Explainable AI (xAI)

Each reasoning step is isolated into an explicit OpenTelemetry span — not raw logs, but structured CoT records with step index, token range, and span parent.

Hierarchical decomposition separates fact retrieval → logical inference → contradiction detection → decision synthesis into independently inspectable layers.

Frontier models structurally struggle to falsify their internal CoTs under adversarial conditions — captured traces are faithful computation records, not post-hoc rationalisations.

Every memory read and write is logged with vector keys, similarity scores, and tier label — producing an immutable record of exactly what context shaped each decision.

Diff snapshots between decision checkpoints surface experience-following loops — where past errors silently replay — before they become compliance incidents.

Regulators demand "Show Me" evidence. Memory telemetry provides the immutable decision lineage that auditors need — not policy documents, but timestamped vector provenance.

Every MCP tool invocation and A2A delegation is recorded with full input/output payloads, selection rationale, and parent trace context — not just success/failure status.

Inter-agent JSON-RPC streams and delegation records are stitched into the parent trajectory graph, preserving cross-agent causality across orchestrator → sub-agent boundaries.

Hallucinated parameters, wrong tool selection, and silent sub-agent failures become detectable anomalies rather than invisible execution black-boxes.

Critical decisions surface the Minimal Explanation Packet (MEP) directly to human reviewers — with memory provenance, CoT summary, and tool call rationale presented at the right abstraction level.

Every approval, override, and correction is timestamped and cryptographically linked to the originating agent span — creating an unbreakable human accountability chain.

EU AI Act Art. 14 mandates human oversight for high-risk systems. HITL records provide the signed evidence artefact that review occurred, who acted, and what changed.

The system structurally perturbs or omits individual CoT steps to verify whether the agent alters its downstream tool choices — confirming that captured rationale is causally linked, not confabulated.

Successful runs tag memory write spans as verified quality benchmarks; failed runs mark their history inputs for deletion — preventing long-term error degradation across multi-turn sessions.

Alternative-path analysis produces a "what would have changed the outcome" record — the artefact regulators and safety teams need to understand model decision boundaries.