Add installable ovllm verifier and signed publish loop

.github/workflows/publish-verified-model.yml

@@ -1,8 +1,86 @@
 name: Publish Verified Model
+
 on:
   workflow_dispatch:
+    inputs:
+      model_name:
+        description: "Name to put in ovllm_manifest.json"
+        required: true
+        default: "gpt10m-shakespeare"
+      hf_repo_id:
+        description: "Optional Hugging Face repo id, e.g. owner/gpt10m-shakespeare"
+        required: false
+        default: ""
+      publish_to_hf:
+        description: "Upload the signed directory to Hugging Face"
+        required: true
+        default: false
+        type: boolean
+
+permissions:
+  contents: read
+  id-token: write
+
 jobs:
-  placeholder-job:
+  sign-and-verify:
     runs-on: ubuntu-latest
+
     steps:
-      - run: echo "This is just a temporary placeholder"
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+          cache: "pip"
+
+      - name: Install dependencies
+        run: |
+          python -m pip install -r requirements.txt
+          python -m pip install -e .
+
+      - name: Train real model (smoke steps)
+        run: |
+          python run_experiment.py \
+            --model gpt10m \
+            --dataset shakespeare \
+            --precision fp32 \
+            --deterministic on \
+            --steps 8 \
+            --batch-size 8 \
+            --block-size 64 \
+            --keep-artifact
+
+      - name: Prepare publish directory
+        run: |
+          ovllm prepare-publish \
+            --weights artifacts/gpt10m_shakespeare_fp32_deton_s99.safetensors \
+            --out dist/gpt10m-shakespeare \
+            --name "${{ inputs.model_name }}"
+
+      - name: Sign with GitHub Actions OIDC
+        env:
+          SIGSTORE_IDENTITY: "https://github.com/${{ github.repository }}/.github/workflows/publish-verified-model.yml@${{ github.ref }}"
+          SIGSTORE_IDENTITY_PROVIDER: "https://token.actions.githubusercontent.com"
+        run: |
+          ovllm sign dist/gpt10m-shakespeare \
+            --identity "$SIGSTORE_IDENTITY" \
+            --identity-provider "$SIGSTORE_IDENTITY_PROVIDER" \
+            --use-ambient-credentials
+
+      - name: Verify signed artifact
+        run: ovllm verify dist/gpt10m-shakespeare --skip-replay
+
+      - name: Upload signed model directory
+        uses: actions/upload-artifact@v4
+        with:
+          name: gpt10m-shakespeare-signed
+          path: dist/gpt10m-shakespeare
+
+      - name: Publish to Hugging Face
+        if: ${{ inputs.publish_to_hf && inputs.hf_repo_id != '' }}
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+        run: |
+          ovllm publish-hf "${{ inputs.hf_repo_id }}" dist/gpt10m-shakespeare

.gitignore

@@ -5,6 +5,10 @@ venv/
 
 __pycache__/
 *.pyc
+build/
+dist/
+*.egg-info/
+.ovllm-cache/
 
 
 .vscode/
@@ -39,3 +43,6 @@ data/*.zip
 data/*.tar.gz
 data/cifar-10-batches-py/
 !data/shakespeare_sample.txt
+
+# Local smoke tests and temp files
+openverifiable-smoke/

CONTRIBUTING.md

@@ -40,12 +40,12 @@ Dependencies are pinned and hash-locked. This is deliberate: it's both a reprodu
 git clone https://github.com/<your-username>/OpenVerifiableLLM
 cd OpenVerifiableLLM
 
-# install the exact, locked dependency set
-uv sync
+# install dependencies and the editable CLI
+pip install -r requirements.txt
+pip install -e .
 
 # confirm things work
-pytest tests/falsifiability
-ruff check .
+python -m unittest discover -s tests
 ```
 
 If you need to add a dependency, add it through the lockfile so the hash-locked set stays authoritative. An unpinned dependency undermines both reproducibility and the project's supply-chain posture.
@@ -58,8 +58,8 @@ If you need to add a dependency, add it through the lockfile so the hash-locked
 
 **Tests.**
 
-- Run the test suite and `ruff` locally before opening a PR. The determinism checks in CI must pass; a PR that breaks them won't merge.
-- Code that changes verification behavior, serialization, RNG/optimizer state handling, or the manifest should include tests for both the success case and the failure case. For verification-affecting changes, extend the falsifiability suite so a clean run passes and the relevant tampered run fails.
+- Run the test suite (`python -m unittest discover -s tests`) locally before opening a PR. The determinism checks in CI must pass; a PR that breaks them won't merge.
+- Code that changes verification behavior, serialization, RNG/optimizer state handling, or the manifest should include tests for both the success case and the failure case. For verification-affecting changes, extend the test suite so a clean run passes and the relevant tampered run fails.
 - A change must not silently weaken the bit-exact reproducibility guarantee on the supported single-GPU stack. If a change trades determinism for performance, make that tradeoff explicit and document it.
 
 **Docs.** If a change alters behavior, inputs, the manifest format, or the verification contract, update the relevant docs in the same PR. The manifest schema is a versioned contract that other parts of the project depend on, so changes there need maintainer sign-off and a migration note.
@@ -68,7 +68,7 @@ If you need to add a dependency, add it through the lockfile so the hash-locked
 
 - Reference the issue it addresses (`Closes #123`) if there is one.
 - Describe what changed and why, and flag anything you'd like reviewers to look at closely.
-- Confirm the test suite and `ruff` pass.
+- Confirm the test suite passes.
 - Keep each PR to one logical change. Small, focused PRs get reviewed and merged faster than large, multi-purpose ones.
 
 Maintainers review on a volunteer basis, so a clear, well-tested, well-scoped PR is the best way to get a quick response. Expect some back-and-forth; review comments are about the work, not about you.