harbor: k-anonymity floor defaults to 5, not a no-op (feedback #2)

min_aggregate_cases defaulted to 1 — a vacuous floor that let an agent read
held-out validation labels one case at a time via single-case aggregate
evals. Default to 5 in both the sidecar policy and the build AgentAccessSpec
so a build that omits it is safe rather than unfloored (shipped builds
already set 5 explicitly).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: varunursekar

vero/src/vero/harbor/build/config.py

@@ -23,7 +23,9 @@ class AgentAccessSpec(EvaluationModel):
     partition: str
     disclosure: DisclosureLevel = DisclosureLevel.AGGREGATE
     expose_case_resources: bool = False
-    min_aggregate_cases: int = Field(default=1, ge=1)
+    # k-anonymity floor for aggregate subset evals; 5 by default so an omitted
+    # value is safe rather than an unfloored (single-case) leak.
+    min_aggregate_cases: int = Field(default=5, ge=1)
     total_runs: int | None = Field(default=None, ge=0)
     total_cases: int | None = Field(default=None, ge=0)
 

vero/src/vero/harbor/sidecar.py

@@ -72,7 +72,10 @@ class SidecarEvaluationPolicy(EvaluationModel):
     disclosure: DisclosureLevel = DisclosureLevel.AGGREGATE
     expose_case_resources: bool = False
     agent_evaluable: bool = True
-    min_aggregate_cases: int = Field(default=1, ge=1)
+    # k-anonymity floor: aggregate subset evals must cover >= this many cases so
+    # a single held-out label can't be read off one case at a time. Default 5
+    # (not 1) so a build that omits it is safe rather than unfloored.
+    min_aggregate_cases: int = Field(default=5, ge=1)
     parameters: dict[str, JsonValue] = Field(default_factory=dict)
     allowed_parameters: list[str] = Field(default_factory=list)
     limits: EvaluationLimits | None = None

vero/tests/test_v05_harbor_build.py

@@ -657,3 +657,8 @@ def test_compiler_uses_published_version_outside_a_source_checkout(
     assert not (output / "environment/vero").exists()
     dockerfile = (output / "environment/Dockerfile").read_text(encoding="utf-8")
     assert "scale-vero[harbor]==0.5.0" in dockerfile
+
+
+def test_agent_access_defaults_to_safe_k_anonymity_floor():
+    # Omitting min_aggregate_cases must yield a real floor (5), not a no-op (1).
+    assert AgentAccessSpec(partition="validation").min_aggregate_cases == 5

vero/tests/test_v05_harbor_sidecar.py

@@ -698,3 +698,12 @@ async def run_as_different_owner(command, cwd=None, timeout=30, env=None):
         )
         == ""
     )
+
+
+def test_sidecar_policy_defaults_to_safe_k_anonymity_floor():
+    from vero.harbor.sidecar import SidecarEvaluationPolicy
+
+    policy = SidecarEvaluationPolicy(
+        backend_id="b", evaluation_set_name="s"
+    )
+    assert policy.min_aggregate_cases == 5