Files
stack/tests/pfs/test_citecheck.py
kert 481cd3f2e5 feat(pfs): dossier assembly — cited facts, computed gaps, model prose behind a sentence-level citation check
pfs.cite (Cite + builders; rule_label moved from llm.lineage and re-exported),
pfs.citecheck (sentence split, number normalization, check_prose; plus
grounding_problems: words, codes and CMS-vs-ours attribution of model prose),
pfs.dossier (facts with resolved cites from every Task 1-12 table, timeline,
stakeholder panels with names from pfs.commenter, stance accuracy beside every
count, computed coverage gaps, ledgered model prose with template fallback,
pfs.dossier table), llm.search item_keys filter, `stack pfs dossier`.
Lineage rows reach the dossier only when a final rule's own plain sentence
states them. HOME-VISITS live: 558 facts, 28 timeline events, check passed,
prose model:qwen2.5:14b+template (8 of 29 units on templates).
2026-11-12 23:07:03 -05:00

215 lines
6.3 KiB
Python
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
"""pfs.citecheck — the sentence-level citation check (dossier task 13)."""
from __future__ import annotations
import pytest
from pfs.cite import Cite
from pfs.citecheck import (
Problem,
check_prose,
grounding_problems,
number_tokens,
split_sentences,
)
from pfs.dossier import Fact
def fact(text, values=None, fid="F1"):
return Fact(
id=fid,
kind="round",
year=2023,
codes=(),
text=text,
values=values or {},
cite=Cite(kind="fr_para", label="x", url="u"),
)
F = {
"F1": fact(
"For CY2023, the RUC recommended a work RVU of 1.00 for 99341 and CMS "
"finalized 1.00 (accept).",
{"final": 1.0},
),
"F2": fact(
"For CY2023, the RUC recommended a work RVU of 1.65 for 99342 and CMS "
"finalized 1.65 (accept).",
{"final": 1.65},
"F2",
),
"F3": fact(
"In CY2023 the payment for 99350 was $131.17 and the total changed −16.5%.",
{},
"F3",
),
}
def test_uncited_sentence_rejected():
assert check_prose("CMS accepted the RUC values.", F)[0].kind == "uncited"
def test_number_from_uncited_fact_rejected():
[p] = check_prose("99341 was finalized at 1.65 [F1].", F)
assert p.kind == "number_mismatch"
assert "1.65" in p.detail
def test_equivalent_number_formats_accepted():
assert check_prose("In CY2023 99341 was set at 1.0 work RVUs [F1].", F) == []
def test_unknown_fact_rejected():
[p] = check_prose("CMS finalized 1.00 for 99341 [F9].", F)
assert p.kind == "unknown_fact"
assert "F9" in p.detail
def test_abbreviations_do_not_split():
text = (
"The U.S. statute, e.g. section 1833, i.e. the Act, No. 5 and p. 49 "
"set 1.00 for 99341 [F1]."
)
assert split_sentences(text) == [text]
assert len(split_sentences("A one [F1]. B two [F2].")) == 2
def test_tags_after_the_period_stay_with_their_sentence():
assert split_sentences("CMS finalized 1.00 for 99341. [F1] Next [F2].") == [
"CMS finalized 1.00 for 99341. [F1]",
"Next [F2].",
]
assert check_prose("CMS finalized 1.00 for 99341 in CY2023. [F1]", F) == []
def test_money_and_percent_normalized():
assert number_tokens("$131.17") == {"131.17"}
assert number_tokens("−16.5%") == number_tokens("-16.5%") == {"-16.5%"}
assert number_tokens("–16.5%") == {"-16.5%"}
assert check_prose("99350 paid 131.17 in CY2023 and moved -16.5% [F3].", F) == []
def test_number_tokens_normalization():
assert number_tokens("1.00 3.60 +2 1,234 25th 2.0") == {
"1",
"3.6",
"2",
"1234",
"25th",
}
# bracketed segments (fact tags, labels) carry no numbers
assert number_tokens("x [F12] [87 FR 69608 ¶1672] y") == set()
# a range is two codes, never a negative number
assert number_tokens("99324–99328") == {"99324", "99328"}
# CY years and docket ids
assert number_tokens("CY2023") == {"2023"}
assert number_tokens("CMS-2022-0113") == {"2022", "0113"}
# HCPCS letters codes are whole tokens; modifiers carry no number
assert number_tokens("G0318 and B1") == {"G0318"}
def test_values_count_as_supported_numbers():
facts = {"F1": fact("CMS finalized the value for 99341 (accept).", {"final": 1.0})}
assert check_prose("99341 was finalized at 1.00 [F1].", facts) == []
facts = {"F1": fact("x", {"share": "12.5%", "label": "n/a"})}
assert check_prose("It was 12.5% [F1].", facts) == []
def test_several_tags_in_one_bracket():
assert check_prose("99341 and 99342 got 1.00 and 1.65 [F1, F2].", F) == []
def test_problem_carries_the_sentence():
p = check_prose("Uncited one. Fine 99341 1.00 [F1].", F)
assert p == [Problem("Uncited one.", "uncited", "no [F#] tag")]
@pytest.mark.parametrize("text", ["", " "])
def test_empty_prose_has_no_problems(text):
assert check_prose(text, F) == []
def test_zero_and_empty_values():
assert number_tokens("-0.00 and +0") == {"0"}
facts = {"F1": fact("x", {"none": None, "flag": True, "n": 0})}
assert check_prose("It was 0 [F1].", facts) == []
def _kind(text, kind="round", fid="F1", codes=()):
return Fact(fid, kind, 2023, codes, text, {}, Cite("fr_para", "x", "u"))
G = {
"F1": _kind(
"For CY2008, the RUC recommended a work RVU of 2.27 for 99326 and CMS finalized 2.63 (refine).",
codes=("99326",),
),
"F2": _kind(
"For CY2008, CMS finalized a work RVU of 3.28 for 99350.",
fid="F2",
codes=("99350",),
),
"F3": _kind(
"CMS reported that commenters supported the proposal (CY2005 final rule).",
"cms_summary",
"F3",
),
"F4": _kind(
"Of 3 CY2023 commenters we classified on the work RVU valuation, 1 supported.",
"stance_split",
"F4",
),
}
def test_grounded_restatement_passes():
assert (
grounding_problems(
"For CY2008, CMS finalized work RVUs for 99326 and 99350 [F1][F2].", G
)
== []
)
assert grounding_problems("No tag here.", G) == [] # left to check_prose
def test_invented_words_are_ungrounded():
[p] = grounding_problems(
"The CPT 2019 codebook introduced new codes for office visits [F1].", G
)
assert p.kind == "ungrounded" and "introduc" in p.detail
def test_a_range_spanning_uncited_codes_is_rejected():
[p] = grounding_problems(
"CMS finalized work RVUs for 99326 through 99350 [F1][F2].", G
)
assert p.kind == "code_mismatch" and "99341" in p.detail
def test_cms_summary_stays_attributed_and_apart():
[p] = grounding_problems(
"Commenters supported the proposal (CY2005 final rule) [F3].", G
)
assert p.kind == "unattributed"
assert (
grounding_problems(
"CMS reported that commenters supported the proposal [F3].", G
)
== []
)
[m] = grounding_problems(
"CMS reported that commenters supported; we classified 3 [F3][F4].", G
)
assert m.kind == "merged"
[u] = grounding_problems(
"Of 3 CY2023 commenters, 1 supported the work RVU valuation [F4].", G
)
assert u.kind == "unattributed" and "classified" in u.detail
def test_a_range_the_fact_writes_may_be_restated():
facts = {"F1": _kind("The Panel deleted codes 99324-99328.", "formation")}
assert grounding_problems("The Panel deleted codes 99324-99328 [F1].", facts) == []