test(llm): live CPT-run guard checks against a plain frozenset (3.8 min to 2 s) (S1)
Some checks failed
CI / lint (push) Successful in 57s
CI / notebooks-smoke (push) Successful in 2m19s
Deploy / changes (push) Successful in 17s
Infra CI / notebooks (push) Successful in 1m35s
CI / test (push) Failing after 5m17s
Infra CI / zotero (push) Successful in 32s
Infra CI / docs (push) Successful in 1m35s
Infra CI / api (push) Successful in 1m30s
Infra CI / llm (push) Successful in 1m25s
Infra CI / bcda (push) Successful in 2m16s
Deploy / notebooks (push) Has been skipped
Deploy / zotero (push) Has been skipped
Infra CI / mc (push) Successful in 49s
Deploy / docs (push) Has been skipped
Deploy / api (push) Successful in 2m1s
Deploy / bcda (push) Successful in 2m33s
Deploy / mc (push) Has been skipped
Deploy / llm (push) Successful in 1m45s
Deploy / report (push) Successful in 16s

This commit is contained in:
kert
2026-11-13 03:07:31 -05:00
parent 749e033c61
commit 9ca93c65a6

View File

@@ -105,23 +105,28 @@ def ama_index():
tail and elements, plus section guidelines. Built once per session."""
from market.edgar import AmaIndex
# SELECT DISTINCT: the runs are the same, far fewer rows to tokenize
con = duckdb.connect(str(REPLICA), read_only=True)
try:
texts = [
t
for (t,) in con.execute(
"SELECT descriptor FROM pfs.cpt_code UNION ALL "
"SELECT DISTINCT t FROM ("
"SELECT descriptor AS t FROM pfs.cpt_code UNION ALL "
"SELECT stem FROM pfs.cpt_code UNION ALL "
"SELECT tail FROM pfs.cpt_code UNION ALL "
"SELECT unnest(elements) FROM pfs.cpt_code UNION ALL "
"SELECT guideline FROM pfs.cpt_section"
"SELECT guideline FROM pfs.cpt_section)"
).fetchall()
if t
]
finally:
con.close()
assert texts, "pfs.cpt_code is empty; nothing to guard against"
return AmaIndex.of(texts)
# AmaIndex.of builds the runs; a plain frozenset copy is what the checks
# intersect with, because set.isdisjoint on the frozenset *subclass* takes
# a slow generic path (about 60 ms per call against 181k runs).
return frozenset(AmaIndex.of(texts))
def _strings(x):
@@ -149,13 +154,19 @@ def _prose_segments(text):
def test_live_no_eight_word_cpt_runs(cfg, slices, ama_index):
"""No served string shares an 8-word run with licensed CPT text."""
from llm import policy
from market.edgar import shares_ama_run
from market.edgar import _WORD
def shares_run(text, n=8): # same tokenization as market.edgar.shares_ama_run
w = _WORD.findall(text.lower())
return not {tuple(w[i : i + n]) for i in range(len(w) - n + 1)}.isdisjoint(
ama_index
)
served = {**slices, "contents": policy.load_contents(cfg)}
hits = []
for name, body in served.items():
for text in _strings(body):
if any(shares_ama_run(seg, ama_index) for seg in _prose_segments(text)):
if any(shares_run(seg) for seg in _prose_segments(text)):
hits.append(name)
break
assert not hits, f"8-word CPT run shared in slices {hits}"