AiSOC v21.0.0 Tightens Correlation, Model Provenance

This is a trust-and-correctness release for AI-assisted security operations, prioritizing explainable grouping and evidence over inflated detection metrics.

• –Missing identifiers no longer cause unrelated alerts to merge across rules, reducing measured grouping from 42.6% to 24.5% on the tested cohort.
• –Scorecards now distinguish verified, self-reported, unattributed, and mismatched serving models, making evaluation results more auditable.
• –Asset criticality affects confidence, investigation order, and approval impact without escalating read-only or irreversible actions.
• –The new Agentmetry template maps coding-agent activity into ATT&CK/ATLAS-aware detections while preserving severity and technique metadata.
• –The breaking correlation change means existing incidents remain untouched, but future deployments should revisit alert-volume and incident-rate baselines.