Diagnostics · The Series · Reading Map
Nine connected pieces, from one car’s messy evidence to a fleet-wide campaign — and back around to preventing the fault entirely. A map, and where to start.
One long argument told in ten parts. It begins with a single vehicle and a pile of untrustworthy clues, builds a disciplined method for turning evidence into a cause, asks when you actually know enough to stop, handles the cases that break the easy assumptions, scales the whole thing to a fleet of thousands, makes the fleet methods statistically honest, proves cause with real experiments, pushes detection down onto the vehicle, lays the data foundation underneath it all, and finally closes the loop by engineering the fault out so it never returns, and finally wiring the whole method into a scalable, multi-agent software product. Read it in order, or jump to the part you need.
Three movements — the core method, then fleet scale, then proof, foundations, and prevention — with the loop closing back to the start.
I · The core method
Grade, anchor, and fuse messy, unequal-trust clues into one defensible cause — with a worked battery case and downloadable mockup inputs.
The sufficiency gate, value of information, the sequential stopping test, and what to do when trusted sources disagree.
When the comfortable assumptions break: multiple simultaneous faults, intermittent events, and sensors that lie.
II · At fleet scale
One code across a population: de-mix the mixture, let the cohort boundary name the cause, and turn it into a remediation scope.
The numbers under the figures: significance, multiple-comparisons, CUSUM surveillance, Weibull forecasting, verify-the-fix, and phantom patterns.
III · Proof, foundations & prevention
From observation to proof: DAGs, the staged rollout as a randomized experiment, difference-in-differences, and instrumental variables.
Real-time detection at the edge — observers, residuals, false-alarm budgets, certification — feeding the offline fleet pipeline.
The unglamorous foundation: joining records by VIN, aligning clocks, surviving missing data, and the pipeline that assembles the table.
Closing the loop: CAPA/8D, FMEA feedback, control plans, poka-yoke, and the escape points — so the fault is engineered out for good.
IV · Build the system
The architecture of an AI-powered, multi-agent fleet-diagnosis tool — data, detection, agents, tools, memory, and the human in the loop — built to be practical, scalable, and a product.
New to the topic? Read 1 → 2 → 3 for the reasoning discipline, then 4 → 5 for fleets. Already do this work and want the sharp edges? Jump to 5 (rigor), 6 (proof), or 9 (prevention). Building the system rather than doing the analysis? Start with 8 (data) and 7 (edge), then 10 (the whole architecture).