benchmark-reliability 0.3.0__tar.gz → 0.3.1__tar.gz

benchmark_reliability-0.3.1/MANIFEST.in

@@ -0,0 +1,13 @@
+# verification/ holds the archived output of
+#   python scripts/compute_prepare_hashes.py --verify
+# under two interpreters, plus the superseded 3.8 run that motivated the
+# canonical content hash. It is evidence for the prepare-level digests rather
+# than package code, so it lives outside src/brf and is included here rather
+# than via package-data.
+include verification/*.txt
+include verification/*.md
+
+# The alternate reference style stays available for reuse of the class file
+# even though the manuscript selects sn-nature.
+include sn-nature.bst
+include sn-mathphys-num.bst

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: benchmark-reliability
-Version: 0.3.0
+Version: 0.3.1
 Summary: Benchmark Reliability Framework (BRF) - dataset-level reliability auditing with built-in benchmark registry
 Author-email: zhanglizhuo <zhanglizhuo@gmail.com>
 License: MIT
@@ -24,6 +24,10 @@ Requires-Dist: matplotlib>=3.5
 Requires-Dist: pandas>=1.1
 Requires-Dist: scipy>=1.5
 Requires-Dist: openml>=0.12
+Provides-Extra: full
+Requires-Dist: kagglehub>=0.1; extra == "full"
+Requires-Dist: openpyxl>=3.1; extra == "full"
+Requires-Dist: pyreadstat>=1.2; extra == "full"
 
 # benchmark-reliability
 

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "benchmark-reliability"
-version = "0.3.0"
+version = "0.3.1"
 description = "Benchmark Reliability Framework (BRF) - dataset-level reliability auditing with built-in benchmark registry"
 readme = "README.md"
 license = { text = "MIT" }
@@ -33,6 +33,21 @@ dependencies = [
     "openml>=0.12",
 ]
 
+# Seven of the 51 datasets need a reader that is not part of the core stack:
+# kagglehub for 5 Kaggle-hosted sources, pyreadstat for the PISA 2015 .sav
+# files, openpyxl for the Taiwan .xlsx files. They are optional rather than
+# required because pyreadstat needs the system-level libreadstat, which would
+# otherwise be imposed on every install for the sake of 44 datasets.
+# pyreadstat also needs libreadstat present on the system, e.g.
+#   Debian/Ubuntu: apt-get install libreadstat-dev
+#   macOS:         brew install readstat
+[project.optional-dependencies]
+full = [
+    "kagglehub>=0.1",
+    "openpyxl>=3.1",
+    "pyreadstat>=1.2",
+]
+
 [project.scripts]
 brf = "brf.cli:main"
 
@@ -48,7 +63,7 @@ where = ["src"]
 include = ["brf", "brf.metrics", "brf.phase", "brf.registry", "brf.registry.sources", "brf.report"]
 
 [tool.setuptools.package-data]
-"brf.registry" = ["manifest.yaml", "registry_v2.1.json"]
+"brf.registry" = ["manifest.yaml", "registry_v2.1.json", "prepare_hashes.json"]
 
 [tool.setuptools.exclude-package-data]
 "brf.registry" = ["cache/*", "cache/**/*"]

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/benchmark_reliability.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: benchmark-reliability
-Version: 0.3.0
+Version: 0.3.1
 Summary: Benchmark Reliability Framework (BRF) - dataset-level reliability auditing with built-in benchmark registry
 Author-email: zhanglizhuo <zhanglizhuo@gmail.com>
 License: MIT
@@ -24,6 +24,10 @@ Requires-Dist: matplotlib>=3.5
 Requires-Dist: pandas>=1.1
 Requires-Dist: scipy>=1.5
 Requires-Dist: openml>=0.12
+Provides-Extra: full
+Requires-Dist: kagglehub>=0.1; extra == "full"
+Requires-Dist: openpyxl>=3.1; extra == "full"
+Requires-Dist: pyreadstat>=1.2; extra == "full"
 
 # benchmark-reliability
 

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/benchmark_reliability.egg-info/SOURCES.txt

@@ -1,3 +1,4 @@
+MANIFEST.in
 README.md
 pyproject.toml
 setup.py
@@ -22,6 +23,7 @@ src/brf/phase/visualization.py
 src/brf/registry/__init__.py
 src/brf/registry/cli.py
 src/brf/registry/manifest.yaml
+src/brf/registry/prepare_hashes.json
 src/brf/registry/registry_v2.1.json
 src/brf/registry/verify.py
 src/brf/registry/sources/__init__.py
@@ -75,4 +77,8 @@ src/brf/report/json_export.py
 src/brf/report/latex_export.py
 tests/test_analyzer.py
 tests/test_metrics.py
-tests/test_phase.py
+tests/test_phase.py
+verification/README.md
+verification/prepare_verify_py313.txt
+verification/prepare_verify_py38.txt
+verification/prepare_verify_py38_legacy.txt

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/benchmark_reliability.egg-info/requires.txt

@@ -4,3 +4,8 @@ matplotlib>=3.5
 pandas>=1.1
 scipy>=1.5
 openml>=0.12
+
+[full]
+kagglehub>=0.1
+openpyxl>=3.1
+pyreadstat>=1.2

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/brf/__init__.py

@@ -1,6 +1,6 @@
 from .analyzer import BRFAnalyzer
 
-__version__ = "0.3.0"
+__version__ = "0.3.1"
 
 __all__ = ["BRFAnalyzer", "__version__"]
 

benchmark_reliability-0.3.1/src/brf/registry/prepare_hashes.json

@@ -0,0 +1,53 @@
+{
+  "abalone": "3f7c6c66fdec85a6b3f2793bdad83bb0620ade38c00b9ac384cd7292b91da970",
+  "air_quality": "4ee091d7c6cd29c22aca75bd181ba12d4fceaea5ee8882931703727485e5d854",
+  "airfoil": "48bcb721e3e8fce370b6ad80ba3ae1735eaee511e1da68e78a48c43b1e9a4701",
+  "assistments": "9bd474c2e85e0778c6b7e669d03976ee6d5832307ade272d94767f85f636fb11",
+  "auto_mpg": "a8b5a265bf2aada37c94fe3385beed198d875732ede5d4d47c888579ff8b76e7",
+  "bike_sharing": "6759530a726ce5bb9712a52317120e273c81a2418adad82723505ecdf20949ec",
+  "boston_housing": "1626353c69b1b1bc79bc74a2fff33b644c433b9565c281a528622d8b20ac64c2",
+  "breast_cancer": "2b89a1181eb63e05dc149950986a466a98b4fa1814e1bb4b1feaa12b40fad485",
+  "california_housing": "e8ab522c9d4aba60bda5eb9ac90291d46725fdbc0368dd5f4d115b807b718e9c",
+  "ccpp": "84f5f83018d7030fa0a088c7f3b3fc6c8535be21b9cffb67c7a92ed0c0718ffd",
+  "climate_weather": "f14ba7b89eda5fea2871b67485fde7188002b9047d63f17d55e6420beca61a8c",
+  "college_scorecard": "2de9c254016f6d939a5bd58fc53f7d31dd4d5dea64dec25d36847bb3ebca5762",
+  "colleges_aaup": "a98f7f115c3fc70742c222180f2f92d56735bb42321f4a213b109482dca0422e",
+  "colleges_usnews": "cae3862477696e189ac044e36029d464bc4800e1f1686f0d6003573c29dc94c5",
+  "concrete": "3c4b71924aa1f5b7f80da0fd170fa741fe9cd671cee93294605988be620f325b",
+  "cpu_act": "12e266a8e018e51f119c188eaeb626aa82aa732f83a2a7ba0909faf5df9eea3d",
+  "credit_card": "035c55b1ffe395caf589eaea48be15c178e68c39464ccd385c268f41ba00706f",
+  "customer_churn": "53e26607e6d3e85f22b1469219cf90cdac066cfcd752f28b832d18a8305a38e5",
+  "diabetes": "e471bce7f98d55bc2f8bf13e93f0c924550f0c630bffbd48558e277d3e22c930",
+  "electricity": "b344eff67643623a8a9692fb4b764e1912c7a8c947d30105ed7046ce62d422ac",
+  "energy_building": "c045e11976d5c01f531afdf3f042b913054661bf5123361cfeaacedb3f709a71",
+  "energy_efficiency": "3d75f02739b5e4f360a29bccc2c19ab0206f96979f7dbf05a19f76fd1455e02f",
+  "entrance_exam": "43bca05374772861b4cf6eefa46f629c4751ab2374f647ef20177ed3ac3313ff",
+  "german_credit": "185e97d0640048b765abc03dbd21c5fba8917a8bb1f295000dbd7ac7e484df26",
+  "higher_ed": "59be6940c05cd59bfc651f7af74dee83e11b506a9f7052f88516806f804a1944",
+  "kaggle_students_performance": "2076cb97e81c74e97af541879064fadf8eb69de9b5640e0d4abde9e1053d5dfa",
+  "kdd_cup_2010": "b7d8e2aef91cbc6a3e7f538c8fa2ab47281ac76c80bbb31c17ec05391557bb9c",
+  "law_school": "b1f33959f2f4a3620c983a46de72027307ecdfa2f391c9b0d6dc526f65b8c14b",
+  "linnerud": "f00cf92d24f6f7fc1d8acc780c8503c5d6346186f152afcb9bdb6c0611fbe502",
+  "mathe": "ac788168060595a034c22849561c28567f54727b713bdcc40c0ed8232440b8e3",
+  "mm_tba": "c530bd64a241ee4012a336dad710cc0b3ada4171817a02e5f7efeeb2a20e0df9",
+  "nursery": "afd19fc4496c390437c76818f2d4e54d8cb46dd019cdd053f46058c6070e49e2",
+  "oli": "a7d2ccf40aaf65ab022c6bfef458461811ee1e66d100a40a3287f01ee9313780",
+  "olympics": "4429c726958ccc094104f1caeefa39a259181a3fbb5bd65d7aba49f55dcfa95d",
+  "oulad": "9aa535d041c9f39091e7e9f5220b6d51bc1f3a42bc27b0086c66b4cf059f61b3",
+  "pisa2015": "1a0e76d0cd8dcf070c41415efc6f0fa6e293539353566182d594ceb4c306e091",
+  "pollution": "5768d5b4509b398f17f90d4da5e46b34b22a5b19ab4ab112feb667be46d7dd07",
+  "real_estate": "1c67624e20a3948cb76612a4c96755b3b2f48d743c3891881ee2b42ebea25bb6",
+  "seoul_bike": "03d8e2b26f08a21aa737c901b5dddef9a11649eb0b596c920b138bfa564a1938",
+  "student_absences": "abf4c2a41c532386182ddc2962c03f86b00dfa4eaceb2c0b2e1450526a381084",
+  "student_depression": "f6f0f65bebafaea796404e79a90d7c8ff2b81ecfacbf0655fd8cd644e9fecb9d",
+  "student_dropout": "c01207db62dc1ac9f098243fda359274447159ebab08ec6dccfda57645fa4ef5",
+  "student_health": "0fa081e63bb3f259fb4c0a7cf74a2cf5a9f136ecf9c221421d7bc1e4ed72330b",
+  "students_exam_scores": "336969cba9ef0bb7152f49385e27b5cf1d27b88978482faaa305db21e3a3a1ca",
+  "tae": "356cb12fda2ff5525ce4967d39a992c1bf2b41d5df98dea056d0a815ebee9a0a",
+  "turkiye": "fef1729ce8edf310ba326531a065f2912abea5884087d817ca5c47dea828d0ff",
+  "uci_student": "7f5a7696641b35866db2ece4c04524f1090408f519abdd490fc34279a6fc55e6",
+  "uci_student_math": "2cbf9c120016bb28c88e2e1ac05da4d40de25490427659e990d0204df2a06354",
+  "wine_quality": "e26239cad8e85992ff400ec40f7cd4447d206512bd335caa7900a22cbc51e424",
+  "xapi_edu": "6470f36422d4ba1e1a4e431bc7fd0c5d589902586d3043d6ce7b16ce9eacfe4e",
+  "yacht": "527653450af99bdbe85ce23f493b2faf9f207bef2b9be26931b0c6bfa8c105ef"
+}

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/brf/registry/sources/__init__.py

@@ -95,15 +95,39 @@ class DatasetSource(ABC):
     def prepare(self):
         """Return (X, y, groups, metadata_card)."""
 
-    def _compute_prepare_sha256(self, X, y, groups):
-        import io
+    def _canonical_array_bytes(arr):
+        """Return (dtype_tag, payload) for an array in an interpreter-stable form."""
+        a = np.asarray(arr)
+        if a.dtype.kind in "OUS":
+            return "str", "\n".join("" if v is None else str(v) for v in a.ravel()).encode("utf-8")
+        kind = {"i": "int64", "u": "uint64", "b": "bool",
+                "f": "float64", "c": "complex128"}.get(a.dtype.kind)
+        if kind is None:
+            raise TypeError(f"unsupported dtype kind {a.dtype.kind!r} for canonical hashing")
+        return kind, np.ascontiguousarray(a, dtype=kind).tobytes()
+
+
+    def compute_prepare_sha256(X, y, groups):
+        """SHA-256 over a canonical encoding of the prepared data.
+
+        This hashes the data, not a serialisation of it. Hashing the bytes that
+        numpy.save() emits is not equivalent: object-dtype arrays are pickled, and
+        that byte stream is not stable across interpreters. Measured on this
+        registry, identical data hashed differently on CPython 3.8 and 3.13 for 26
+        of 51 datasets, because the pickled form of the object-dtype `groups` array
+        differed in length (8663 vs 8646 bytes for abalone) while its values were
+        element-for-element equal.
+
+        Each array is therefore reduced to an explicitly declared dtype, its shape,
+        and its raw bytes, so the digest depends only on the values.
+        """
         sha = hashlib.sha256()
-        for arr in (X, y, groups):
-            buf = io.BytesIO()
-            np.save(buf, np.asarray(arr))
-            sha.update(buf.getvalue())
+        for name, arr in (("X", X), ("y", y), ("groups", groups)):
+            dtype_tag, payload = _canonical_array_bytes(arr)
+            sha.update(f"{name}|{dtype_tag}|{np.asarray(arr).shape}|".encode("utf-8"))
+            sha.update(payload)
+            sha.update(b"\x00")
         return sha.hexdigest()
-
     def verify_prepare(self):
         if not self.prepare_sha256:
             return True

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/brf/registry/sources/kaggle_students_performance.py

@@ -19,7 +19,14 @@ class KaggleStudentsPerformanceSource(DatasetSource):
     grouping_description = "Race/Ethnicity (5 groups: A-E)"
 
     def download(self):
-        import kagglehub
+        try:
+            import kagglehub
+        except ImportError as exc:
+            raise ImportError(
+                "kagglehub is required for this dataset but is not installed. "
+                "Install the optional extra with: "
+                "pip install 'benchmark-reliability[full]'"
+            ) from exc
         import shutil
         dest_dir = self._ensure_cache_dir()
         csv_path = dest_dir / "StudentsPerformance.csv"

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/brf/registry/sources/kdd_cup_2010.py

@@ -26,6 +26,9 @@ class KDDCup2010Source(DatasetSource):
         train_file = dest_dir / "algebra_2005_2006_train.txt"
         if train_file.exists():
             return train_file
+        # A prior download extracts into a subdir; reuse it offline.
+        for cand in dest_dir.rglob("algebra_2005_2006_train.txt"):
+            return cand
 
         urls = [self.source_url] + self.fallback_urls
         for url in urls:

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/brf/registry/sources/mm_tba.py

@@ -49,7 +49,14 @@ class MMTBASource(DatasetSource):
         return dest_dir
 
     def prepare(self):
-        import openpyxl
+        try:
+            import openpyxl
+        except ImportError as exc:
+            raise ImportError(
+                "openpyxl is required for this dataset but is not installed. "
+                "Install the optional extra with: "
+                "pip install 'benchmark-reliability[full]'"
+            ) from exc
 
         root = self.download()
         # Handle figshare zip structure: MM-TBA/MM-TBA/...

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/brf/registry/sources/pisa2015.py

@@ -38,7 +38,14 @@ class PISA2015Source(DatasetSource):
         return sav_path
 
     def prepare(self):
-        import pyreadstat
+        try:
+            import pyreadstat
+        except ImportError as exc:
+            raise ImportError(
+                "pyreadstat is required for this dataset but is not installed. "
+                "Install the optional extra with: "
+                "pip install 'benchmark-reliability[full]'"
+            ) from exc
         pvt_cols = [f"PV{i}SCIE" for i in range(1, 6)]
         usecols = ["CNT", "ESCS", "WEALTH", "HOMEPOS"] + pvt_cols
         df, _ = pyreadstat.read_sav(str(self.download()), usecols=usecols)

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/brf/registry/sources/student_absences.py

@@ -19,7 +19,14 @@ class StudentAbsencesSource(DatasetSource):
     grouping_description = "School (2: GP/MS)"
 
     def download(self):
-        import kagglehub
+        try:
+            import kagglehub
+        except ImportError as exc:
+            raise ImportError(
+                "kagglehub is required for this dataset but is not installed. "
+                "Install the optional extra with: "
+                "pip install 'benchmark-reliability[full]'"
+            ) from exc
         dest_dir = self._ensure_cache_dir()
         csv_path = dest_dir / "student-mat.csv"
         if csv_path.exists():

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/brf/registry/sources/student_dropout.py

@@ -24,6 +24,10 @@ class StudentDropoutSource(DatasetSource):
         csv_path = dest_dir / "student_dropout.csv"
         if csv_path.exists():
             return csv_path
+        # A prior download keeps the extracted csv under its original name
+        # (e.g. data.csv); reuse it offline instead of re-downloading.
+        for f in dest_dir.glob("**/*.csv"):
+            return f
         resp = urllib.request.urlopen(self.source_url, timeout=60)
         with zipfile.ZipFile(io.BytesIO(resp.read())) as z:
             z.extractall(str(dest_dir))

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/brf/registry/sources/student_health.py

@@ -19,7 +19,14 @@ class StudentHealthSource(DatasetSource):
     sha256 = "3fb408fe4b66d4cdb90ad4ec5e3fb6de94ba662d0471a3d46c03da483069f76b"
 
     def download(self):
-        import kagglehub
+        try:
+            import kagglehub
+        except ImportError as exc:
+            raise ImportError(
+                "kagglehub is required for this dataset but is not installed. "
+                "Install the optional extra with: "
+                "pip install 'benchmark-reliability[full]'"
+            ) from exc
         dest_dir = self._ensure_cache_dir()
         csv_path = dest_dir / "student-por.csv"
         if csv_path.exists():

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/brf/registry/sources/students_exam_scores.py

@@ -25,7 +25,14 @@ class StudentsExamScoresSource(DatasetSource):
             return csv_path
         
         
-        import kagglehub
+        try:
+            import kagglehub
+        except ImportError as exc:
+            raise ImportError(
+                "kagglehub is required for this dataset but is not installed. "
+                "Install the optional extra with: "
+                "pip install 'benchmark-reliability[full]'"
+            ) from exc
         path = Path(kagglehub.dataset_download("desalegngeb/students-exam-scores"))
         for f in path.glob("*.csv"):
             if "Original" in f.name:

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/brf/registry/sources/xapi_edu.py

@@ -26,7 +26,14 @@ class XAPIEduSource(DatasetSource):
         
         # Try Kaggle via kagglehub first
         try:
-            import kagglehub
+            try:
+                import kagglehub
+            except ImportError as exc:
+                raise ImportError(
+                    "kagglehub is required for this dataset but is not installed. "
+                    "Install the optional extra with: "
+                    "pip install 'benchmark-reliability[full]'"
+                ) from exc
             path = kagglehub.dataset_download("aljarah/xAPI-Edu-Data")
             import os
             for f in os.listdir(path):

{benchmark_reliability-0.3.0 → benchmark_reliability-0.3.1}/src/brf/registry/verify.py

@@ -12,12 +12,38 @@ from .sources import REGISTRY_SOURCES, list_sources
 _PREPARE_HASHES_PATH = Path(__file__).resolve().parent / "prepare_hashes.json"
 
 
+def _canonical_array_bytes(arr):
+    """Return (dtype_tag, payload) for an array in an interpreter-stable form."""
+    a = np.asarray(arr)
+    if a.dtype.kind in "OUS":
+        return "str", "\n".join("" if v is None else str(v) for v in a.ravel()).encode("utf-8")
+    kind = {"i": "int64", "u": "uint64", "b": "bool",
+            "f": "float64", "c": "complex128"}.get(a.dtype.kind)
+    if kind is None:
+        raise TypeError(f"unsupported dtype kind {a.dtype.kind!r} for canonical hashing")
+    return kind, np.ascontiguousarray(a, dtype=kind).tobytes()
+
+
 def compute_prepare_sha256(X, y, groups):
+    """SHA-256 over a canonical encoding of the prepared data.
+
+    This hashes the data, not a serialisation of it. Hashing the bytes that
+    numpy.save() emits is not equivalent: object-dtype arrays are pickled, and
+    that byte stream is not stable across interpreters. Measured on this
+    registry, identical data hashed differently on CPython 3.8 and 3.13 for 26
+    of 51 datasets, because the pickled form of the object-dtype `groups` array
+    differed in length (8663 vs 8646 bytes for abalone) while its values were
+    element-for-element equal.
+
+    Each array is therefore reduced to an explicitly declared dtype, its shape,
+    and its raw bytes, so the digest depends only on the values.
+    """
     sha = hashlib.sha256()
-    for arr in (X, y, groups):
-        buf = io.BytesIO()
-        np.save(buf, np.asarray(arr))
-        sha.update(buf.getvalue())
+    for name, arr in (("X", X), ("y", y), ("groups", groups)):
+        dtype_tag, payload = _canonical_array_bytes(arr)
+        sha.update(f"{name}|{dtype_tag}|{np.asarray(arr).shape}|".encode("utf-8"))
+        sha.update(payload)
+        sha.update(b"\x00")
     return sha.hexdigest()
 
 

benchmark_reliability-0.3.1/verification/README.md

@@ -0,0 +1,55 @@
+# Archived prepare-verification logs
+
+Raw output of
+
+```
+python scripts/compute_prepare_hashes.py --verify
+```
+
+## Current state (canonical content hash)
+
+| log | interpreter | OK | MISMATCH | ERROR |
+|---|---|---|---|---|
+| `prepare_verify_py313.txt` | CPython 3.13.11 (numpy 2.4.4, pandas 2.3.3) | 51 | 0 | 0 |
+| `prepare_verify_py38.txt` | CPython 3.8.20 (numpy 1.24.3, pandas 2.0.3) | 36 | 0 | 15 |
+
+`MISMATCH` is 0 on both. The 15 errors on 3.8 are missing optional readers in
+that deliberately minimal environment (9 `sklearn`, 3 `kagglehub`, and one each
+of `scipy`, `pyreadstat`, `openpyxl`); 36 + 15 = 51. They are installation
+state, not hash disagreement.
+
+## Why the legacy log exists
+
+`prepare_verify_py38_legacy.txt` is the same command under the previous
+implementation and shows `10 OK / 27 MISMATCH / 15 ERROR`. It is kept because it
+is the evidence for the change that produced the current scheme.
+
+`compute_prepare_sha256` used to hash the bytes `numpy.save()` emits. That is a
+hash of a serialisation, not of the data. `groups` is an `object`-dtype array,
+so `numpy.save` pickles it, and the pickled byte stream is not stable across
+interpreters. For `abalone`, `X` and `y` were element-for-element equal across
+3.8 and 3.13 (`np.array_equal` `True`, max absolute difference `0.0`, same shape,
+same `float64` dtype), yet the digests differed because the serialised `groups`
+array was 8663 bytes on 3.8 and 8646 on 3.13 (`ec91fc24...` vs `98b59954...`).
+The dtype and element type (`str`) matched on both sides and
+`pickle.DEFAULT_PROTOCOL` was 4 on both, so the difference was in the encoding
+rather than in protocol selection.
+
+The function now reduces each array to an explicitly declared dtype, its shape,
+and its raw bytes, so the digest depends on the values only.
+
+## Reproducing the 3.8 run
+
+Python 3.8 is not in the base environment:
+
+```
+conda create -n py38chk python=3.8 -y
+conda activate py38chk
+pip install --no-deps dist/benchmark_reliability-0.3.1-py3-none-any.whl
+conda install numpy            # 1.24.3, last numpy supporting 3.8
+pip install pandas==2.0.3      # last pandas supporting 3.8
+```
+
+scikit-learn, scipy, matplotlib, openml, kagglehub, pyreadstat and openpyxl
+were left uninstalled, which accounts for all 15 errors and affects none of the
+36 verifications.

benchmark_reliability-0.3.1/verification/prepare_verify_py313.txt

@@ -0,0 +1,55 @@
+# brf registry prepare verification — canonical content hash — Python 3.13.11
+# generated 2026-10-01 07:25:34 CST
+# command: python scripts/compute_prepare_hashes.py --verify
+#
+  abalone... OK (3f7c6c66fdec85a6...)
+  air_quality... OK (4ee091d7c6cd29c2...)
+  airfoil... OK (48bcb721e3e8fce3...)
+  assistments... OK (9bd474c2e85e0778...)
+  auto_mpg... OK (a8b5a265bf2aada3...)
+  bike_sharing... OK (6759530a726ce5bb...)
+  boston_housing... OK (1626353c69b1b1bc...)
+  breast_cancer... OK (2b89a1181eb63e05...)
+  california_housing... OK (e8ab522c9d4aba60...)
+  ccpp... OK (84f5f83018d7030f...)
+  climate_weather... OK (f14ba7b89eda5fea...)
+  college_scorecard... OK (2de9c254016f6d93...)
+  colleges_aaup... OK (a98f7f115c3fc707...)
+  colleges_usnews... OK (cae3862477696e18...)
+  concrete... OK (3c4b71924aa1f5b7...)
+  cpu_act... OK (12e266a8e018e51f...)
+  credit_card... OK (035c55b1ffe395ca...)
+  customer_churn... OK (53e26607e6d3e85f...)
+  diabetes... OK (e471bce7f98d55bc...)
+  electricity... OK (b344eff67643623a...)
+  energy_building... OK (c045e11976d5c01f...)
+  energy_efficiency... OK (3d75f02739b5e4f3...)
+  entrance_exam... OK (43bca05374772861...)
+  german_credit... OK (185e97d0640048b7...)
+  higher_ed... OK (59be6940c05cd59b...)
+  kaggle_students_performance... OK (2076cb97e81c74e9...)
+  kdd_cup_2010... OK (b7d8e2aef91cbc6a...)
+  law_school... OK (b1f33959f2f4a362...)
+  linnerud... OK (f00cf92d24f6f7fc...)
+  mathe... OK (ac788168060595a0...)
+  mm_tba... OK (c530bd64a241ee40...)
+  nursery... OK (afd19fc4496c3904...)
+  oli... OK (a7d2ccf40aaf65ab...)
+  olympics... OK (4429c726958ccc09...)
+  oulad... OK (9aa535d041c9f390...)
+  pisa2015... OK (1a0e76d0cd8dcf07...)
+  pollution... OK (5768d5b4509b398f...)
+  real_estate... OK (1c67624e20a3948c...)
+  seoul_bike... OK (03d8e2b26f08a21a...)
+  student_absences... OK (abf4c2a41c532386...)
+  student_depression... OK (f6f0f65bebafaea7...)
+  student_dropout... OK (c01207db62dc1ac9...)
+  student_health... OK (0fa081e63bb3f259...)
+  students_exam_scores... OK (336969cba9ef0bb7...)
+  tae... OK (356cb12fda2ff552...)
+  turkiye... OK (fef1729ce8edf310...)
+  uci_student... OK (7f5a7696641b3586...)
+  uci_student_math... OK (2cbf9c120016bb28...)
+  wine_quality... OK (e26239cad8e85992...)
+  xapi_edu... OK (6470f36422d4ba1e...)
+  yacht... OK (527653450af99bdb...)

benchmark_reliability-0.3.1/verification/prepare_verify_py38.txt

@@ -0,0 +1,55 @@
+# brf registry prepare verification — canonical content hash — Python 3.8.20
+# generated 2026-10-01 07:25:34 CST
+# command: python scripts/compute_prepare_hashes.py --verify
+#
+  abalone... OK (3f7c6c66fdec85a6...)
+  air_quality... OK (4ee091d7c6cd29c2...)
+  airfoil... OK (48bcb721e3e8fce3...)
+  assistments... OK (9bd474c2e85e0778...)
+  auto_mpg... OK (a8b5a265bf2aada3...)
+  bike_sharing... OK (6759530a726ce5bb...)
+  boston_housing... OK (1626353c69b1b1bc...)
+  breast_cancer... ERROR: No module named 'sklearn'
+  california_housing... ERROR: No module named 'sklearn'
+  ccpp... OK (84f5f83018d7030f...)
+  climate_weather... ERROR: No module named 'sklearn'
+  college_scorecard... OK (2de9c254016f6d93...)
+  colleges_aaup... OK (a98f7f115c3fc707...)
+  colleges_usnews... ERROR: No module named 'scipy'
+  concrete... OK (3c4b71924aa1f5b7...)
+  cpu_act... OK (12e266a8e018e51f...)
+  credit_card... ERROR: No module named 'sklearn'
+  customer_churn... ERROR: No module named 'sklearn'
+  diabetes... ERROR: No module named 'sklearn'
+  electricity... ERROR: No module named 'sklearn'
+  energy_building... OK (c045e11976d5c01f...)
+  energy_efficiency... OK (3d75f02739b5e4f3...)
+  entrance_exam... OK (43bca05374772861...)
+  german_credit... ERROR: No module named 'sklearn'
+  higher_ed... OK (59be6940c05cd59b...)
+  kaggle_students_performance... ERROR: No module named 'kagglehub'
+  kdd_cup_2010... OK (b7d8e2aef91cbc6a...)
+  law_school... OK (b1f33959f2f4a362...)
+  linnerud... ERROR: No module named 'sklearn'
+  mathe... OK (ac788168060595a0...)
+  mm_tba... ERROR: No module named 'openpyxl'
+  nursery... OK (afd19fc4496c3904...)
+  oli... OK (a7d2ccf40aaf65ab...)
+  olympics... OK (4429c726958ccc09...)
+  oulad... OK (9aa535d041c9f390...)
+  pisa2015... ERROR: No module named 'pyreadstat'
+  pollution... OK (5768d5b4509b398f...)
+  real_estate... OK (1c67624e20a3948c...)
+  seoul_bike... OK (03d8e2b26f08a21a...)
+  student_absences... ERROR: No module named 'kagglehub'
+  student_depression... OK (f6f0f65bebafaea7...)
+  student_dropout... OK (c01207db62dc1ac9...)
+  student_health... ERROR: No module named 'kagglehub'
+  students_exam_scores... OK (336969cba9ef0bb7...)
+  tae... OK (356cb12fda2ff552...)
+  turkiye... OK (fef1729ce8edf310...)
+  uci_student... OK (7f5a7696641b3586...)
+  uci_student_math... OK (2cbf9c120016bb28...)
+  wine_quality... OK (e26239cad8e85992...)
+  xapi_edu... OK (6470f36422d4ba1e...)
+  yacht... OK (527653450af99bdb...)

benchmark_reliability-0.3.1/verification/prepare_verify_py38_legacy.txt

@@ -0,0 +1,112 @@
+# brf registry prepare verification — canonical content hash — Python 3.8.20
+# generated 2026-10-01 07:25:34 CST
+# command: python scripts/compute_prepare_hashes.py --verify
+#
+# SUPERSEDED. Produced before compute_prepare_sha256 was changed to hash
+# canonical bytes. Under the old np.save() byte-stream scheme this run reported 26
+# false MISMATCHes against data that was element-for-element identical to the
+# 3.13 run. Retained as the evidence that motivated the canonical encoding.
+
+  abalone... MISMATCH
+    expected: 5b6b32a69477bea5675cb302e1d6a945...
+    actual:   32041fecd1f68c3c7c1f37716a1546c2...
+  air_quality... OK (e78a0ea54f9dc86d...)
+  airfoil... OK (be1cb863965a847b...)
+  assistments... MISMATCH
+    expected: fb0d83a83cd4eecf4eafa52f7c6100ee...
+    actual:   0890b8928350b4a946f931a3b2e258fd...
+  auto_mpg... MISMATCH
+    expected: 41551b5d3f9a5feb37fc3a6532991dfd...
+    actual:   b4907d473fe1733c5f987cebfce72a0a...
+  bike_sharing... MISMATCH
+    expected: 5a55a9c0ef9a1ab85fa02e167d6254ea...
+    actual:   76c62f8826fe0e841e4d7a1b8f672b5e...
+  boston_housing... OK (ff3c536970b60eec...)
+  breast_cancer... ERROR: No module named 'sklearn'
+  california_housing... ERROR: No module named 'sklearn'
+  ccpp... OK (50d712fc6096d18b...)
+  climate_weather... ERROR: No module named 'sklearn'
+  college_scorecard... MISMATCH
+    expected: ae9a804d4def1a5aee8ef39f6a96693e...
+    actual:   ad66a2c2fed247d753dc7e0f051956f0...
+  colleges_aaup... MISMATCH
+    expected: 71e1b7db1f017580fdc8ecbae641e47f...
+    actual:   089661e485c416b01bd82f7a04c1e164...
+  colleges_usnews... ERROR: No module named 'scipy'
+  concrete... OK (d827c72762e42f6a...)
+  cpu_act... MISMATCH
+    expected: 8507aed1b25e7f879780e7a60f59cb06...
+    actual:   329e1ff0462193a08089ac584dd70f00...
+  credit_card... ERROR: No module named 'sklearn'
+  customer_churn... ERROR: No module named 'sklearn'
+  diabetes... ERROR: No module named 'sklearn'
+  electricity... ERROR: No module named 'sklearn'
+  energy_building... MISMATCH
+    expected: a587b2900e50f0493714966c16e6af6d...
+    actual:   0f14736ee89295f4279c9d30d4ae84ca...
+  energy_efficiency... OK (56a8509c2811ad32...)
+  entrance_exam... MISMATCH
+    expected: 73daf60a69b19a9ac2ff9a2d610e1af0...
+    actual:   066807ff742bc4f8b0e7f18e34f6b3c5...
+  german_credit... ERROR: No module named 'sklearn'
+  higher_ed... MISMATCH
+    expected: 71b480e5744f32b139fca955a86af930...
+    actual:   d1ef86ac2d396ed20f8fcc701049bcb5...
+  kaggle_students_performance... ERROR: No module named 'kagglehub'
+  kdd_cup_2010... MISMATCH
+    expected: f5379171e250daf624767701640e6243...
+    actual:   1a18aff48cad57cc155cd8928ceb96d5...
+  law_school... MISMATCH
+    expected: d99966e6e23e8774b53634c292df0877...
+    actual:   12945ab4890ddd8803fafd3af08cf2ef...
+  linnerud... ERROR: No module named 'sklearn'
+  mathe... MISMATCH
+    expected: 39ef0dae8f556e1ca526bb9551c3b7a6...
+    actual:   11d772e492f9b6caa11ed30cb64ee1cb...
+  mm_tba... ERROR: No module named 'openpyxl'
+  nursery... MISMATCH
+    expected: 3ba69383b901493d654b4e38cce34beb...
+    actual:   866a472bfbe942924097e848238487cb...
+  oli... MISMATCH
+    expected: 584b1624cf141fac677fb6d81d43eb9f...
+    actual:   4e0bd241e17c139603ad18a752f796fc...
+  olympics... MISMATCH
+    expected: 2f499125556a7f4221b0a9029d82a6ec...
+    actual:   337b4dc3f214fe0aaf73f83e7ed6873e...
+  oulad... MISMATCH
+    expected: 160907b3b6f20ee6ec6960c15559f080...
+    actual:   f1b56a2e4f40cff7e34663eae01c0552...
+  pisa2015... ERROR: No module named 'pyreadstat'
+  pollution... OK (02b01fb8648921c0...)
+  real_estate... OK (33b6e420f4961959...)
+  seoul_bike... MISMATCH
+    expected: 3e349da8a18d80a4462c34627e04ccbb...
+    actual:   36673a5b15a5a29c47d0eee19f816024...
+  student_absences... ERROR: No module named 'kagglehub'
+  student_depression... MISMATCH
+    expected: 959f6b8a2ff303fe598525f322b5aa14...
+    actual:   8a6e5bf75ce598602461b28dd2e1f49e...
+  student_dropout... MISMATCH
+    expected: 44f0577f5da770ade95934c476243714...
+    actual:   dbbfb63f5818677e00ac26bb1435641c...
+  student_health... ERROR: No module named 'kagglehub'
+  students_exam_scores... MISMATCH
+    expected: 209266b41fbda02c661e7c350deec769...
+    actual:   7be836a4e719f04e6db4df617a877b65...
+  tae... OK (d05565e68408f503...)
+  turkiye... MISMATCH
+    expected: 54dc857070e80e8217d7e14c92653b56...
+    actual:   29466167772a3dabb964fa689ea20579...
+  uci_student... MISMATCH
+    expected: fca0d5cb95050125554cd5ca841c2b0e...
+    actual:   3056d3cb28d419c249ad14984c388fec...
+  uci_student_math... MISMATCH
+    expected: 121c895f712f8492be1a932a5810d6ab...
+    actual:   d91e9ed9827d3bcf42c676662204a486...
+  wine_quality... MISMATCH
+    expected: fab3c99e6cae7f91afc9a5ceee877955...
+    actual:   f32af70bccc532bd553c35a671982896...
+  xapi_edu... MISMATCH
+    expected: 2b6bae74d0261408955484b4f4790c04...
+    actual:   d7329eb246a08a980061fbccd7d701dc...
+  yacht... OK (3c571dcc684e1a38...)