The five-case RCA series and the agentic framework that automates it both assume a luxury the real bay rarely affords: data clean enough to grade your own detective work. This monograph is about the other situation — the ordinary one. You have a diagnostic trouble code with a description that may or may not mean what it says, a few thousand lines of the ECU’s C source, a measurement log of uncertain provenance, a folder of dealer and customer symptom reports written in plain English, and a maintenance and repair history that is part signal and part red herring. None of it is fully trustworthy, and each piece is untrustworthy in a different way.
The mistake everyone makes is to pour all of it into one pot and stir. The right procedure treats the unequal trust of the inputs as a first-class part of the method, not an afterthought. This piece lays out that procedure — eight steps — and the two layers the existing series under-specifies: provenance grading (knowing what each source can and cannot tell you before you use it) and Bayesian evidence fusion (combining sources by reliability-weighted likelihood, counting independent corroboration only, never double-counting a single cause echoed across ten complaints).
It is the natural methodological companion to The Anatomy of a Root Cause (the worked battery case), Fault Detection and Isolation (the detection layer), and the statistical-methods series (the techniques that turn measurements into evidence). Where those cover technique, this one covers epistemics — how to reason honestly when your evidence disagrees with itself.
What it covers
Twelve sections, about eighteen minutes of careful reading.
§ 1–2 — The evidence pile and the trap. Why a real investigation is a heap of unequal sources. Why averaging them, or trusting the loudest, fails. The principle that organizes everything else: grade before you fuse.
§ 3 — Provenance grading. A trust table for the six input families — DTC, C source, telemetry, dealer/customer testimony, maintenance history, repair history — and what each can and cannot establish. The reliability weight assigned to each.
§ 4 — Anchor to one timeline and one cohort. Why most “clues” only become evidence once they are placed in time and population. Defining affected vs. unaffected before interpreting anything.
§ 5 — Frame as abduction. Hypotheses written as causal chains — root cause → mechanism → observable signature — not as correlations. If a hypothesis predicts no signature, it cannot be tested.
§ 6 — The impostor. Correlation that is a co-effect of the true cause. The partial-correlation gate that breaks it. Why the most convincing feature in the room is the prime suspect for being a fake.
§ 7 — The referee. Physics and code as the confirmers. Tracing the C path to prove the suspect logic actually executes under the recorded conditions. The simulator-of-record that must regenerate the observed signature.
§ 8 — Fuse by weight, not by vote. The Bayesian fusion loop: each datum updates the posterior over hypotheses, scaled by trust tier and likelihood ratio, with independence enforced so a single root cause echoed across many reports counts once.
§ 9 — Accept on convergence. The triangulation test. The root cause is the one mechanism that simultaneously explains the code pattern, the measurement signature, the executable path, and the population selectivity — without contradiction from the testimony.
§ 10 — Report the mechanism with residual uncertainty. Output is a causal chain, a confidence, and the single disconfirming test that would still overturn it — then the handoff to prognosis.
§ 11 — Failure modes & traps. Testimony cascades, double-counting, trusting the DTC’s prose, confirming a code path that cannot execute, confound from the repair history itself.
§ 12 — The procedure as a checklist. The whole method on one page, runnable against any investigation.
Part II · §§ 13–19 — The method in action. A full worked battery case with downloadable mockup inputs — a DTC and freeze-frame, the ECU’s C source carrying a signed/unsigned cold-guard bug, fleet telemetry, customer complaints, and service history. The procedure is run end to end: a master flowchart and the method as an algorithm, the evidence graded, a timeline and cohort plot, four hypothesis chains, the impostor broken by partial correlation, the bug confirmed by tracing the C path to the freeze-frame and regenerating the signature, trust-weighted fusion to a posterior, convergence on a single cause, and a remaining-useful-life forecast with and without the fix. Ten figures, two algorithms, and a downloadable input bundle.
§§ 20–21 — Practical notes & references. Eight field rules of thumb for working with unequal-trust evidence, and a grouped further-reading list spanning abduction, causal inference and confounding, Bayesian evidence fusion, model-based diagnosis, measurement and provenance, and root-cause-analysis practice.
Read it
The monograph lives at its own URL with the warm-paper editorial layout of the RCA series — rust accent, teal for confirmed evidence, slate for neutral text — with a provenance-grading table, evidence-fusion diagrams, the Bayesian fusion loop written out as an algorithm, and a full worked battery case (Part II) built on downloadable mockup inputs.
Mockup inputs
↓ rca-uncertain-evidence-mockup.zip — the fabricated inputs for the Part II walkthrough: dtc_freeze_frame.json, bms_balance.c (the cold-guard bug), fleet_telemetry.csv, customer_complaints.md, service_history.csv, and a README with the ground truth so the analysis can grade itself.
Where it sits
The diagnosis series, in order:
I · The core method
- Root-Cause Analysis Under Uncertain, Heterogeneous Evidence — (this post)
- When Is the Evidence Enough?
- Harder Cases in Diagnosis
II · At fleet scale
III · Proof, foundations & prevention
- Proving Cause with Experiments
- Diagnosis on the Vehicle
- The Data Plumbing Behind Fleet Diagnosis
- From Root Cause to No Cause
Full map: The Diagnosis Series index.
← Back to Autonomy