An agent over the public MetroPT-3 dataset: seven months of the air production unit on a Metro do Porto train, with four air leaks dated by the operator. A historian, an event log, a topology graph and a note base, served to an LLM as MCP tools.

The layer that dataset supports and most public datasets do not is reliability: four recurrent events on one repairable unit, on an operating-time clock, with the right censoring carried explicitly. MTTR is not estimable because one report has no maintenance record and another dates the repair a month before its own failure.

Two data findings decide those numbers before any model does. The logging gaps are the machine off rather than lost telemetry, which moves the failure rate by 22%. And ten acquisition freezes, where five analogue channels hold a single value each while a digital channel toggles on a fixed 40 s square wave, fabricate 63 load cycles per hour and outrank every reported failure until they are masked.

Life data analysis, reliability growth, accelerated testing, RBD, FMEA and fault trees, built against Abernethy's The New Weibull Handbook.

Over the usual 105 °C → 40 °C extrapolation the rule embeds Ea = 0.707 eV, within 4% of the only published value for that exact capacitor technology. So it is right by coincidence, not by construction: nothing in it stops you carrying it to a neighbouring technology, where the same error runs from 2.40× optimistic to 0.13× conservative.

Three competing manufacturers confirm the bound in their own technical notes. Closed-form, reproducible with the standard library, and a scope error in an earlier version of the analysis is corrected in the open rather than quietly.

It publishes the sufficient statistics instead of the individual lifetimes, which keeps the underlying data closed while leaving every calculation verifiable: verify.py reproduces every published number with no access to the raw lifetimes and no dependencies.

A neuro-symbolic, glass-box system for multi-document audit. The most expensive error there is a false-positive divergence: a disagreement the system invents by hallucinating an answer the document never gave.

Reported transparently, negative results and limitations included.

Reliability modelling for a PhD thesis on shell-and-tube heat exchangers. OREDA-2015 reports only aggregate exposure time and failure counts, never unit-level histories, so we reconstruct statistically consistent time-to-failure sequences via a Poisson renewal process, select a lifetime model by likelihood-based criteria, and derive RCM maintenance intervals.

Because the shape is an artefact, it is carried as an explicit sensitivity assumption instead of reported as a finding.