anle2024 · Oct 30, 2026 · Oct 31, 2026 · Nov 1, 2026 · Nov 2, 2026 · Nov 2, 2026
diff --git a/.github/workflows/kumo-engine-weekly.yml b/.github/workflows/kumo-engine-weekly.yml
@@ -36,6 +36,7 @@ jobs:
       matrix: ${{ steps.config.outputs.matrix }}
       manifest_sha256: ${{ steps.config.outputs.manifest_sha256 }}
       reference_sha256: ${{ steps.config.outputs.reference_sha256 }}
+      p1: ${{ steps.config.outputs.p1 }}
     steps:
       - uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6
         with:
@@ -70,8 +71,22 @@ jobs:
                   if not isinstance(path, str) or not PurePosixPath(path).is_absolute() or any(c in path for c in '\r\n'):
                       raise SystemExit('Bind absolute data paths and the shared host lock')
               profiles.append(profile)
+          p1 = value.get('p1', {})
+          if os.environ['MODE'] == 'candidate':
+              runner = p1['runner']
+              if (not isinstance(runner, dict) or set(runner) != {'group', 'labels'}
+                      or any(not isinstance(v, str) or not v for v in runner.values())):
+                  raise SystemExit('Bind the existing P1 runner group and label')
+              for key in ('input_root', 'reference', 'host_lock', 'work_root'):
+                  path = p1[key]
+                  if not isinstance(path, str) or not PurePosixPath(path).is_absolute() or any(c in path for c in '\r\n'):
+                      raise SystemExit('Bind absolute P1 input, reference, lock, and work paths')
+              for key in ('manifest_sha256', 'reference_sha256'):
+                  if not re.fullmatch('[0-9a-f]{64}', p1.get(key, '')):
+                      raise SystemExit('Bind the reviewed P1 manifest and reference checksums')
           with open(os.environ['GITHUB_OUTPUT'], 'a') as output:
               output.write('matrix=' + json.dumps(dict(include=profiles)) + '\n')
+              output.write('p1=' + json.dumps(p1) + '\n')
               for key in ('manifest_sha256', 'reference_sha256'):
                   output.write(key + '=' + str(value.get(key, '')) + '\n')
           PYCODE
@@ -89,6 +104,7 @@ jobs:
       - run: python packages/kumo-engine/test/test_weekly_check.py
       - run: python packages/kumo-engine/test/test_weekly_reference.py
       - run: python packages/kumo-engine/test/test_weekly_workflow.py
+      - run: python packages/kumo-engine/test/test_weekly_p1.py
 
   wheel:
     if: github.event_name == 'workflow_dispatch' && inputs.mode == 'candidate'
@@ -317,6 +333,92 @@ jobs:
           shutil.rmtree(temporary / 'weekly-wheel', ignore_errors=True)
           PY
 
+  p1:
+    name: P1 CPU sampling
+    needs: [cpu, configure, gpu, wheel]
+    if: ${{ !cancelled() && github.event_name == 'workflow_dispatch' && inputs.mode == 'candidate' && needs.cpu.result == 'success' && needs.configure.result == 'success' && needs.wheel.result == 'success' }}
+    runs-on: ${{ fromJSON(needs.configure.outputs.p1).runner }}
+    timeout-minutes: 120
+    env:
+      P1_CONFIG: ${{ needs.configure.outputs.p1 }}
+      CUDA_VISIBLE_DEVICES: ''
+      KUMO_ACCELERATOR: cpu
+      OMP_NUM_THREADS: '8'
+      MKL_NUM_THREADS: '8'
+      OPENBLAS_NUM_THREADS: '8'
+      PYTHONDONTWRITEBYTECODE: '1'
+      PIP_CONFIG_FILE: /dev/null
+    steps:
+      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6
+        with:
+          persist-credentials: false
+      - uses: actions/setup-python@ece7cb06caefa5fff74198d8649806c4678c61a1 # v6
+        with:
+          python-version: '3.12'
+      - uses: actions/download-artifact@3e5f45b2cfb9172054b4087a40e8e0b5a5461e7c # v8
+        with:
+          name: weekly-wheel-${{ github.run_id }}-${{ github.run_attempt }}
+          path: ${{ runner.temp }}/weekly-p1-wheel
+      - name: Prepare an isolated P1 environment
+        run: |
+          python - <<'PY'
+          import json, os, runpy, subprocess
+          from pathlib import Path
+          api = runpy.run_path('packages/kumo-engine/scripts/weekly_check.py')
+          config = json.loads(os.environ['P1_CONFIG'])
+          wheels = Path(os.environ['RUNNER_TEMP']) / 'weekly-p1-wheel'
+          [wheel] = wheels.glob('*.whl')
+          build = api['read'](wheels / 'build.json')
+          api['check_build'](build, build['version'], api['digest'](wheel))
+          root = Path(config['work_root']) / ('weekly-p1-' + os.environ['GITHUB_RUN_ID'] + '-' + os.environ['GITHUB_RUN_ATTEMPT'])
+          root.mkdir(exist_ok=False)
+          with open(os.environ['GITHUB_ENV'], 'a') as output:
+              output.write(f'WEEKLY_P1_ROOT={root}\n')
+          subprocess.run(['python', '-m', 'venv', str(root / 'venv')], check=True)
+          python = str(root / 'venv/bin/python')
+          subprocess.run([python, '-I', '-m', 'pip', '--isolated', 'install', '--index-url', 'https://pypi.org/simple', str(wheel)], check=True)
+          subprocess.run([python, '-I', '-m', 'pip', '--isolated', 'check'], check=True)
+          PY
+      - name: Run P1 against its approved reference
+        run: |
+          python - <<'PY'
+          import json, os, runpy, subprocess
+          from pathlib import Path
+          api = runpy.run_path('packages/kumo-engine/scripts/weekly_check.py')
+          config = json.loads(os.environ['P1_CONFIG'])
+          root = Path(os.environ['WEEKLY_P1_ROOT'])
+          wheels = Path(os.environ['RUNNER_TEMP']) / 'weekly-p1-wheel'
+          [wheel] = wheels.glob('*.whl')
+          build = api['read'](wheels / 'build.json')
+          api['check_build'](build, build['version'], api['digest'](wheel))
+          subprocess.run([str(root / 'venv/bin/python'), '-I', 'packages/kumo-engine/scripts/weekly_p1.py',
+              '--mode', 'candidate', '--input-root', config['input_root'], '--manifest-sha256', config['manifest_sha256'],
+              '--reference', config['reference'], '--reference-sha256', config['reference_sha256'],
+              '--host-lock', config['host_lock'], '--output', str(root / 'evidence'),
+              '--wheel', str(wheel), '--wheel-sha256', build['wheel_sha256']], check=True)
+          PY
+      - uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7
+        if: always() && env.WEEKLY_P1_ROOT != ''
+        with:
+          name: weekly-p1-${{ github.run_id }}-${{ github.run_attempt }}
+          path: ${{ env.WEEKLY_P1_ROOT }}/evidence
+          if-no-files-found: error
+          retention-days: 7
+      - name: Remove this P1 job's temporary files
+        if: always() && env.WEEKLY_P1_ROOT != ''
+        run: |
+          python - <<'PY'
+          import json, os, shutil
+          from pathlib import Path
+          config = json.loads(os.environ['P1_CONFIG'])
+          root = Path(os.environ['WEEKLY_P1_ROOT']).resolve()
+          expected = 'weekly-p1-' + os.environ['GITHUB_RUN_ID'] + '-' + os.environ['GITHUB_RUN_ATTEMPT']
+          if root.parent != Path(config['work_root']).resolve() or root.name != expected:
+              raise SystemExit('Unexpected P1 cleanup directory')
+          shutil.rmtree(root)
+          shutil.rmtree(Path(os.environ['RUNNER_TEMP']) / 'weekly-p1-wheel', ignore_errors=True)
+          PY
+
   collect:
     needs: [configure, gpu]
     if: ${{ !cancelled() && needs.configure.result == 'success' }}

diff --git a/docs/development/engine-weekly-ci.md b/docs/development/engine-weekly-ci.md
@@ -36,8 +36,9 @@ restored weights with the saved checkpoint. Checkpoint-format failures, missing
 rows, changed weights, and excess score drift fail the case. A candidate also reloads its own saved model.
 
 P1 means five CPU sampling operations on commerce/100M. B9 means prediction with
-the retained P2 model on one GPU and then two matching GPUs. They are separate
-changes and do not run in this workflow.
+the retained P2 model on one GPU and then two matching GPUs. The separate
+[P1 job](engine-weekly-p1.md) follows the eleven cases during a manual candidate run.
+B9 remains a separate change.
 
 ## Inputs and reference approval
 

diff --git a/docs/development/engine-weekly-p1-calibration.json b/docs/development/engine-weekly-p1-calibration.json
@@ -0,0 +1,231 @@
+{
+  "format": "kumo-engine-weekly-p1-calibration-v1",
+  "approval": "PENDING_REVIEW",
+  "date_utc": "2026-11-02",
+  "physical_fits": 0,
+  "operations": 15,
+  "runtime": {
+    "version": "1.0.0",
+    "wheel_sha256": "1bc15ecb7696bce216b05aa883d6897d2b8ce4342c6c2f9d62b7cc28b314539e",
+    "dependencies": {
+      "cloudpickle": "3.1.2",
+      "mpmath": "1.3.0",
+      "nvidia-cuda-runtime-cu12": "12.9.79",
+      "optuna": "4.9.0",
+      "multidict": "7.1.0",
+      "psutil": "7.2.2",
+      "nvidia-nvjitlink": "13.4.92",
+      "certifi": "2026.10.19",
+      "torchmetrics": "1.8.1",
+      "nvidia-cusparse": "12.6.3.3",
+      "packaging": "26.3",
+      "cuda-pathfinder": "1.8.3",
+      "nvidia-cudnn-cu13": "9.24.0.43",
+      "nvidia-cusolver": "12.0.4.66",
+      "aiohappyeyeballs": "2.7.1",
+      "Werkzeug": "3.1.9",
+      "xxhash": "4.0.2",
+      "yarl": "1.25.1",
+      "tensorboard": "2.21.0",
+      "filelock": "4.2.4",
+      "nvidia-nvtx": "13.0.85",
+      "threadpoolctl": "3.7.0",
+      "absl-py": "2.5.1",
+      "fonttools": "4.67.0",
+      "antlr4-python3-runtime": "4.9.3",
+      "nvidia-cublas": "13.1.1.3",
+      "cycler": "0.12.1",
+      "pillow": "13.0.0",
+      "fsspec": "2026.9.0",
+      "nvidia-cufile": "1.15.1.6",
+      "frozenlist": "1.8.0",
+      "nvidia-curand-cu12": "10.3.10.19",
+      "typing-inspection": "0.4.4",
+      "aiohttp": "3.14.5",
+      "numpy": "2.5.4",
+      "propcache": "0.5.4",
+      "Markdown": "3.11.2",
+      "python-dateutil": "2.9.0.post0",
+      "matplotlib": "3.11.2",
+      "setuptools": "84.2.0",
+      "joblib": "1.6.0",
+      "tensorboard-data-server": "0.7.2",
+      "scikit-learn": "1.9.1",
+      "cuda-toolkit": "13.0.3.0",
+      "nvidia-cuda-runtime": "13.0.96",
+      "nvidia-cuda-nvrtc-cu12": "12.9.86",
+      "kiwisolver": "1.5.1",
+      "cuda-bindings": "13.4.4",
+      "pytorch-frame": "0.3.0",
+      "charset-normalizer": "3.5.2",
+      "nvidia-cuda-cupti": "13.0.85",
+      "tzdata": "2026.5",
+      "scipy": "1.18.1",
+      "tabulate": "0.8.10",
+      "nvidia-curand": "10.4.0.35",
+      "grpcio": "1.84.0",
+      "urllib3": "2.8.0",
+      "nvidia-cufft": "12.0.0.61",
+      "nvidia-nvshmem-cu13": "3.4.5",
+      "Mako": "1.4.3",
+      "aiosignal": "1.4.0",
+      "torch": "2.14.1",
+      "protobuf": "7.36.2",
+      "requests": "2.34.2",
+      "sympy": "1.14.0",
+      "pyparsing": "3.3.3",
+      "torch-geometric": "2.8.1",
+      "PyYAML": "6.0.3",
+      "tqdm": "4.70.1",
+      "pytz": "2026.5",
+      "narwhals": "2.27.1",
+      "MarkupSafe": "3.0.4",
+      "networkx": "3.7",
+      "pydantic": "2.14.1",
+      "nvidia-cusparselt-cu13": "0.8.1",
+      "nvidia-cuda-nvrtc": "13.0.88",
+      "nvidia-cublas-cu12": "12.9.2.10",
+      "pip": "24.0",
+      "nvidia-nccl-cu13": "2.30.7",
+      "attrs": "26.1.0",
+      "triton": "3.8.0",
+      "kumo-engine": "1.0.0",
+      "alembic": "1.20.0",
+      "six": "1.17.0",
+      "colorlog": "6.12.0",
+      "pytorch-lightning": "2.6.5",
+      "pyarrow": "24.0.0",
+      "SQLAlchemy": "2.1.5",
+      "typing_extensions": "4.16.0",
+      "pydantic_core": "2.50.1",
+      "idna": "3.20",
+      "faiss-gpu": "1.14.3",
+      "duckdb": "1.5.5",
+      "Jinja2": "3.1.6",
+      "annotated-types": "0.8.0",
+      "sqlglot": "30.17.0",
+      "lightning-utilities": "0.15.3",
+      "pandas": "2.3.3",
+      "contourpy": "1.4.0"
+    }
+  },
+  "workload": {
+    "n": 300000,
+    "unique_customers": 100000,
+    "offset": 0,
+    "hops": 3,
+    "batch": 256,
+    "procs": 2,
+    "threads": 8,
+    "plan": [
+      24,
+      8,
+      6
+    ],
+    "cache": "no_drop_uncontrolled",
+    "features": true,
+    "expected_work": {
+      "batches": 1172,
+      "requested_seeds": 300000
+    },
+    "historical_sampled_work": {
+      "nodes": 173052351,
+      "edges": 196315170
+    }
+  },
+  "graph_row_counts": {
+    "merchants": 6848,
+    "purchases": 17144193,
+    "merchant_views": 63497928,
+    "searches": 20142246,
+    "favorites": 796998,
+    "reviews": 1959145,
+    "customers": 2960000
+  },
+  "identical_work_per_operation": {
+    "seeds": 300000,
+    "batches": 1172,
+    "nodes": 173052351,
+    "edges": 196315170,
+    "fetched_rows": 173052351,
+    "membership_sha256": "008c24bd53606f71928bc1fd8e7dd055680cb322de0246747951b481b0564f57"
+  },
+  "sampling_function_sha256": "5a1b7b26e48fa95c6eef3747fcf5d8cb65add55e0b7c1b8e4eaa3b47726fae70",
+  "measurement_manifest_sha256": "2656adba90834d2c89b91673bc092a5f9f6eb20a4d6e216c48e9a036ee511752",
+  "workload_sha256": "8dc9e7ad80efed943310140bfac8d9235dc13045cacd0d591dbf46f4e4f5f85a",
+  "observations": [
+    {
+      "block": 1,
+      "status": "PASS",
+      "physical_fits": 0,
+      "raw_result_sha256": "fe22afcc846424b3646ac28ff370e01e5ad28e6b8b85c17ddc71da59f86eeff7",
+      "runner_sha256": "1b0722762d12fa546222dbcc3e25c0dc2309457dfa727b6edcb8f98a54e83bae",
+      "python": "3.12.3",
+      "command_elapsed": "PT7M37.684S",
+      "operation_seconds": [
+        90.02141279296484,
+        88.75617220095592,
+        88.70439774199622,
+        91.41735572600737,
+        89.70340646797558
+      ],
+      "median_seconds": 89.70340646797558
+    },
+    {
+      "block": 2,
+      "status": "PASS",
+      "physical_fits": 0,
+      "raw_result_sha256": "287c75361763cf00b5f4060eee901cc0703ed0b6e6dbd7eaa3cbac1699eeb1f0",
+      "runner_sha256": "1b0722762d12fa546222dbcc3e25c0dc2309457dfa727b6edcb8f98a54e83bae",
+      "python": "3.12.3",
+      "command_elapsed": "PT7M33.947S",
+      "operation_seconds": [
+        88.51640437496826,
+        90.26576620602282,
+        86.10522801801562,
+        89.66752364404965,
+        90.85604079504265
+      ],
+      "median_seconds": 89.66752364404965
+    },
+    {
+      "block": 3,
+      "status": "PASS",
+      "physical_fits": 0,
+      "raw_result_sha256": "5a1b7ff7c2f707dc866f6b8ded63f3f7da43708ab1b1b3275e18fe11e52b81cb",
+      "runner_sha256": "e432e55964006acc55232f64f00aebb06ac6fa6a72400366eb6cebd6e7f00cf6",
+      "python": "3.12.3",
+      "command_elapsed": "PT7M34.122S",
+      "operation_seconds": [
+        90.26548215199728,
+        89.95532578800339,
+        85.50048555596732,
+        89.96577449498,
+        90.01619070599554
+      ],
+      "median_seconds": 89.96577449498
+    }
+  ],
+  "proposed_timing_policy": {
+    "atol_seconds": 5.0,
+    "rtol": 0.1,
+    "baseline_block": 1,
+    "baseline_median_seconds": 89.70340646797558,
+    "maximum_candidate_median_seconds": 103.67374711477315,
+    "observed_block_median_range_seconds": [
+      89.66752364404965,
+      89.96577449498
+    ],
+    "rule": "candidate median <= reference median * 1.1 + 5 seconds",
+    "rationale": "The three observed medians differ by less than 0.4 percent. The proposed 10 percent plus 5 seconds is operational headroom, not a measured confidence interval."
+  },
+  "proposed_input_manifest_sha256": "a6f6abd0f27d874c460e3cede732020822f5c1cc814b41e0efddfded63d05fc4",
+  "proposed_reference_sha256": "4b08fee40490cff81f9b8bf465a22aa90aed7d21e7ad67bb435acfaed808e8db",
+  "limits": [
+    "The host provides 192 CPUs. P1 uses two processes with eight sampler threads each.",
+    "Cache state is uncontrolled. These three blocks do not establish a universal timing budget.",
+    "The installed published wheel is 1.0.0. No GitHub dispatch or candidate-wheel comparison ran.",
+    "Raw results and their original input manifest remain unchanged. A separate reference record binds the proposed timing policy to the same data and work.",
+    "A maintainer must review the reference and policy before configuring their checksums for future manual weekly runs."
+  ]
+}