Tuple[Discrepancy, ...] = ()
+    study_description: str = ""
+
+    def __post_init__(self) -> None:
+        object.__setattr__(self, "discrepancies", tuple(self.discrepancies or ()))
+        expected = derive_verdict(self.discrepancies)
+        if self.verdict != expected:
+            raise PopulationError(
+                f"PopulationMatch verdict {self.verdict!r} was not derived from its discrepancies "
+                f"(derive_verdict says {expected!r}). The verdict has exactly one home (D-9)."
+            )
+        if self.verdict not in POPULATION_MATCH_VERDICTS:
+            raise PopulationError(f"verdict {self.verdict!r} must be one of {POPULATION_MATCH_VERDICTS}")
+
+    @classmethod
+    def from_discrepancies(cls, discrepancies: Sequence[Discrepancy],
+                           study_description: str = "") -> "PopulationMatch":
+        return cls(verdict=derive_verdict(discrepancies), discrepancies=tuple(discrepancies),
+                   study_description=study_description)
+
+
+# --------------------------------------------------------------------- parsing
+def parse_criterion(key: str, spec: Any) -> Criterion:
+    """Parse ONE entry of the flat `criteria` dict.
+
+    WIRE FORM — deliberate deviation from FR-2's bare-scalar example, recorded here
+    because it is a contract: the value is a SPEC OBJECT, not a scalar. A scalar
+    cannot state the criterion's `kind`, and the kind is what decides 'none' (you
+    would have been excluded) versus 'partial' (the effect was simply not measured
+    from where you stand). Defaulting it would let the wire format silently pick the
+    verdict — the exact class of failure this slice exists to prevent.
+
+        {"bmi_min": {"value": 30, "kind": "baseline",
+                     "verbatim": "men with obesity (BMI >= 30 kg/m2)",
+                     "locator": "[Methods, Participants]"}}
+    """
+    if not isinstance(spec, Mapping):
+        raise PopulationError(
+            f"criterion {key!r} must be an object carrying at least `kind` and `verbatim` "
+            f"(got {type(spec).__name__}). A bare scalar cannot state the criterion kind, and "
+            f"the kind decides the verdict — see parse_criterion's docstring."
+        )
+    unknown_keys = set(spec) - {"value", "op", "kind", "verbatim", "locator", "unstated_reason"}
+    if unknown_keys:
+        raise PopulationError(f"criterion {key!r} carries unknown keys {sorted(unknown_keys)}")
+    return Criterion(
+        field=key,
+        op=str(spec.get("op") or _op_for_key(key)),
+        value=spec.get("value"),
+        kind=str(spec.get("kind", "")),
+        verbatim=str(spec.get("verbatim", "")),
+        locator=spec.get("locator"),
+        unstated_reason=spec.get("unstated_reason"),
+    )
+
+
+def parse_study_population(data: Any) -> StudyPopulation:
+    """Parse the opaque JSON dict that `ed25519_verifier` carries but never interprets.
+
+    The direction of the dependency matters: the verifier NEVER imports this module
+    (a crypto module whose correctness depends on a semantics module is the
+    fail-open shape `source_tier_ceiling` was already bitten by). It validates the
+    SHAPE; this function is the only place that reads the MEANING.
+    """
+    if not isinstance(data, Mapping):
+        raise PopulationError("study_population must be a JSON object")
+    criteria_raw = data.get("criteria") or {}
+    if not isinstance(criteria_raw, Mapping):
+        raise PopulationError("study_population.criteria must be an object keyed by field name")
+    criteria = tuple(parse_criterion(key, spec) for key, spec in criteria_raw.items())
+    return StudyPopulation(
+        description=str(data.get("description", "")),
+        criteria=criteria,
+        unstated_reason=data.get("unstated_reason"),
+        locator=data.get("locator"),
+    )
+
+
+# ------------------------------------------------------------------ evaluation
+def _as_number(field: str, who: str, value: Any) -> float:
+    try:
+        return float(value)
+    except (TypeError, ValueError):
+        raise PopulationError(
+            f"{who} value {value!r} for {field!r} is not comparable as a number, but the criterion "
+            f"uses an ordered operator. Refusing loudly rather than guessing which side it falls on."
+        ) from None
+
+
+def _satisfied(criterion: Criterion, patient_value: Any) -> Tuple[bool, str]:
+    """(satisfied, direction-when-violated). One generic evaluator, no per-case branch (FR-5)."""
+    op = criterion.op
+    if op in _ORDERED_OPS:
+        patient = _as_number(criterion.field, "patient", patient_value)
+        required = _as_number(criterion.field, "study", criterion.value)
+        if op == ">=":
+            return patient >= required, "below"
+        if op == "<=":
+            return patient <= required, "above"
+        if op == ">":
+            return patient > required, "below"
+        return patient < required, "above"
+    if op == "range":
+        patient = _as_number(criterion.field, "patient", patient_value)
+        low = _as_number(criterion.field, "study", criterion.value[0])
+        high = _as_number(criterion.field, "study", criterion.value[1])
+        if patient < low:
+            return False, "below"
+        if patient > high:
+            return False, "above"
+        return True, "outside-set"
+    if op == "in":
+        allowed = [str(v).strip().lower() for v in criterion.value]
+        return str(patient_value).strip().lower() in allowed, "outside-set"
+    return str(patient_value).strip().lower() == str(criterion.value).strip().lower(), "outside-set"
+
+
+def evaluate(study_population: StudyPopulation, patient: PatientProfile) -> List[Discrepancy]:
+    """ONE generic evaluator over (op, value) pairs — the four real reversals and the
+    synthetic cases all flow through here with different DATA (FR-5)."""
+    if not isinstance(study_population, StudyPopulation):
+        raise PopulationError("evaluate() needs a StudyPopulation (use parse_study_population first)")
+    if not isinstance(patient, PatientProfile):
+        raise PopulationError("evaluate() needs a PatientProfile")
+
+    discrepancies: List[Discrepancy] = []
+
+    if study_population.unstated_reason and not study_population.criteria:
+        return [
+            Discrepancy(
+                field=UNSTATED_POPULATION_FIELD,
+                kind="criterion-unstated",
+                patient_value="n/a",
+                study_requirement=f"not stated: {study_population.unstated_reason}",
+                verbatim=study_population.description,
+                locator=study_population.locator,
+                direction="absent",
+            )
+        ]
+
+    for criterion in study_population.criteria:
+        if criterion.unstated_reason is not None:
+            discrepancies.append(
+                Discrepancy(
+                    field=criterion.field,
+                    kind="criterion-unstated",
+                    patient_value=_render_value(patient.get(criterion.field)),
+                    study_requirement=criterion.requirement_text(),
+                    verbatim=criterion.verbatim,
+                    locator=criterion.locator,
+                    direction="absent",
+                )
+            )
+            continue
+        if not patient.has(criterion.field):
+            discrepancies.append(
+                Discrepancy(
+                    field=criterion.field,
+                    kind="patient-value-missing",
+                    patient_value="не указан / not recorded",
+                    study_requirement=criterion.requirement_text(),
+                    verbatim=criterion.verbatim,
+                    locator=criterion.locator,
+                    direction="absent",
+                )
+            )
+            continue
+        patient_value = patient.get(criterion.field)
+        ok, direction = _satisfied(criterion, patient_value)
+        if ok:
+            continue
+        discrepancies.append(
+            Discrepancy(
+                field=criterion.field,
+                kind="eligibility-excluded" if criterion.kind == "eligibility" else "baseline-out-of-range",
+                patient_value=_render_value(patient_value),
+                study_requirement=criterion.requirement_text(),
+                verbatim=criterion.verbatim,
+                locator=criterion.locator,
+                direction=direction,
+            )
+        )
+    return discrepancies
+
+
+def _render_value(value: Any) -> str:
+    if value is None:
+        return "не указан / not recorded"
+    return str(value)
+
+
+def match(study_population: StudyPopulation, patient: PatientProfile) -> PopulationMatch:
+    """The one entry point a caller needs: evaluate, then derive."""
+    discrepancies = evaluate(study_population, patient)
+    return PopulationMatch.from_discrepancies(discrepancies, study_population.description)
+
+
+# ------------------------------------------------------------------- rendering
+_KIND_PROSE = {
+    "eligibility-excluded": "this patient would NOT have been enrolled",
+    "baseline-out-of-range": "enrollable, but the effect was not measured from this starting value",
+    "criterion-unstated": "the source does not state this axis",
+    "patient-value-missing": "the patient profile does not carry this axis",
+}
+
+
+def render_population_match(match_result: PopulationMatch) -> str:
+    """THE ONLY renderer. Iterates DISCREPANCY_KEYS — an allowlist over emitted
+    structure — so a field added to Discrepancy without being added to the tuple is
+    simply not printed. There is no path that prints a verdict without its
+    discrepancies (D-10)."""
+    if not isinstance(match_result, PopulationMatch):
+        raise PopulationError("render_population_match() needs a PopulationMatch")
+    lines = [f"POPULATION_MATCH: {match_result.verdict}"]
+    if match_result.study_description:
+        lines.append(f"  study population: {match_result.study_description}")
+    if not match_result.discrepancies:
+        lines.append("  no discrepancies — every stated criterion is satisfied by a known patient value")
+        return "\n".join(lines)
+    for discrepancy in match_result.discrepancies:
+        row = {key: getattr(discrepancy, key) for key in DISCREPANCY_KEYS}
+        lines.append(
+            "  {field} — patient {patient_value}; study requires {study_requirement} "
+            "({kind}, {direction})".format(**row)
+        )
+        lines.append("      {prose}".format(prose=_KIND_PROSE[row["kind"]]))
+        quote = '      "{verbatim}"'.format(verbatim=row["verbatim"])
+        if row["locator"]:
+            quote += "  {locator}".format(locator=row["locator"])
+        lines.append(quote)
+    return "\n".join(lines)
+
+
+def match_from_fact(fact_study_population: Any, patient_values: Mapping[str, Any]) -> PopulationMatch:
+    """Convenience for the report gate: opaque dict + raw patient dict → verdict."""
+    return match(parse_study_population(fact_study_population), PatientProfile(patient_values))
+
+
+def load_field_cases(path: str) -> Dict[str, Any]:
+    """Read the committed fixture file. Kept here so tests and the README example
+    load the SAME artifact rather than each restating the four cases."""
+    with open(path, "r", encoding="utf-8") as handle:
+        return json.load(handle)
+
+
+__all__ = [
+    "CRITERION_FIELDS", "CRITERION_KINDS", "CRITERION_OPS", "DISCREPANCY_KINDS",
+    "DISCREPANCY_KEYS", "DISCREPANCY_DIRECTIONS", "POPULATION_MATCH_VERDICTS", "UNIT_SUFFIXES",
+    "UNSTATED_POPULATION_FIELD",
+    "PopulationError", "Criterion", "StudyPopulation", "PatientProfile", "Discrepancy",
+    "PopulationMatch", "derive_verdict", "evaluate", "match", "match_from_fact",
+    "normalize_field", "parse_criterion", "parse_study_population", "render_population_match",
+    "load_field_cases",
+]