Fix stochastic sweep seed integrity and add actual demo

.github/workflows/ci.yml

@@ -0,0 +1,26 @@
+name: CI
+
+on:
+  push:
+  pull_request:
+
+jobs:
+  test:
+    strategy:
+      matrix:
+        python-version: ["3.11", "3.12", "3.13"]
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+      - uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python-version }}
+          cache: pip
+      - run: python -m pip install -e ".[dev]"
+      - run: python -m pytest
+      - name: Install media verifier
+        run: sudo apt-get update && sudo apt-get install -y ffmpeg
+      - name: Verify source-pinned demo
+        run: python scripts/verify_seed_integrity_demo.py

README.md

@@ -12,6 +12,14 @@ We identify *loop-action bias* as a mechanistic failure mode of outcome-only cre
 | Real ALFWorld, Haiku (scorer diff.) | 13.3% | **34.8%** | = branch-aware |
 | Real ALFWorld, Qwen-7B (scorer diff.) | 3.4% | **31.8%** | = branch-aware |
 
+## Actual workflow demo
+
+![CreditLab stochastic seed integrity demo](media/demo/seed-integrity.gif)
+
+[Watch the MP4](media/demo/seed-integrity.mp4) · [Inspect the capture receipt](media/demo/capture.json)
+
+This provider-free demo executes the checked-in `stochastic/v1` manifest through the real environment adapter. It shows that one seed is repeatable while seeds 0 and 1 produce distinct stochastic paths. The capture is pinned to its exact runtime source commit and guarded by `scripts/verify_seed_integrity_demo.py`. It does not exercise a language model or validate the paper's performance claims.
+
 ## Reproducing the Results
 
 ### Prerequisites

media/demo/capture.json

@@ -0,0 +1,133 @@
+{
+  "artifacts": {
+    "seed-integrity.gif": {
+      "bytes": 252820,
+      "sha256": "d3eb157c762bb3c4677ea37fcb66916e21837ed5b48c9a95fe51cf5aafbf00d3"
+    },
+    "seed-integrity.mp4": {
+      "bytes": 164354,
+      "sha256": "45044b69f3d3d0650361d4c1e515d882bc2273106fb03ce7b9b4aa9ab05e2988"
+    }
+  },
+  "benchmark": "stochastic/v1",
+  "cross_seed_paths_distinct": true,
+  "same_seed_replay_matches": true,
+  "schema_version": 1,
+  "scope": "provider-free stochastic manifest fixture; no model or paper-performance claim",
+  "seeds": [
+    0,
+    1
+  ],
+  "source_blobs": {
+    "benchmarks/stochastic/v1/tasks.yaml": "e2826242df4cc340b67eba5c05b700059e5a806c",
+    "scripts/generate_seed_integrity_demo.py": "81e1cefd3b0d5f0f0f4da68ce6d0e31cc5d18a8a",
+    "src/creditlab/analysis/sweep.py": "1bf34600f9c4dd140702bc3370ac903dd4383a34",
+    "src/creditlab/core/registry.py": "0bd254dd96460beea41035fb2c0cf9eb8d0f3b72",
+    "src/creditlab/envs/manifest_graph.py": "6d7f4b70422001e97ab84869df060d4e24f32b24"
+  },
+  "source_commit": "b39ab74c07cc56f30e96fd8d955170e012167356",
+  "source_tree": "ffc8450ecf5b4a271e3ca3fc04ae1f45f86aeeb9",
+  "task_id": "stoch_gate_train_01",
+  "traces": {
+    "0": [
+      {
+        "action": "unlock_gate",
+        "from": "start",
+        "observation": "The lock resists. The gate stays shut.",
+        "terminal": false,
+        "to": "start",
+        "turn": 1
+      },
+      {
+        "action": "unlock_gate",
+        "from": "start",
+        "observation": "The lock resists. The gate stays shut.",
+        "terminal": false,
+        "to": "start",
+        "turn": 2
+      },
+      {
+        "action": "unlock_gate",
+        "from": "start",
+        "observation": "The first gate opens.",
+        "terminal": false,
+        "to": "gate_2",
+        "turn": 3
+      },
+      {
+        "action": "unlock_gate",
+        "from": "gate_2",
+        "observation": "The second gate opens.",
+        "terminal": false,
+        "to": "gate_3",
+        "turn": 4
+      },
+      {
+        "action": "unlock_gate",
+        "from": "gate_3",
+        "observation": "The third gate opens.",
+        "terminal": false,
+        "to": "gate_4",
+        "turn": 5
+      },
+      {
+        "action": "unlock_gate",
+        "from": "gate_4",
+        "observation": "The final gate opens. Task complete.",
+        "terminal": true,
+        "to": "success",
+        "turn": 6
+      }
+    ],
+    "1": [
+      {
+        "action": "unlock_gate",
+        "from": "start",
+        "observation": "The first gate opens.",
+        "terminal": false,
+        "to": "gate_2",
+        "turn": 1
+      },
+      {
+        "action": "unlock_gate",
+        "from": "gate_2",
+        "observation": "The lock resists again.",
+        "terminal": false,
+        "to": "gate_2",
+        "turn": 2
+      },
+      {
+        "action": "unlock_gate",
+        "from": "gate_2",
+        "observation": "The lock resists again.",
+        "terminal": false,
+        "to": "gate_2",
+        "turn": 3
+      },
+      {
+        "action": "unlock_gate",
+        "from": "gate_2",
+        "observation": "The second gate opens.",
+        "terminal": false,
+        "to": "gate_3",
+        "turn": 4
+      },
+      {
+        "action": "unlock_gate",
+        "from": "gate_3",
+        "observation": "The third gate opens.",
+        "terminal": false,
+        "to": "gate_4",
+        "turn": 5
+      },
+      {
+        "action": "unlock_gate",
+        "from": "gate_4",
+        "observation": "The final gate opens. Task complete.",
+        "terminal": true,
+        "to": "success",
+        "turn": 6
+      }
+    ]
+  }
+}

pyproject.toml

@@ -27,6 +27,9 @@ wandb = [
 paper = [
   "matplotlib>=3.8,<4",
 ]
+demo = [
+  "pillow>=10,<13",
+]
 alfworld = [
   "alfworld==0.4.2",
 ]