Enum WarningCode 

pub enum WarningCode {
    MsumNonNegativity,
    ProbabilityDomainViolation,
    FoldInRecursivePath,
    EceBinningBias,
    UncalibratedLLMLogprobs,
    UncalibratedNeuralPredicate,
    SharedNeuralInputArgument,
    SharedNeuralFeatureValue,
    PositiveComplementCorrelation,
    CrossPredicateCorrelation,
    SharedRetrievalContext,
}

Variants

MsumNonNegativity

ProbabilityDomainViolation

FoldInRecursivePath

Phase B F1: a clause has a recursive IS-ref and a FOLD aggregate but no ALONG clause. Almost always a semantic mistake — FOLD groups by KEY columns, not by path. (Stress Corpus B3.)

EceBinningBias

Phase C C4: VALIDATE METRICS ece was requested; the equal-width binning ECE is biased in the small-sample regime (Kumar et al. NeurIPS 2019). Use DEBIASED_ECE instead for an unbiased estimator. The bare ECE value is still reported.

UncalibratedLLMLogprobs

Phase B G1-lite: a CREATE MODEL declares no CALIBRATION (or CALIBRATION None) AND the xervo_alias heuristically looks like an LLM provider (generate/..., chat/..., llm/...). Raw LLM logprobs are not calibrated probabilities (rollout D-10). Treat as a documentation hint until Xervo exposes calibration_source.

UncalibratedNeuralPredicate

Phase C C4: a rule body invokes a CREATE MODEL whose output is PROB AND which declares no CALIBRATION (or CALIBRATION None). The fitted probability flows into the probabilistic stack (MNOR / MPROD / complement) — without calibration, the downstream aggregates compound the miscalibration. Run a CALIBRATE statement to fit a transform, or explicitly mark the choice with CALIBRATION none to acknowledge the risk (the warning still fires for the explicit-none case to keep the acknowledgement visible — same pattern as Phase A’s FuzzyNotProbabilistic, rollout D-9).

SharedNeuralInputArgument

Phase C F2a: two or more neural-model invocations in the same rule share an INPUT VARIABLE argument (e.g. model_a(s) and model_b(s)). Under independence-by-default composing the probabilities via MNOR/MPROD is likely wrong since both share the random variable s. Suppressed when ALL invocations involved carry the @independent annotation on their CREATE MODEL declaration. Rollout D-8.

SharedNeuralFeatureValue

Phase C F2b: two or more neural-model invocations in the same rule share an equivalent FEATURE VALUE expression (e.g. model_a(s.tier) and model_b(s.tier)). Different from F2a — even when binding variables differ, the feature input is structurally identical so the same correlation concern applies. Suppression by @independent annotation.

PositiveComplementCorrelation

Phase D F3 case 3: a rule body has both a positive IS-ref and an IS NOT (complement) to different rules on the same subject variable. When the positive and negated rules share base facts, the independence assumption that underlies the probabilistic complement / aggregation is violated. This is a structural over-detection (the MVP fires whenever the pattern matches, even if no actual base overlap exists at runtime); a future refinement will gate on runtime support-set intersection.

CrossPredicateCorrelation

Phase D F3 case 2: a rule body has two or more positive IS-refs to different PROB-bearing rules on the same subject variable. The implicit p AND q conjunction assumes independence between p and q, which is wrong when the two rules share base facts. Structural over-detection (the MVP fires whenever the pattern matches, even if no actual support overlap exists at runtime); a future refinement will gate on runtime support-set intersection.

SharedRetrievalContext

Phase D F3 case 4 (F2c): two or more neural-model invocations in the same rule receive retrieval-backed features (similar_to(prop, _) / semantic_match(prop, _)) over the same node property. The two models condition on the same retrieval evidence, so the implicit independence assumption that underlies composition via MNOR/MPROD/etc. is suspect. Suppressed when all involved models carry @independent. Structural over-detection; a future refinement could gate on cosine similarity of the pre-embedded query vectors (queries are constants per apply_model_invocations call).

Trait Implementations

impl Clone for WarningCode

fn clone(&self) -> WarningCode

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for WarningCode

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl PartialEq for WarningCode

fn eq(&self, other: &WarningCode) -> bool

Tests for self and other values to be equal, and is used by ==.

fn ne(&self, other: &Rhs) -> bool

Tests for !=. The default implementation is almost always sufficient, and should not be overridden without very good reason.

impl StructuralPartialEq for WarningCode