Widen supported Python versions to 3.9-3.12

.github/workflows/ci.yml

@@ -1,10 +1,12 @@
 # CI — lint + the full default test suite (unit + integration + regression) on the
-# pure-Python core (no Julia/[vle] extra installed). `pytest -q`'s own addopts
-# (`-m 'not vle'`) deselects the Julia-backed tests automatically -- confirmed none of
-# the non-`vle`-marked tests import anything Julia-touching at module import time, so
-# this runs clean without the [vle] extra. The frozen stack (TF 2.16.2 / GPflow 2.9.2 /
-# Py 3.9) is what's pinned in pyproject.toml; Julia-backed VLE + regression tests that
-# need it stay a separate, gated job for whenever CI grows a Julia-capable runner.
+# pure-Python core (no Julia/[vle] extra installed), across every Python version this
+# package claims to support (3.9-3.12; see pyproject.toml's requires-python and
+# docs/installation.md for why 3.13+ isn't -- upstream, GPflow requires numpy<2 and no
+# numpy 1.x supports 3.13). `pytest -q`'s own addopts (`-m 'not vle'`) deselects the
+# Julia-backed tests automatically -- confirmed none of the non-`vle`-marked tests
+# import anything Julia-touching at module import time, so this runs clean without the
+# [vle] extra. Julia-backed VLE + regression tests that need it stay a separate, gated
+# job for whenever CI grows a Julia-capable runner.
 name: CI
 
 on:
@@ -14,7 +16,7 @@ on:
     branches: [main]
 
 jobs:
-  test:
+  lint:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
@@ -27,5 +29,21 @@ jobs:
           pip install -e ".[dev]"
       - name: Lint (ruff)
         run: ruff check .
+
+  test:
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.9", "3.10", "3.11", "3.12"]
+    steps:
+      - uses: actions/checkout@v4
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+      - name: Install (core + dev)
+        run: |
+          python -m pip install --upgrade pip
+          pip install -e ".[dev]"
       - name: Run the default test suite (unit + integration + regression, not vle)
         run: pytest -q

.readthedocs.yaml

@@ -7,12 +7,16 @@
 # the autodoc_mock_imports fallback if this ever becomes too slow/flaky on RTD's
 # builders). Julia/Clapeyron (the `[vle]` extra) are never installed here -- the docs
 # build never needs them (see docs/vle_example.md).
+#
+# Built on Python 3.12 (the top of the supported 3.9-3.12 range; see
+# pyproject.toml's requires-python) so autodoc's real import of the TF-backed stack
+# genuinely exercises the newest supported interpreter, not just the oldest.
 version: 2
 
 build:
   os: ubuntu-24.04
   tools:
-    python: "3.9"
+    python: "3.12"
 
 sphinx:
   configuration: docs/conf.py

CHANGELOG.md

@@ -7,7 +7,19 @@ entry below summarizes the library relative to the paper's original research cod
 
 ## [Unreleased]
 
-Nothing yet.
+### Added
+
+- Support for Python 3.10, 3.11, and 3.12, alongside the existing 3.9 (`pyproject.toml`'s
+  `requires-python` widened from `>=3.9,<3.10` to `>=3.9,<3.13`; Python 3.9 alone is EOL
+  as of October 2025). Verified with a full test-suite run -- including the exact-value
+  regression pins (`tests/integration/data/synthetic_baseline.json`, atol 1e-10) and the
+  paper reference-value regressions -- on all four versions; no dependency version
+  changed. `.github/workflows/ci.yml` now runs the default suite across a 3.9-3.12
+  matrix. `.readthedocs.yaml` builds on Python 3.12.
+- A documented explanation of why Python 3.13+ isn't supported: GPflow requires
+  `numpy<2` in every release, and no NumPy 1.x publishes a Python 3.13 wheel -- an
+  upstream constraint, not a limitation of this package's own pins. See
+  `docs/installation.md`.
 
 ## [0.1.2] - 2026-08-06
 

README.md

@@ -31,6 +31,11 @@ The **core library is pure Python** (GPflow / TensorFlow / NumPy / SciPy) with n
 dependency. Julia + Clapeyron are only needed for the `vle_distillation` example, which
 isn't part of the PyPI package -- see "From source" below.
 
+Supports **Python 3.9-3.12**. Python 3.13+ isn't available: this package depends on
+GPflow, and GPflow requires `numpy<2` in every release -- no NumPy 1.x publishes a
+Python 3.13 wheel. That's an upstream constraint, not something this package can work
+around; see [`docs/installation.md`](docs/installation.md) for the full explanation.
+
 **macOS note:** set `export PYTHON_JULIACALL_HANDLE_SIGNALS=yes` before importing
 `juliacall`, or Julia crashes with a bus error (SIGBUS).
 

docs/index.md

@@ -8,9 +8,9 @@ design with Bayesian hierarchical Gaussian-process (GP) surrogates. Prior physic
 knowledge is encoded through priors on the GP hyperparameters; new data is chosen by
 maximizing the predictive differential entropy of the resulting hierarchical
 posterior, so *hyperparameter uncertainty* -- not just predictive variance -- drives
-data acquisition. That distinction is the paper's contribution and the reason the
-library exists: a plain GP's predictive variance ignores uncertainty in the
-hyperparameters themselves, which matters most exactly when data is scarce.
+data acquisition. Many other adaptive sampling methods in literature use the GP's predictive variance 
+and ignore uncertainty in the hyperparameters themselves, which (we argue) matters most exactly when data is scarce.
+BITS for GAPS addresses this limitation.
 
 ```{admonition} Reference
 K. D. Jones and A. W. Dowling, "BITS for GAPS: Bayesian Information-Theoretic

docs/installation.md

@@ -10,13 +10,26 @@ optional VLE/distillation example, and even then only when you actually call int
 pip install bits_for_gaps
 ```
 
-```{admonition} Frozen dependency stack
+```{admonition} Frozen dependency stack, Python 3.9-3.12
 :class: note
-The core pins an exact, verified-working stack rather than floating version ranges:
-Python 3.9, NumPy 1.26, SciPy 1.13, GPflow 2.9.2, TensorFlow 2.16.2, TensorFlow
-Probability 0.24.0. This is a deliberate reproducibility choice -- GPflow's TensorFlow
-dependency makes casual version bumps risky, so modernizing the stack is left as a
-separate, later effort.
+The core pins an exact, verified-working dependency stack rather than floating version
+ranges: NumPy 1.26, SciPy 1.13, GPflow 2.9.2, TensorFlow 2.16.2, TensorFlow Probability
+0.24.0. This is a deliberate reproducibility choice -- GPflow's TensorFlow dependency
+makes casual version bumps risky, so modernizing the stack is left as a separate, later
+effort. The package itself supports **Python 3.9 through 3.12**, verified with a real
+test-suite run (including the exact-value regression pins) on every one of those four
+versions.
+```
+
+```{admonition} Why not Python 3.13+?
+:class: important
+This is an **upstream constraint**, not a project choice: `bits_for_gaps` depends on
+GPflow, and GPflow requires `numpy<2` in every release it has ever published, including
+the latest (2.11.1). No 1.x release of NumPy publishes a Python 3.13 wheel -- the first
+NumPy version that does is 2.1.0. So GPflow (and therefore this package) cannot run on
+Python 3.13 at all, regardless of what this package's own pins say. TensorFlow's newest
+release similarly has no Python 3.14 wheels yet. Python 3.12 is the ceiling until GPflow
+adds NumPy 2 support upstream -- no timeline is promised or implied.
 ```
 
 Verify it imports (no Julia touched):
@@ -30,7 +43,8 @@ python -c "import bits_for_gaps; print(bits_for_gaps.__version__)"
 ```bash
 git clone https://github.com/dowlinglab/bits_for_gaps
 cd bits_for_gaps
-conda env create -f environment.yml   # Python 3.9 + the pinned stack
+conda env create -f environment.yml   # Python 3.12 + the pinned stack (see the file's
+                                       # header comment for using 3.9 instead)
 conda activate bits_for_gaps
 pip install -e .
 ```

environment.yml

@@ -5,14 +5,25 @@
 #   conda activate bits_for_gaps
 #   pip install -e ".[dev,vle]"
 #
+# Python 3.12 (moved here from 3.9 -- see CHANGELOG.md's Python-3.9-3.12 entry): the
+# package supports 3.9-3.12, and this dev/reproduction env now uses the top of that
+# range so day-to-day development and `paper/reproduce.py`/`paper/full_reproduction.py`
+# runs exercise the newest supported interpreter. Verified with a from-scratch,
+# ~28-minute `paper/full_reproduction.py` rerun on 3.12: HMC R-hat/ESS, the
+# hyperparameter posterior, and the McCabe-Thiele column all reproduce the paper's
+# statistical/qualitative behavior (see paper/REPRODUCTION.md). If you specifically
+# need the interpreter the archived published results (`paper/data/`, `paper/reference/`)
+# were originally produced on, change `python=3.12` below to `python=3.9` -- every
+# other pin in this file is unaffected either way.
+#
 # IMPORTANT (macOS): export PYTHON_JULIACALL_HANDLE_SIGNALS=yes before running
 # anything that imports juliacall, or Julia crashes with a bus error (SIGBUS).
 name: bits_for_gaps
 channels:
   - conda-forge
   - defaults
 dependencies:
-  - python=3.9
+  - python=3.12
   - pip
   - numpy=1.26.*
   - scipy=1.13.*

paper/REPRODUCTION.md

@@ -213,3 +213,53 @@ considerably tighter than "qualitative" on the deterministic parts of the pipeli
 (HMC diagnostics, hyperparameter posterior, entropy field). Reproduce via
 `paper/full_reproduction.py` (module docstring has the exact constants used); not
 gated in CI (see above).
+
+### Cross-Python-version check: the same rerun on Python 3.12
+
+The numbers above are from a Python 3.9 run (this package's original dev environment).
+When widening supported Python versions to 3.9-3.12, `paper/full_reproduction.py` was
+rerun from scratch on **3.12** (same TF 2.16.2/GPflow 2.9.2 pins -- only the
+interpreter changed) to confirm the science, not just the test suite, holds across
+versions. Completed in 0.47 h (28 min, inside the documented ~25-30 min range).
+
+**Iterations 1-2 reproduced the 3.9 run's max entropy to 6+ significant figures**
+(1.4576908 vs. 1.4577 at iteration 1; 1.2470073 vs. 1.2470 at iteration 2) -- striking,
+given no bitwise-reproducibility guarantee was made across interpreters. From
+iteration 3 on, the two runs' design trajectories diverge (expected: each iteration's
+`predict_grid`-style posterior-predictive diagnostic draws from TensorFlow's ambient,
+unseeded RNG -- see `bits_for_gaps/mixture.py`'s module docstring -- so the *same*
+non-reproducibility this package already documents *within* one Python version also
+separates two versions' runs once it fires). What matters for this check is that the
+statistical conclusions survive that divergence:
+
+| | R-hat (iter 15) | ESS (iter 15) |
+|---|---|---|
+| Python 3.9 fresh run (above) | 1.00523, 1.00730, 1.00879 | 1468.29, 2428.09, 653.15 |
+| Python 3.12 fresh run | 1.00527, 1.00736, 1.00879 | 1468.06, 2424.94, 653.15 |
+
+Both comfortably under the R-hat < 1.1 threshold, both healthy ESS -- HMC convergence
+is not Python-version-sensitive.
+
+| | mean | median |
+|---|---|---|
+| Python 3.9 fresh run (above) | 1.35645, 0.86239, 3.19502 | 1.28619, 0.81949, 3.04407 |
+| Python 3.12 fresh run | 1.35606, 0.86216, 3.19502 | 1.28501, 0.81881, 3.04407 |
+
+Same ordering (`lengthscale_2` > `lengthscale_1`) as every other run of this model.
+
+Max entropy still decays from **1.458** (iteration 1) to **-0.224** (iteration 15),
+crossing zero in the same iteration range as the 3.9 run. Test RMSE still drops from
+**4.385** (iteration 1) to **0.856** (iteration 15) -- a ~5.1x reduction, matching the
+3.9 fresh run's ~4.9x (4.337 -> 0.887) and the paper's own ~6.5x (4.34 -> 0.67); same
+non-monotonic-between-iterations trend, expected for an entropy- (not error-) driven
+acquisition. Both the Wilson and the 15-iteration-adaptive surrogate McCabe-Thiele
+columns converged, tracking each other within 0.027 mole fraction (liquid) / 0.014
+(vapor) at every stage -- inside the 0.03/0.014 band the 3.9 run and the regression
+suite's own tolerance already use.
+
+**Conclusion: Python 3.12 gives statistically equivalent science to Python 3.9**, with
+the deterministic parts of the pipeline (HMC diagnostics, hyperparameter posterior
+ordering) landing close enough to look almost version-independent, and the
+RNG-sensitive parts (entropy trajectory past iteration 2, test RMSE, column stage
+values) landing in the same regime as every other independent rerun of this model,
+3.9 included.

pyproject.toml

@@ -7,7 +7,7 @@ name = "bits_for_gaps"
 dynamic = ["version"]
 description = "Bayesian Information-Theoretic Sampling for hierarchical GAussian Process Surrogates (BITS for GAPS)"
 readme = "README.md"
-requires-python = ">=3.9,<3.10"
+requires-python = ">=3.9,<3.13"
 license = { file = "LICENSE" }
 authors = [
     { name = "Alexander W. Dowling", email = "adowling@nd.edu" },
@@ -23,14 +23,20 @@ classifiers = [
     "License :: OSI Approved :: BSD License",
     "Programming Language :: Python :: 3",
     "Programming Language :: Python :: 3.9",
+    "Programming Language :: Python :: 3.10",
+    "Programming Language :: Python :: 3.11",
+    "Programming Language :: Python :: 3.12",
     "Topic :: Scientific/Engineering",
     "Topic :: Scientific/Engineering :: Artificial Intelligence",
     "Operating System :: OS Independent",
 ]
 
 # Frozen reproduction baseline: an exact, verified-working stack rather than floating
 # version ranges, since GPflow's TensorFlow dependency makes casual version bumps
-# risky. tensorflow-macos on Apple Silicon; plain tensorflow elsewhere.
+# risky. tensorflow-macos on Apple Silicon; plain tensorflow elsewhere. Verified on
+# Python 3.9-3.12 (see docs/installation.md); the numpy<2 pin is also why 3.13+ isn't
+# available -- GPflow requires numpy<2 in every release, and no numpy 1.x supports
+# Python 3.13 (an upstream constraint -- see docs/installation.md).
 dependencies = [
     "numpy>=1.26,<2",
     "scipy>=1.13,<1.14",