Skip to content

rewrite

rewrite

Classes:

Name Description
RiskTolerance

Risk tolerance presets for leakage mass thresholds.
PrivacyGoal

Structured privacy and utility goal for rewrite mode.
EvaluationCriteria

Criteria for privacy leakage evaluation and repair.

RiskTolerance

Bases: str, Enum

Risk tolerance presets for leakage mass thresholds.

Each preset bundles a coherent set of repair and review thresholds:

  • minimal — Tight leakage threshold (0.6), flags for review aggressively. Good for medical, legal, and financial data.
  • low — Default. Moderate leakage threshold (1.0). Good for most privacy-sensitive data.
  • moderate — Relaxed leakage threshold (1.5), lower review bar.
  • high — High leakage threshold (2.0), does not auto-repair individual high-sensitivity leaks.

PrivacyGoal pydantic-model

Bases: BaseModel

Structured privacy and utility goal for rewrite mode.

Fields:

Validators:

protect pydantic-field

What to protect (e.g. direct identifiers, quasi-identifiers).

preserve pydantic-field

What to preserve (e.g. utility, semantic meaning).

to_prompt_string()

Serialize goal into prompt-ready text.

Source code in src/anonymizer/config/rewrite.py 
def to_prompt_string(self) -> str:
    """Serialize goal into prompt-ready text."""
    return f"PROTECT: {self.protect}\nPRESERVE: {self.preserve}"

EvaluationCriteria pydantic-model

Bases: BaseModel

Criteria for privacy leakage evaluation and repair.

risk_tolerance controls the leakage threshold that triggers repair, whether individual high-sensitivity leaks trigger repair, and the thresholds for flagging records for human review. See RiskTolerance for preset descriptions.

max_repair_iterations caps how many repair rounds are attempted (each round = one LLM call per failing row). Set to 0 to disable repair while still producing evaluation metrics.

Fields:

risk_tolerance = RiskTolerance.low pydantic-field

Preset controlling repair and review thresholds.

max_repair_iterations = 3 pydantic-field

Maximum repair rounds. Set to 0 to disable repair.

repair_threshold property

Leakage mass above which a row is sent for repair.

repair_any_high_leak property

Whether any single high-sensitivity leak triggers repair.

flag_utility_below property

Flag for human review if utility score is below this.

flag_leakage_above property

Flag for human review if leakage mass exceeds this.

sensitivity_weights property

Weights for high/medium/low sensitivity levels in leakage mass computation.