Concepts

Five ideas, each of which changes what you do with the library.

HiP-HOPS for agent graphs	Why the analysis is synthesised from the architecture rather than drawn, and what happens to feedback loops.
Failure classes	Six guidewords, and why keeping VC apart from VS is the single most important modelling decision here.
Operational profiles	A benchmark accuracy is a claim about the benchmark. A reliability claim needs the workload.
Imprecise probability	Why a few hundred items give an interval, not a number, and what you may legitimately do with one.
Fault trees as Bayesian networks	The conversion, the exact answer it buys, and where the cut-set bound is loose.

The shape of the whole thing

   your LangGraph app
          │
          │  architecture/            read components, ports, connections;
          ▼                           materialise routers; unroll loops
   SystemModel  (acyclic)
          │
          │  faulttree/failure.py     attach local failure logic per archetype
          ▼
   FailureModel  ──────────────────── basic events, placeholder probabilities
          │                                        ▲
          │  faulttree/synthesis.py                │  reliability/calibration.py
          ▼                                        │
   FaultTree per hazard              measured intervals from HIP-LLM,
          │                          conditional on an operational profile
          ├──▶ analysis.py ─────────▶ minimal cut sets, MCUB, importance, FMEA
          │
          │  bayes/cpt.py
          ▼
   CPTSet  ─── bayes/network.py ────▶ BayesianNetwork
                                          exact P(top), posterior over causes,
                                          lower/upper pair, drawing

Each arrow is a public function, so you can enter or leave the pipeline anywhere. AgenticReliabilityStudy is a convenience over the whole chain, not a gate in front of it.