This is the post the five-case series was building toward. The cases established the method — abduction disciplined by partial correlation and physics, the refusal of convincing impostors, the cluster-geometry discipline, the difference between a fault-correction and a system-design accommodation. This post is the architecture that automates the method while keeping the human in the loop at every step that matters.
The framework is built around a tight idea: agents are bad at being the analyst, but very good at being the junior analyst with infinite patience. They are excellent at fetching telemetry, running scripts, generating hypotheses, simulating them, and presenting the evidence in a form the senior analyst can adjudicate. They are terrible at being the senior analyst — at committing to a root cause and authorizing the fix. The framework’s design discipline is to keep all of the first-kind work inside the agent layer and all of the second-kind work behind a human gate. The result is a closed-loop system that does the analyst’s day in minutes while leaving the analyst’s decisions exactly where they belong.
This monograph is the architectural reading of the framework. It is the natural companion to the HITL in Agentic AI series (this is the worked-out example of those principles in a specific domain), the Physics Sandbox post (the physics-oracle pattern is the simulator at the bottom of the framework), and the Why Agentic AI Fails field guide (the gate placement is exactly the architecture that prevents the three classic failure modes).
What it covers
Twelve sections, about twenty-six minutes of careful reading.
§ 1–2 — The problem and the principle. Why current RCA workflows are a bottleneck. Why a single super-agent fails at this task. The principle that organizes the rest of the architecture: agents fetch and propose, humans commit.
§ 3–4 — The agent team. Six specialized agents — perceiver, grounder, hypothesizer, simulator, summarizer, gatekeeper — and what each one does. The interfaces between them. The hand-off discipline.
§ 5–6 — The tool surface. What the agents have access to. Telemetry queries (read-only). Simulator runs (read-only). Document retrieval (read-only). The actuator layer (write, but gated). The deterministic tools that the LLM agents wrap.
§ 7 — Memory architecture. Short-term context (the current investigation). Long-term institutional memory (the knowledge graph of past investigations). The retrieval discipline.
§ 8 — The gate placement. Where the human is in the loop. The before-you-commit gate. The before-you-fix gate. The before-you-deploy gate. The audit-after-the-fact mechanism for actions that happen between gates.
§ 9 — The closed loop. How the framework’s output feeds the next investigation. The labeling-becomes-dataset discipline. The case library that grows with use.
§ 10 — Failure modes and mitigations. What can go wrong. Agent hallucinations on telemetry (mitigation: physics-refereed verification). Confidence amplification across agents (mitigation: the gatekeeper’s adversarial role). Specification gaming of the gate (mitigation: the interface design that surfaces optimization pressure).
§ 11 — Where this fits in the literature. The intellectual neighbors. Cross-references to HITL theory, agentic-AI engineering, and the broader RCA / prognostics literature.
§ 12 — A worked walkthrough. The pointer to the framework battery walkthrough, which runs this entire framework on the battery case and shows the actual outputs at each stage.
Read it
The monograph lives at its own URL with a warm-paper architecture-document layout — navy accent for the agentic-architecture motif, color-coded blocks for the six agents, interactive block diagrams throughout.
The five case studies that ground the framework
This architecture would be unmoored without concrete examples. The five cases are the proof — each one in a different root-cause family, each one demonstrating the method this framework automates:
- Battery Degradation — software / firmware
- Preconditioning-Mode NVH — electromagnetic / structural
- Module Weld Resistance — tooling / manufacturing
- HV Isolation Faults — environmental
- Brake Rotor Corrosion — usage / behavior
And the worked walkthrough of the framework running on Case 1.
All-cases source bundle
↓ source.zip — all five RCA Python generators + all five datasets in a single archive. The combined source for the entire series.
← Back to Autonomy