Differentially Private Score Release

director_ai.core.federated_privacy.score_release.DifferentialPrivacyScoreReleaser

DifferentialPrivacyScoreReleaser(*, epsilon: float, sensitivity: float = 1.0, accountant: PrivacyAccountant | None = None, seed: int | None = None, allow_insecure_seed: bool = False)

Laplace mechanism wrapper for public or cross-tenant score release.

Internal guard decisions should continue to use raw scores. This class exists for dashboards, analytics exports, public benchmark summaries, or other disclosure surfaces where releasing the exact score could leak membership information about the underlying retrieval corpus.

epsilon property

epsilon: float

Return the per-release privacy loss parameter.

sensitivity property

sensitivity: float

Return the declared score sensitivity.

release_score

release_score(score: float, *, label: str, tenant_id: str = '', threshold: float | None = None) -> PrivacyScoreRelease

Release a clamped, Laplace-noised coherence score.

director_ai.core.federated_privacy.score_release.PrivacyScoreRelease dataclass

PrivacyScoreRelease(released_score: float, noise: float, epsilon_spent: float, sensitivity: float, mechanism: str, label: str, tenant_id: str = '', threshold: float | None = None, raw_score_included: bool = False)

Tenant-safe differentially private score release record.

approved_at_threshold property

approved_at_threshold: bool | None

Return the threshold decision for the released score, if configured.

to_dict

to_dict() -> dict[str, Any]

Return a tenant-safe payload with the raw score suppressed.

Boundary

DifferentialPrivacyScoreReleaser adds calibrated Laplace noise to coherence scores before they are exposed outside the runtime trust boundary. It is for dashboards, analytics exports, cross-tenant aggregates, and public summaries where an exact score could leak membership information about the retrieval corpus.

It is not used for internal enforcement decisions. Halt and approval logic should continue to use raw scores inside the protected runtime.

from director_ai.core import DifferentialPrivacyScoreReleaser
from director_ai.core.federated_privacy import PrivacyAccountant

accountant = PrivacyAccountant(max_epsilon=5.0)
releaser = DifferentialPrivacyScoreReleaser(
    epsilon=0.25,
    sensitivity=1.0,
    accountant=accountant,
)

release = releaser.release_score(
    score.raw_score,
    label="public-dashboard",
    tenant_id="tenant-a",
    threshold=0.6,
)
publish(release.to_dict())

The release payload omits the raw score and carries mechanism, epsilon, sensitivity, and budget metadata. Use PrivacyAccountant to cap cumulative privacy loss across score disclosures.