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:

python
1from calibrated_explanations.plugins.base import validate_plugin_meta
2validate_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:

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:

python
1# Required: predict_proba must match VennAbers surface exactly
2def predict_proba(
3    self, x, *, output_interval: bool = False, classes=None, bins=None
4) -> np.ndarray: ...
5# Shapes: (n_samples, n_classes) when output_interval=False
6#         (n_samples, n_classes, 3) when output_interval=True (predict, low, high)
7
8def is_multiclass(self) -> bool: ...
9def is_mondrian(self) -> bool: ...

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

python
1def predict_probability(self, x) -> np.ndarray: ...  # shape (n_samples, 2): (low, high)
2def predict_uncertainty(self, x) -> np.ndarray: ...  # shape (n_samples, 2): (width, confidence)
3def pre_fit_for_probabilistic(self, x, y) -> None: ...
4def compute_proba_cal(self, x, y, *, weights=None) -> np.ndarray: ...
5def 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:

python
1# OK — passive types
2from calibrated_explanations.core.exceptions import ValidationError
3
4# NOT OK — implementation details
5from calibrated_explanations.core.calibrated_explainer import CalibratedExplainer  # red flag

Check with:

bash
1grep -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:

python
1import warnings, logging
2_LOGGER = logging.getLogger("calibrated_explanations.plugins.<name>")
3
4# BAD — silent fallback
5if something_failed:
6    use_legacy_path()
7
8# GOOD — visible fallback
9if something_failed:
10    msg = "MyPlugin: <reason>. Falling back to legacy path."
11    _LOGGER.info(msg)
12    warnings.warn(msg, UserWarning, stacklevel=2)
13    use_legacy_path()

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

Heavy optional dependencies must be imported lazily:

python
1# BAD
2import matplotlib.pyplot as plt   # top-level in a module reachable from package root
3
4# GOOD
5def render(self, ...):
6    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.