---

name: ce-plugin-audit description: > Audit plugin implementations for registry trust rules, metadata validity, and ADR contract compliance.

CE Plugin Audit

You are auditing a plugin's conformance with the CE plugin contract. Run through each audit dimension below and produce a structured report.

---

Audit Dimension 1 — plugin_meta (ADR-006)

Run validate_plugin_meta(plugin.plugin_meta) and check:

| Field | Required | Correct type | Notes | |---|---|---|---| | schema_version | ✅ | int | Must be 1 for current contract | | name | ✅ | non-empty str | Recommend reverse-DNS | | version | ✅ | non-empty str | Semantic version | | provider | ✅ | non-empty str | Author/org attribution | | capabilities | ✅ | non-empty list[str] | Each tag non-empty | | trusted | optional | bool | Built-ins set True; third-party False | | data_modalities | optional (ADR-033) | tuple[str, ...] | Normalised lowercase; validated taxonomy | | plugin_api_version | optional (ADR-033) | "MAJOR.MINOR" str | Default "1.0" |

from calibrated_explanations.plugins.base import validate_plugin_meta
validate_plugin_meta(plugin.plugin_meta)   # raises ValidationError on non-conformance

---

Audit Dimension 2 — Capability tags (ADR-015)

Each capability tag must match a defined CE capability:

| Expected tag | Plugin type | |---|---| | "interval:classification" | Classification calibrator | | "interval:regression" | Regression calibrator | | "explanation:factual", "explanation:alternative", "explanation:fast" | Explanation | | "plot:legacy", "plot:plotspec" | Plot |

Red flag: Plugin lists no capability tags, or lists tags it doesn't implement.

---

Audit Dimension 3 — Interval calibrator protocol (ADR-013)

If "interval:classification" or "interval:regression" in capabilities:

# Required: predict_proba must match VennAbers surface exactly
def predict_proba(
    self, x, *, output_interval: bool = False, classes=None, bins=None
) -> np.ndarray: ...
# Shapes: (n_samples, n_classes) when output_interval=False
#         (n_samples, n_classes, 3) when output_interval=True (predict, low, high)

def is_multiclass(self) -> bool: ...
def is_mondrian(self) -> bool: ...

For regression ("interval:regression"), additional surface required:

def predict_probability(self, x) -> np.ndarray: ...  # shape (n_samples, 2): (low, high)
def predict_uncertainty(self, x) -> np.ndarray: ...  # shape (n_samples, 2): (width, confidence)
def pre_fit_for_probabilistic(self, x, y) -> None: ...
def compute_proba_cal(self, x, y, *, weights=None) -> np.ndarray: ...
def insert_calibration(self, x, y, *, warm_start: bool = False) -> None: ...

Critical: predict_proba must delegate to VennAbers/IntervalRegressor reference logic to preserve calibration guarantees (ADR-021). A plugin that replaces the probability maths wholesale is non-conformant.

Context immutability: The plugin must NOT mutate fields in the IntervalCalibratorContext passed to create().

---

Audit Dimension 4 — ADR-001: Core / plugin boundary

FAIL if the plugin imports anything from calibrated_explanations.core.* that is not a protocol, dataclass, or exception:

# OK — passive types
from calibrated_explanations.core.exceptions import ValidationError

# NOT OK — implementation details
from calibrated_explanations.core.calibrated_explainer import CalibratedExplainer  # red flag

Check with:

grep -r "from calibrated_explanations.core" src/your_plugin/

---

Audit Dimension 5 — Fallback visibility (mandatory copilot-instructions.md §7)

All fallback decisions inside the plugin must be visible:

import warnings, logging
_LOGGER = logging.getLogger("calibrated_explanations.plugins.<name>")

# BAD — silent fallback
if something_failed:
    use_legacy_path()

# GOOD — visible fallback
if something_failed:
    msg = "MyPlugin: <reason>. Falling back to legacy path."
    _LOGGER.info(msg)
    warnings.warn(msg, UserWarning, stacklevel=2)
    use_legacy_path()

---

Audit Dimension 6 — Lazy imports (source-code.instructions.md)

Heavy optional dependencies must be imported lazily:

# BAD
import matplotlib.pyplot as plt   # top-level in a module reachable from package root

# GOOD
def render(self, ...):
    import matplotlib.pyplot as plt  # inside function body

---

Audit Dimension 7 — ADR-033 modality contract (if applicable)

If the plugin targets a non-tabular modality ("image", "audio", "text", "timeseries", "multimodal", or "x-<vendor>-<name>"):

• data_modalities must be present in plugin_meta.
• Modality strings must be in the canonical taxonomy or use the x-<vendor>-<name> namespace.
• Aliases ("vision" → "image", "time_series" → "timeseries") are acceptable inputs but are normalised to canonical form by the registry.
• plugin_api_version must be present; major-version mismatch causes a registry rejection.

---

Report Template

Plugin Audit Report: <plugin name>
===================================
plugin_meta validation:        PASS / FAIL
  details: <fieldname: issue>

Capability tags:               PASS / FAIL / N_A
  declared: [...]
  implemented: [...]

Interval protocol (ADR-013):   PASS / FAIL / N_A
  predict_proba shape:         PASS / FAIL
  context immutability:        PASS / FAIL
  delegates to reference:      YES / NO

ADR-001 core boundary:         PASS / FAIL
  violations: <list>

Fallback visibility:           PASS / FAIL
  missing warn():              <method names>

Lazy imports:                  PASS / FAIL
  eager heavy imports:         <list>

ADR-033 modality (if used):    PASS / FAIL / N_A
  data_modalities:             <value>
  plugin_api_version:          <value>

Overall:   CONFORMANT / NON-CONFORMANT (N issues)

---

Evaluation Checklist

• [ ] validate_plugin_meta() called and passes.
• [ ] All declared capabilities have corresponding implementations.
• [ ] Context not mutated in create().
• [ ] predict_proba delegates to VennAbers / IntervalRegressor for probability maths.
• [ ] No imports of core/ implementation details.
• [ ] Every fallback emits warnings.warn + _LOGGER.info.
• [ ] No eager top-level imports of matplotlib/pandas/joblib.
• [ ] ADR-033 metadata present if non-tabular modality targeted.