chore: align Flow dependencies and reward contract compatibility

.github/workflows/ci.yml

@@ -129,7 +129,7 @@ jobs:
 
       - name: Install uv for immutable lock check
         if: needs.scope.outputs.code_changed == 'true'
-        uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9 # v9.0.0
+        uses: astral-sh/setup-uv@20cfd1bf945f4377ade1205e4dbc17946fc9a30d # v10.0.1
         with:
           version: "0.12.5"
           enable-cache: false

README.md

@@ -466,15 +466,14 @@ runtime uses. It returns a signed `RewardEvidenceReceiptV1`: the terminal
 effect landed, or it didn't, or the store couldn't be read and the episode is
 unscored. Unscored is never 0.
 
-The worker is not in a published release yet, and the release carrying it has
-no date. These two commands are what will work once it lands:
+The reward worker ships in version 1.35.1. You can install that release:
 
 ```bash
-pip install 'openadapt-flow[reward]'
+pip install 'openadapt-flow[reward]==1.35.1'
 openadapt-flow serve-reward --seed-mockmed --port 8788
 ```
 
-To run it today, install from the repository head:
+For development, install from the repository head:
 
 ```bash
 git clone https://github.com/OpenAdaptAI/openadapt-flow

docs/REWARD_WORKER.md

@@ -146,16 +146,14 @@ check.
 
 ## The MockMed run
 
-The reward worker is not in a published release yet, and the release
-carrying it has no date. These two commands are what will work once it
-lands:
+The reward worker ships in version 1.35.1. You can install that release:
 
 ```bash
-pip install 'openadapt-flow[reward]'
+pip install 'openadapt-flow[reward]==1.35.1'
 openadapt-flow serve-reward --seed-mockmed --port 8788
 ```
 
-Until then, run it from a checkout of the repository head:
+For development, install from the repository head:
 
 ```bash
 git clone https://github.com/OpenAdaptAI/openadapt-flow

openadapt_flow/reward/calibration.py

@@ -26,6 +26,7 @@
 import math
 import random
 from dataclasses import dataclass
+from decimal import Decimal
 from typing import Any, Callable, Iterable, Mapping, Optional, Sequence
 
 from openadapt_types.process_capability import _digest_payload
@@ -55,6 +56,19 @@ def binomial_cdf(k: int, n: int, p: float) -> float:
     return min(1.0, total)
 
 
+def confidence_delta(confidence: float) -> float:
+    """Complement the declared decimal confidence without subtraction drift.
+
+    JSON records 0.95 as 0.95. Binary subtraction would instead produce
+    0.050000000000000044, which exceeds a strict certificate policy of 0.05.
+    The bound and its certificate must use the same declared probability.
+    """
+
+    if not 0.0 < confidence < 1.0:
+        raise ValueError("confidence must lie in (0, 1)")
+    return float(Decimal("1") - Decimal(str(confidence)))
+
+
 def clopper_pearson_upper(
     failures: int, trials: int, *, confidence: float = 0.95
 ) -> float:
@@ -67,9 +81,7 @@ def clopper_pearson_upper(
         raise ValueError("trials must be positive")
     if not 0 <= failures <= trials:
         raise ValueError("failures must lie in [0, trials]")
-    if not 0.0 < confidence < 1.0:
-        raise ValueError("confidence must lie in (0, 1)")
-    alpha = 1.0 - confidence
+    alpha = confidence_delta(confidence)
     if failures == trials:
         return 1.0
     if failures == 0:

openadapt_flow/reward/seed.py

@@ -39,6 +39,7 @@
 
 from openadapt_flow.reward.calibration import (
     CalibrationResult,
+    confidence_delta,
     corpus_digest_for,
     corpus_from_effects,
     extradup_trials,
@@ -305,7 +306,7 @@ def write_bundle(
             issued_at_policy_update=0,
             expiry_policy_updates=CERTIFICATE_EXPIRY_UPDATES,
             epsilon=calibration.epsilon,
-            delta=1.0 - calibration.confidence,
+            delta=confidence_delta(calibration.confidence),
         )
         _write(directory / CERTIFICATE_FILE, certificate.model_dump(mode="json"))
         _write(directory / CALIBRATION_FILE, calibration.as_metadata())

openadapt_flow/reward/worker.py

@@ -471,7 +471,7 @@ def _issue(
             tier,
             self.certificate,
             episode.policy_update,
-            scoring=self.contract.scoring,
+            contract=self.contract,
         )
         state = certificate_state(self.certificate, episode.policy_update)
         receipt_id = _new_id("reward_receipt")

public-artifacts.json

@@ -93,7 +93,7 @@
     },
     {
       "path": ".github/workflows/ci.yml",
-      "sha256": "46976c2a8daa9a936a4db9f64b72719f4618ac3e29ae02d3cc8035dd61c4b356"
+      "sha256": "50810ca5284b7b108049b18c7b6752d2e7c544dc35f65b3e246ad7fe2f24199a"
     },
     {
       "path": ".github/workflows/citrix-workspace-standin.yml",

pyproject.toml

@@ -44,7 +44,7 @@ dependencies = [
     "cryptography>=42.0",
     # Portable ProcessContract v1 capability, artifact, authentication, and
     # process-receipt contracts. The healthy runtime remains model-neutral.
-    "openadapt-types>=0.17.0,<0.18.0",
+    "openadapt-types>=0.18.0,<0.19.0",
 ]
 
 [project.optional-dependencies]
@@ -77,10 +77,10 @@ dev = [
     # scope and issuer), and reward evidence receipt the reward worker signs.
     # A peer must still negotiate each schema; the dependency alone never
     # upgrades an existing decision surface.
-    "openadapt-types>=0.17.0,<0.18.0",
+    "openadapt-types>=0.18.0,<0.19.0",
     # Engineering-hygiene gates (lint+format, type-check, coverage). Pinned to
     # majors so CI and local dev run the same checkers.
-    "ruff==0.16.5",
+    "ruff==0.16.6",
     "mypy>=1.11",
     "pytest-cov>=5.0",
     "jsonschema>=4.21",
@@ -100,15 +100,15 @@ grounding = ["openadapt-grounding>=0.1.0"]
 console = [
     "fastapi>=0.110",
     "uvicorn>=0.29",
-    "openadapt-types>=0.17.0,<0.18.0",
+    "openadapt-types>=0.18.0,<0.19.0",
 ]
 # Reference Execute server: `openadapt-flow serve-execute`. Same public
 # request schema as Cloud Execute, hosted in this process with a local
 # self-signed receipt. Not the hosted control plane.
 execute = [
     "fastapi>=0.110",
     "uvicorn>=0.29",
-    "openadapt-types>=0.17.0,<0.18.0",
+    "openadapt-types>=0.18.0,<0.19.0",
 ]
 # WindowsBackend: HTTP client for the WAA (Windows Agent Arena) server.
 windows = ["requests>=2.31", "pywin32>=312; platform_system == 'Windows'"]
@@ -118,7 +118,7 @@ windows = ["requests>=2.31", "pywin32>=312; platform_system == 'Windows'"]
 reward = [
     "fastapi>=0.110",
     "uvicorn>=0.29",
-    "openadapt-types>=0.17.0,<0.18.0",
+    "openadapt-types>=0.18.0,<0.19.0",
 ]
 # Native macOS window capture/input. Imported lazily; other platforms never
 # install or import these framework bindings.
@@ -162,8 +162,9 @@ service = ["fastapi>=0.110", "uvicorn[standard]>=0.29"]
 service-mlx = [
     "fastapi>=0.110",
     "uvicorn[standard]>=0.29",
-    "mlx-vlm>=0.6.4,<0.7",
-    "transformers>=5.5,<5.15",
+    "mlx-vlm>=0.6.17,<0.7",
+    "transformers>=5.14,<5.17",
+    "jinja2>=3.1,<4",
 ]
 # Cross-platform desktop recording on-ramp: convert an openadapt-capture
 # session into a flow recording (openadapt_flow.adapters.capture). Optional and
@@ -181,7 +182,7 @@ capture = ["openadapt-capture>=1.2.0"]
 # field-exact against the released schemas. The `interop-types` CI job
 # type-checks and tests the boundaries against the real package; each schema is
 # consumed only by an explicitly negotiated peer.
-interop = ["openadapt-types>=0.17.0,<0.18.0"]
+interop = ["openadapt-types>=0.18.0,<0.19.0"]
 
 [project.scripts]
 openadapt-flow = "openadapt_flow.__main__:main"

tests/test_ci_workflow_contract.py

@@ -23,7 +23,7 @@ def test_lint_job_rejects_a_stale_dependency_lock() -> None:
     lint_job = workflow[lint_start:compatibility_start]
 
     assert (
-        "uses: astral-sh/setup-uv@c771a70e6277c0a99b617c7a806ffedaca235ff9" in lint_job
+        "uses: astral-sh/setup-uv@20cfd1bf945f4377ade1205e4dbc17946fc9a30d" in lint_job
     )
     assert 'version: "0.12.5"' in lint_job
     assert "run: uv lock --locked" in lint_job

tests/test_dependency_security.py

@@ -28,8 +28,9 @@ def _locked_version(name: str) -> tuple[int, ...]:
 def test_mlx_research_extra_keeps_transformers_in_patched_range() -> None:
     """Do not reintroduce the three model-loading/Trainer RCE advisories."""
 
-    assert (5, 5) <= _locked_version("transformers") < (5, 13)
-    assert (0, 6, 4) <= _locked_version("mlx-vlm") < (0, 7)
+    assert (5, 14) <= _locked_version("transformers") < (5, 17)
+    assert (0, 6, 17) <= _locked_version("mlx-vlm") < (0, 7)
+    assert (3, 1) <= _locked_version("jinja2") < (4,)
 
 
 def test_transformers_is_confined_to_the_mlx_research_extra() -> None:
@@ -48,7 +49,7 @@ def test_transformers_is_confined_to_the_mlx_research_extra() -> None:
         for dependency in root["optional-dependencies"]["service-mlx"]
     }
     assert "transformers" not in core_names
-    assert {"mlx-vlm", "transformers"} <= mlx_names
+    assert {"mlx-vlm", "transformers", "jinja2"} <= mlx_names
 
 
 def test_core_runtime_excludes_imagehash_scipy_closure() -> None:

tests/test_reward_worker.py

@@ -2,7 +2,9 @@
 
 from __future__ import annotations
 
+import base64
 import json
+import math
 from pathlib import Path
 from typing import Any
 
@@ -24,6 +26,7 @@
 from openadapt_flow.reward.calibration import (  # noqa: E402
     CorpusRecipe,
     clopper_pearson_upper,
+    confidence_delta,
     corpus_from_effects,
     extradup_trials,
 )
@@ -153,7 +156,7 @@ def test_verified_tier2_is_certified(seeded: dict[str, Any]) -> None:
     assert receipt.reward_components == {"terminal_effect": 1.0}
     assert receipt.certificate_state.value == "current"
     assert receipt.calibration_scope is RewardCalibrationScopeV1.SYNTHETIC
-    assert receipt.production_certified is False
+    assert receipt.certification_refusals(worker.contract, worker.certificate) == ()
     assert receipt.reward_contract_digest == worker.contract.digest
     assert envelope["unscored"] is False
     assert envelope["execute_seal"] is False
@@ -182,6 +185,52 @@ def test_verified_tier2_expired_certificate_is_not_certified(
     assert receipt.scalar_reward == 1.0
 
 
+@pytest.mark.parametrize(
+    ("field", "value", "reason"),
+    [
+        ("epsilon", math.nextafter(0.05, 1.0), "epsilon"),
+        ("delta", math.nextafter(0.05, 1.0), "delta"),
+        ("threshold", 0.6, "threshold"),
+        ("expiry_policy_updates", 1001, "expiry"),
+    ],
+)
+def test_current_certificate_must_satisfy_full_policy(
+    seeded: dict[str, Any], field: str, value: float, reason: str
+) -> None:
+    from openadapt_types.process_capability import canonical_json_bytes
+
+    from openadapt_flow.execute.keys import load_or_create_private_key
+
+    path = seeded["tier2"] / CERTIFICATE_FILE
+    payload = json.loads(path.read_text())
+    if field == "expiry_policy_updates":
+        value = payload[field] + 1
+    payload[field] = value
+    unsigned = {
+        k: v
+        for k, v in payload.items()
+        if k not in {"signature", "signature_algorithm"}
+    }
+    key = load_or_create_private_key(seeded["data_dir"])
+    payload["signature"] = base64.b64encode(
+        key.sign(canonical_json_bytes(unsigned))
+    ).decode("ascii")
+    path.write_text(json.dumps(payload))
+    worker = _worker(seeded)
+    receipt = _receipt(_run(worker, seeded, MOCKMED_HONEST_PATIENT, "strict_policy"))
+    assert receipt.reward_outcome is RewardOutcomeV1.VERIFIED
+    assert receipt.scalar_reward == 1.0
+    assert receipt.certificate_state.value == "current"
+    assert receipt.certified is False
+    assert any(
+        reason in refusal
+        for refusal in receipt.certification_refusals(
+            worker.contract, worker.certificate
+        )
+    )
+    assert worker.verify_receipt(receipt)
+
+
 def test_tier0_is_development_only_never_certified(seeded: dict[str, Any]) -> None:
     worker = _worker(seeded, "tier0")
     assert worker.certificate is None
@@ -530,6 +579,9 @@ def test_certificate_bound_is_recomputable(seeded: dict[str, Any]) -> None:
         calibration["calibration_trials"],
         confidence=calibration["calibration_confidence"],
     )
+    bundle = RewardBundle.load(bundle_dir)
+    assert certificate.delta == bundle.contract.certificate_policy.delta == 0.05
+    assert certificate.unmet(bundle.contract.certificate_policy) == ()
     assert certificate.epsilon == recomputed
     assert certificate.epsilon == pytest.approx(
         1.0 - 0.05 ** (1.0 / CALIBRATION_TRIALS)
@@ -558,6 +610,24 @@ def _mockmed_corpus() -> CorpusRecipe:
     )
 
 
+@pytest.mark.parametrize(
+    ("confidence", "delta"), [(0.95, 0.05), (0.99, 0.01), (0.9, 0.1)]
+)
+def test_confidence_delta_preserves_declared_decimal(
+    confidence: float, delta: float
+) -> None:
+    assert confidence_delta(confidence) == delta
+    assert clopper_pearson_upper(0, 20, confidence=confidence) == 1 - delta ** (1 / 20)
+
+
+@pytest.mark.parametrize("confidence", [0.0, 1.0, -0.1, math.nan, math.inf])
+def test_confidence_delta_refuses_invalid_probability(confidence: float) -> None:
+    with pytest.raises(ValueError, match="confidence must lie"):
+        confidence_delta(confidence)
+    with pytest.raises(ValueError, match="confidence must lie"):
+        clopper_pearson_upper(0, 20, confidence=confidence)
+
+
 def test_clopper_pearson_upper_matches_known_values() -> None:
     # 0 of 15 is the bound the openadapt-evals proof run reports.
     assert clopper_pearson_upper(0, 15) == pytest.approx(0.181036, abs=1e-6)