diff --git a/.github/workflows/cache.yml b/.github/workflows/cache.yml
index e617bc8..3cbf454 100644
--- a/.github/workflows/cache.yml
+++ b/.github/workflows/cache.yml
@@ -23,8 +23,19 @@ jobs:
       - name: Install JAX, Numpyro
         shell: bash -l {0}
         run: |
-          pip install --upgrade "jax[cuda12-local]"
-          pip install numpyro  
+          # CUDA 12 stack. .github/runs-on.yml pins ami-09baf66e396fa7cfd, a CUDA 12
+          # AMI, so do not move to cu130 / jax[cuda13] without bumping the AMI first.
+          # torch and pyro-ppl were never installed here even though bayes_nonconj.md
+          # imports both -- this workflow has never produced a successful build.
+          pip install torch --index-url https://download.pytorch.org/whl/cu128
+          pip install pyro-ppl
+          pip install "jax[cuda12]==0.7.1"
+          pip install numpyro
+          # See the equivalent block in ci.yml for why the conda stack must be held.
+          python -c "import importlib.metadata as md; d={x.metadata['Name'].lower():x.version for x in md.distributions()}; print('\n'.join(f'{p}=={d[p]}' for p in ('numpy','pandas','scipy') if p in d))" > /tmp/conda-pins.txt
+          cat /tmp/conda-pins.txt
+          pip install --only-binary=:all: -c /tmp/conda-pins.txt "arviz<1" "pymc<6" "kaleido<1" "prettytable<3.18"
+          python -c "import numpy, numba; assert numpy.__version__.startswith('1.'), numpy.__version__"
           python scripts/test-jax-install.py
       - name: Check nvidia drivers
         shell: bash -l {0}
@@ -44,4 +55,8 @@ jobs:
         uses: actions/upload-artifact@v6
         with:
           name: build-cache
-          path: _build
\ No newline at end of file
+          path: _build
+          # _config.yml sets execute_notebooks: "cache", so the execution cache
+          # lives in _build/.jupyter_cache. upload-artifact excludes hidden files
+          # by default, which would ship a "cache" containing no cache.
+          include-hidden-files: true
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index b17ca3f..b709608 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -23,10 +23,30 @@ jobs:
       - name: Install JAX, Numpyro
         shell: bash -l {0}
         run: |
-          pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128
+          # CUDA 12 stack. .github/runs-on.yml pins ami-09baf66e396fa7cfd, a CUDA 12
+          # AMI, so do not move to cu130 / jax[cuda13] without bumping the AMI first.
+          pip install torch --index-url https://download.pytorch.org/whl/cu128
           pip install pyro-ppl
-          pip install --upgrade "jax[cuda12-local]"
-          pip install numpyro  
+          pip install "jax[cuda12]==0.7.1"
+          pip install numpyro
+          # arviz and pymc are pre-installed at versions the conda stack can carry, so
+          # the notebooks' own unpinned `!pip install arviz pymc` cells become no-ops.
+          # They must not be allowed to move the conda-provided scientific stack: left
+          # free, xarray drags pandas forward, pandas 3.x requires numpy>=2, and pip
+          # backtracks through pandas until it reaches an sdist that fails to build
+          # (the numpy Cython headers need Cython>=3). numpy>=2 would also break numba
+          # 0.60.0, which caps at numpy<2.1 and is imported by 9 lectures.
+          # The constraints file is derived from the env rather than hardcoded so that
+          # bumping anaconda in environment.yml does not silently invalidate it.
+          # prettytable<3.18 is a real fix, not hygiene: prettytable declares a bare
+          # "wcwidth" with no floor, so pip accepts conda's wcwidth 0.2.x as satisfying
+          # it, but 3.18.0 calls wcwidth.width(), which only exists from wcwidth 0.3.0.
+          # That AttributeError is what broke prob_matrix.md in run 30975106893.
+          # 3.17.0 still uses wcswidth(), which conda's wcwidth has.
+          python -c "import importlib.metadata as md; d={x.metadata['Name'].lower():x.version for x in md.distributions()}; print('\n'.join(f'{p}=={d[p]}' for p in ('numpy','pandas','scipy') if p in d))" > /tmp/conda-pins.txt
+          cat /tmp/conda-pins.txt
+          pip install --only-binary=:all: -c /tmp/conda-pins.txt "arviz<1" "pymc<6" "kaleido<1" "prettytable<3.18"
+          python -c "import numpy, numba; assert numpy.__version__.startswith('1.'), numpy.__version__"
           python scripts/test-jax-install.py
         # Check nvidia drivers
       - name: nvidia Drivers
@@ -56,7 +76,7 @@ jobs:
         uses: actions/upload-artifact@v6
         if: failure()
         with:
-          name: execution-reports
+          name: execution-reports-notebooks
           path: _build/jupyter/reports
       - name: Build PDF from LaTeX
         shell: bash -l {0}
@@ -68,7 +88,7 @@ jobs:
         uses: actions/upload-artifact@v6
         if: failure()
         with:
-          name: execution-reports
+          name: execution-reports-latex
           path: _build/latex/reports
       # Final Build of HTML
       - name: Build HTML
@@ -79,7 +99,7 @@ jobs:
         uses: actions/upload-artifact@v6
         if: failure()
         with:
-          name: execution-reports
+          name: execution-reports-html
           path: _build/html/reports
       - name: Preview Deploy to Netlify
         uses: nwtgck/actions-netlify@v3
diff --git a/.github/workflows/collab.yml b/.github/workflows/collab.yml
index a3f0fe5..8870252 100644
--- a/.github/workflows/collab.yml
+++ b/.github/workflows/collab.yml
@@ -37,6 +37,11 @@ jobs:
         shell: bash -l {0}
         run: |
           pip install jupyter-book==1.0.3 quantecon-book-theme==0.8.2 sphinx-tojupyter==0.3.0 sphinxext-rediraffe==0.2.7 sphinxcontrib-youtube==1.3.0 sphinx-togglebutton==0.3.2 arviz sphinx-proof sphinx-exercise sphinx-reredirects
+          # The Colab image ships plotly 5.24.1 and no kaleido, so back_prop.md's
+          # `!pip install kaleido` cell pulls kaleido 1.x, which is incompatible.
+          # Pinning here makes that cell a no-op. Unlike the conda workflows, this
+          # container is a coherent numpy 2.0.2 stack, so it needs no numpy pin.
+          pip install "kaleido<1"
       # Build of HTML (Execution Testing)
       - name: Build HTML
         shell: bash -l {0}
diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index cae7931..0e39387 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -25,8 +25,17 @@ jobs:
       - name: Install JAX, Numpyro
         shell: bash -l {0}
         run: |
-          pip install --upgrade "jax[cuda12-local]"
-          pip install numpyro  
+          # CUDA 12 stack. .github/runs-on.yml pins ami-09baf66e396fa7cfd, a CUDA 12
+          # AMI, so do not move to cu130 / jax[cuda13] without bumping the AMI first.
+          pip install torch --index-url https://download.pytorch.org/whl/cu128
+          pip install pyro-ppl
+          pip install "jax[cuda12]==0.7.1"
+          pip install numpyro
+          # See the equivalent block in ci.yml for why the conda stack must be held.
+          python -c "import importlib.metadata as md; d={x.metadata['Name'].lower():x.version for x in md.distributions()}; print('\n'.join(f'{p}=={d[p]}' for p in ('numpy','pandas','scipy') if p in d))" > /tmp/conda-pins.txt
+          cat /tmp/conda-pins.txt
+          pip install --only-binary=:all: -c /tmp/conda-pins.txt "arviz<1" "pymc<6" "kaleido<1" "prettytable<3.18"
+          python -c "import numpy, numba; assert numpy.__version__.startswith('1.'), numpy.__version__"
           python scripts/test-jax-install.py
       - name: Check nvidia drivers
         shell: bash -l {0}