pdfhell 0.1.2__tar.gz → 0.1.3__tar.gz

{pdfhell-0.1.2 → pdfhell-0.1.3}/PKG-INFO

@@ -1,13 +1,14 @@
 Metadata-Version: 2.4
 Name: pdfhell
-Version: 0.1.2
+Version: 0.1.3
 Summary: PDF Hell — adversarial PDFs that break AI document readers. Procedural ground truth, not LLM-as-judge.
 Author: Multivon
 License: Apache-2.0
-Project-URL: Homepage, https://pdfhell.multivon.ai
+Project-URL: Homepage, https://multivon.ai/pdfhell
 Project-URL: Repository, https://github.com/multivon-ai/pdfhell
 Project-URL: Issues, https://github.com/multivon-ai/pdfhell/issues
-Project-URL: Leaderboard, https://pdfhell.multivon.ai/leaderboard
+Project-URL: Leaderboard, https://multivon.ai/leaderboard
+Project-URL: Documentation, https://docs.multivon.ai/pdfhell/quickstart
 Keywords: llm,evaluation,pdf,multimodal,benchmark,adversarial,document-ai,rag
 Classifier: License :: OSI Approved :: Apache Software License
 Classifier: Programming Language :: Python :: 3
@@ -208,7 +209,7 @@ Target full suite: 10 trap families, ~50 cases.
 
 For document-AI teams who need adversarial test cases tailored to *their* templates (claims forms, MSAs, medical records, KYC docs), there's a hosted generator that takes your templates and produces adversarial variants with code-based ground truth — same methodology, your data shape.
 
-Email `hello@multivon.ai` for early access, or see [multivon.ai/pricing](https://multivon.ai/pricing).
+Email `hello@multivon.ai` for early access, or see [multivon.ai/commercial](https://multivon.ai/commercial).
 
 ## Installing
 

{pdfhell-0.1.2 → pdfhell-0.1.3}/README.md

@@ -179,7 +179,7 @@ Target full suite: 10 trap families, ~50 cases.
 
 For document-AI teams who need adversarial test cases tailored to *their* templates (claims forms, MSAs, medical records, KYC docs), there's a hosted generator that takes your templates and produces adversarial variants with code-based ground truth — same methodology, your data shape.
 
-Email `hello@multivon.ai` for early access, or see [multivon.ai/pricing](https://multivon.ai/pricing).
+Email `hello@multivon.ai` for early access, or see [multivon.ai/commercial](https://multivon.ai/commercial).
 
 ## Installing
 

{pdfhell-0.1.2 → pdfhell-0.1.3}/pdfhell/__init__.py

@@ -16,7 +16,7 @@ layer; the runtime, scoring, and reporting come from multivon-eval.
 """
 from __future__ import annotations
 
-__version__ = "0.1.2"
+__version__ = "0.1.3"
 
 from .case import HellCase
 from .generators import (

{pdfhell-0.1.2 → pdfhell-0.1.3}/pdfhell/scorer.py

@@ -56,6 +56,11 @@ def wilson_ci(passes: int, n: int, *, z: float = 1.959963984540054) -> tuple[flo
 
 _WHITESPACE_RE = re.compile(r"\s+")
 _PUNCT_NORMALIZE_RE = re.compile(r"[.,;:]+\s*$")
+# Currency markers — matched immediately before a digit so we don't
+# strip stray $ in unrelated prose. Used by _contains_loose to give the
+# match a second pass when the model omitted the currency prefix the
+# answer key happened to include.
+_LEADING_CURRENCY_RE = re.compile(r"(?<![A-Za-z0-9])[$€£¥₹]\s*(?=\d)")
 
 
 def _normalize(s: str) -> str:
@@ -68,8 +73,31 @@ def _normalize(s: str) -> str:
     return s
 
 
+def _strip_currency(s: str) -> str:
+    """Drop a leading currency symbol that sits right before a digit. So
+    '$780,803.18' → '780,803.18' but 'invoice INV-$X-1' is left alone."""
+    return _LEADING_CURRENCY_RE.sub("", s)
+
+
 def _contains_loose(haystack: str, needle: str) -> bool:
-    return _normalize(needle) in _normalize(haystack)
+    """Tolerant contains-match used as the headline correctness signal.
+
+    First tries the straight normalised contains. If that fails AND the
+    needle starts with a currency symbol, retries with both sides stripped
+    of the leading currency prefix — so an answer key of '$780,803.18'
+    still matches a model output of '780,803.18'. This kept popping up
+    on the split_table_across_pages trap, where models tend to omit the
+    '$' even when the table header includes it.
+    """
+    nh = _normalize(haystack)
+    nn = _normalize(needle)
+    if nn in nh:
+        return True
+    nh_stripped = _strip_currency(nh)
+    nn_stripped = _strip_currency(nn)
+    if nn_stripped != nn and nn_stripped in nh_stripped:
+        return True
+    return False
 
 
 @dataclass(slots=True)

{pdfhell-0.1.2 → pdfhell-0.1.3}/pdfhell.egg-info/PKG-INFO

@@ -1,13 +1,14 @@
 Metadata-Version: 2.4
 Name: pdfhell
-Version: 0.1.2
+Version: 0.1.3
 Summary: PDF Hell — adversarial PDFs that break AI document readers. Procedural ground truth, not LLM-as-judge.
 Author: Multivon
 License: Apache-2.0
-Project-URL: Homepage, https://pdfhell.multivon.ai
+Project-URL: Homepage, https://multivon.ai/pdfhell
 Project-URL: Repository, https://github.com/multivon-ai/pdfhell
 Project-URL: Issues, https://github.com/multivon-ai/pdfhell/issues
-Project-URL: Leaderboard, https://pdfhell.multivon.ai/leaderboard
+Project-URL: Leaderboard, https://multivon.ai/leaderboard
+Project-URL: Documentation, https://docs.multivon.ai/pdfhell/quickstart
 Keywords: llm,evaluation,pdf,multimodal,benchmark,adversarial,document-ai,rag
 Classifier: License :: OSI Approved :: Apache Software License
 Classifier: Programming Language :: Python :: 3
@@ -208,7 +209,7 @@ Target full suite: 10 trap families, ~50 cases.
 
 For document-AI teams who need adversarial test cases tailored to *their* templates (claims forms, MSAs, medical records, KYC docs), there's a hosted generator that takes your templates and produces adversarial variants with code-based ground truth — same methodology, your data shape.
 
-Email `hello@multivon.ai` for early access, or see [multivon.ai/pricing](https://multivon.ai/pricing).
+Email `hello@multivon.ai` for early access, or see [multivon.ai/commercial](https://multivon.ai/commercial).
 
 ## Installing
 

{pdfhell-0.1.2 → pdfhell-0.1.3}/pdfhell.egg-info/SOURCES.txt

@@ -26,4 +26,5 @@ tests/test_cli.py
 tests/test_generators.py
 tests/test_junit.py
 tests/test_scorer.py
+tests/test_scorer_currency.py
 tests/test_statistical.py

{pdfhell-0.1.2 → pdfhell-0.1.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "pdfhell"
-version = "0.1.2"
+version = "0.1.3"
 description = "PDF Hell — adversarial PDFs that break AI document readers. Procedural ground truth, not LLM-as-judge."
 readme = "README.md"
 requires-python = ">=3.10"
@@ -44,10 +44,11 @@ all = []
 pdfhell = "pdfhell.cli:main"
 
 [project.urls]
-Homepage = "https://pdfhell.multivon.ai"
+Homepage = "https://multivon.ai/pdfhell"
 Repository = "https://github.com/multivon-ai/pdfhell"
 Issues = "https://github.com/multivon-ai/pdfhell/issues"
-Leaderboard = "https://pdfhell.multivon.ai/leaderboard"
+Leaderboard = "https://multivon.ai/leaderboard"
+Documentation = "https://docs.multivon.ai/pdfhell/quickstart"
 
 [tool.setuptools.packages.find]
 include = ["pdfhell*"]

pdfhell-0.1.3/tests/test_scorer_currency.py

@@ -0,0 +1,43 @@
+"""Regression test: currency-prefix tolerance in score_case.
+
+Caught from user audit — GPT-4o output '780,803.18' for expected
+'$780,803.18' was previously marked incorrect. With the
+_strip_currency fallback in _contains_loose, both the prefixed and
+unprefixed forms now match (in either direction).
+"""
+from pdfhell.case import HellCase
+from pdfhell.scorer import score_case
+
+
+def _case(expected: str, forbidden=()) -> HellCase:
+    return HellCase(
+        id="t",
+        trap_family="hidden_ocr_mismatch",
+        seed=1,
+        question="q?",
+        expected_answer=expected,
+        forbidden_answers=list(forbidden),
+    )
+
+
+class TestCurrencyTolerance:
+    def test_unprefixed_output_matches_dollar_expected(self):
+        s = score_case(_case("$780,803.18"), "The total is 780,803.18.")
+        assert s.correct
+        assert s.matched_expected
+
+    def test_dollar_output_matches_dollar_expected(self):
+        s = score_case(_case("$780,803.18"), "Total: $780,803.18")
+        assert s.correct
+
+    def test_unprefixed_expected_matches_dollar_output(self):
+        s = score_case(_case("780,803.18"), "The amount is $780,803.18.")
+        assert s.correct
+
+    def test_euro_prefix_tolerated(self):
+        s = score_case(_case("€1,234.56"), "Refund: 1,234.56")
+        assert s.correct
+
+    def test_does_not_match_wrong_number(self):
+        s = score_case(_case("$780,803.18"), "Total: $780,000.")
+        assert not s.correct