JuliusBrussee/caveman · error

session-value artifact confidence policy invalid

Error message

session-value artifact confidence policy invalid

What it means

Confidence-policy constants are range-checked: QualityUncertaintyZ must be finite and in (0, 5] (the z-score used for conservative quality estimates) and MaxInversePropensity must be finite and in [1, 100] (the importance-weighting cap). NaN, infinities, or out-of-range values would make exploration corrections unbounded or degenerate, so they are rejected.

Source

Thrown at proxy/routing/session_value.go:188

	}
	if artifact.PolicyVersion <= 0 || artifact.RouterVersion != SessionValueRouterVersion || strings.TrimSpace(artifact.EstimatorVersion) == "" {
		return errors.New("session-value artifact version identity invalid")
	}
	if !validCompactPoolHash(artifact.CandidatePoolHash) || artifact.CandidatePoolHash != candidatePoolHash {
		return errors.New("session-value artifact candidate pool mismatch")
	}
	if !validSHA256Ref(artifact.TrainingManifestHash) || strings.TrimSpace(artifact.TrainingExtractor) == "" || strings.TrimSpace(artifact.OutcomeContractVersion) == "" {
		return errors.New("session-value artifact training lineage invalid")
	}
	if artifact.ValidFrom.IsZero() || artifact.ValidUntil.IsZero() || !artifact.ValidUntil.After(artifact.ValidFrom) || now.Before(artifact.ValidFrom) || !now.Before(artifact.ValidUntil) {
		return errors.New("session-value artifact outside validity window")
	}
	if artifact.RollbackParentHash != "" && (!validSHA256Ref(artifact.RollbackParentHash) || artifact.RollbackParentHash == artifact.ArtifactHash) {
		return errors.New("session-value artifact rollback lineage invalid")
	}
	if !finite(artifact.QualityUncertaintyZ) || artifact.QualityUncertaintyZ <= 0 || artifact.QualityUncertaintyZ > 5 ||
		!finite(artifact.MaxInversePropensity) || artifact.MaxInversePropensity < 1 || artifact.MaxInversePropensity > 100 {
		return errors.New("session-value artifact confidence policy invalid")
	}
	featureNames := SessionValueFeatureNames()
	if len(artifact.FeatureSpecs) != len(featureNames) || len(artifact.Actions) == 0 {
		return errors.New("session-value artifact has no features or actions")
	}
	for i, spec := range artifact.FeatureSpecs {
		if spec.Name != featureNames[i] || (i > 0 && artifact.FeatureSpecs[i-1].Name >= spec.Name) {
			return errors.New("session-value artifact feature vocabulary or order invalid")
		}
		if spec.Name == "turn_index" && !spec.Required {
			return errors.New("session-value artifact must require turn_index")
		}
		if !finite(spec.Mean) || !finite(spec.Scale) || spec.Scale <= 0 || !finite(spec.Min) || !finite(spec.Max) || spec.Min < 0 || spec.Max < spec.Min {
			return fmt.Errorf("session-value feature %q bounds invalid", spec.Name)
		}
	}
	seenActions := map[string]struct{}{}
	artifactPool := make([]Candidate, 0, len(artifact.Actions))

View on GitHub (pinned to 766dce6b13)

Solutions

1. Clamp and validate both constants in the trainer: 0 < QualityUncertaintyZ <= 5 and 1 <= MaxInversePropensity <= 100.
2. Fail the training job when any artifact float is non-finite (math.IsNaN / math.IsInf) instead of emitting it.
3. Regenerate and re-seal the artifact after fixing the estimation code.

Example fix

// before
{
  "quality_uncertainty_z": 0,
  "max_inverse_propensity": 1000
}

// after
{
  "quality_uncertainty_z": 1.96,
  "max_inverse_propensity": 20
}

// Before validation: reject non-finite and out-of-range confidence constants.
func finite(v float64) bool { return !math.IsNaN(v) && !math.IsInf(v, 0) }
if !finite(artifact.QualityUncertaintyZ) || artifact.QualityUncertaintyZ <= 0 || artifact.QualityUncertaintyZ > 5 ||
	!finite(artifact.MaxInversePropensity) || artifact.MaxInversePropensity < 1 || artifact.MaxInversePropensity > 100 {
	return errors.New("confidence policy out of range: need 0 < z <= 5 and 1 <= max_inverse_propensity <= 100")
}

func confidencePolicySane(a routing.SessionValuePolicyArtifact) bool {
	return finite(a.QualityUncertaintyZ) && a.QualityUncertaintyZ > 0 && a.QualityUncertaintyZ <= 5 &&
		finite(a.MaxInversePropensity) && a.MaxInversePropensity >= 1 && a.MaxInversePropensity <= 100
}

When it happens

Trigger: The trainer emitting NaN or Inf into the JSON (division by zero during estimation, empty-data covariance); QualityUncertaintyZ set to 0 or negative; MaxInversePropensity set to 0 (weights everything out) or above 100 (explodes estimator variance).

Common situations: Numerical instability in offline training runs; struct defaults never overwritten by the fitting step; downstream JSON handling turning null/garbage into non-finite floats.

Related errors

• session-value artifact version identity invalid
• session-value artifact schema mismatch
• session-value artifact tenant scope mismatch
• session-value artifact candidate pool mismatch
• session-value artifact outside validity window

AI-assisted analysis of JuliusBrussee/caveman@766dce6b13 (2026-08-18). Data as JSON: /api/errors/8370908e17f9e108. Report an issue: GitHub.