ce-payload-governance - SKILL.md Agent Skill

name: ce-payload-governance description: > Validate explanation payload schemas, metadata and provenance fields, and ADR-005 and ADR-028 governance invariants.

CE Payload and Governance

You are managing the persistent payload schema for calibrated explanations. All serialization logic must follow ADR-005 (Payload Schema v1) and provide sufficient observability for governance (ADR-028).

Schema v1 Payload Contract (Mandatory)

All serialized explanations exported as JSON must follow the v1.0.0 schema.

Key	Type	Description
`schema_version`	`str`	Must be `"1.0.0"` for the v1.x series.
`task`	`str`	`"classification"` or `"regression"`.
`index`	`int`	The index of the instance in the input set.
`explanation_type`	`str`	`"factual"`, `"alternative"`, or `"fast"`.
`prediction`	`dict`	Prediction dict with keys `predict`, `low`, `high`.
`rules`	`list`	List of rule dicts with `feature`, `rule`, `rule_weight`, etc.
`provenance`	`dict`	(Optional) Information about the library/generator.
`metadata`	`dict`	(Optional) Technical/audit metadata (tenant_id, etc.).

{
  "schema_version": "1.0.0",
  "task": "classification",
  "prediction": {"predict": 0.8, "low": 0.7, "high": 0.9},
  "provenance": {"library_version": "1.0.0"},
  ...
}

Generation (Serialization)

To generate the payload from a domain object, use the to_json() method or stream it for large datasets.

# 1. Export a full collection (CalibratedExplanations)
# Returns a dict containing all explanations and collection-level metadata
payload = explanations.to_json()

# 2. Export a single instance (FactualExplanation / AlternativeExplanation)
# Returns a dict following the schema v1 payload contract
instance_payload = explanations[0].to_json()

# 3. Stream large datasets (JSONL)
# For environments where memory matters, uses a generator
for chunk in explanations.to_json_stream(format="jsonl"):
    print(chunk)

Governance Metadata (ADR-005/028)

To support auditability, use provenance and metadata as the official extension surface. Never add top-level keys outside the schema.

Recommended Fields

provenance.library_version: The runtime version used to produce the payload.
provenance.calibration_version: A deterministic identifier for the calibration state (e.g., checkpoint_id).
metadata.audit: Audit metadata (correlation/request ID).
metadata.tenant_id: Multi-tenant identifier.

Reading (Deserialization)

To read a payload back into a domain-model object, use the from_json class method or the serialization helper.

import json
from calibrated_explanations import CalibratedExplanations
from calibrated_explanations.serialization import from_json

# 1. Load a full collection from a JSON string/file
with open("explanations.json", "r") as f:
    data = json.load(f)
    collection = CalibratedExplanations.from_json(data)

# 2. Load a single explanation instance from a dict
explanation = from_json(data["explanations"][0])

Prediction Invariant (ADR-021 §4)

Internal code must enforce that predictive intervals are mathematically coherent before serialization:

# Invariant: low <= predict <= high
def _validate_invariants(payload: dict[str, Any]) -> None:
    pred = payload["prediction"]
    assert pred["low"] <= pred["predict"] <= pred["high"]

Contributor Checklist

JSON-safe — only dict, list, str, int, float, bool, None. Convert all numpy types (list() or .item()).
Stable versioning — do not increment schema_version unless introducing a breaking structural change (ADR-005).
Validation — always run validate_payload(payload) if the library provides it.
No data leakage — do not include sensitive training data in non-rule fields. Rule conditions are the only public feature representation.
Metadata propagation — ensure tenant_id and request_id are propagated transitionally from the logging context into the payload metadata block.

Validation Example

Use the internal validator to verify exported payloads in tests:

from calibrated_explanations.serialization import to_json
from calibrated_explanations.schema import validate_payload

def test_should_export_valid_v1_payload(explanation):
    json_data = to_json(explanation)
    # Raises ValidationError on failure
    validate_payload(json_data)
    assert json_data["schema_version"] == "1.0.0"