Leg B — availability and the beam-trip budget

Leg B is a Monte Carlo over a reliability block diagram: the machine is a series of repairable units, each with an MTBF, an MTTR distribution (exponential, lognormal or fixed), optionally n_parallel identical units of which k_required must be up, and a fault class that says how an event ends. Every unit is an alternating renewal process (a unit under repair cannot fail again until it is back — the model behind the closed form MTBF / (MTBF + MTTR), which the Monte Carlo is pinned against).

Inputs

blocks.csv (availability.blocks in the spec, relative to the spec file; the campaign copies it into legs/availability/blocks.csv when it is created and reads that copy from then on — a job exported to another machine carries it):

name,parent,mtbf_h,mttr_h,mttr_dist,n_parallel,k_required,fault_class,scenario_class
RF_STATION,RF,2000,4,exponential,1,1,downtime,S8
SRF_CAVITY_TRIP,RF,60,0.003,fixed,1,1,auto_rephase,S1
ION_SOURCE,FRONT_END,500,1,exponential,2,1,downtime,

scenario_class names the fault-leg class (S1 … S10) whose recovered-by fractions route the block's events (see below); blank = every event ends the way the block's own fault_class says.

Without a table the campaign writes a surrogate template whose values sit in the range of published linac component figures and are labelled as placeholders — replace them with the project's MTBF/MTTR data before reading anything into the numbers.

Fault classes (availability.fault_classes, on top of the defaults): downtime (the block's own repair time), auto_rephase (a fixed 10 s — the RF re-phases itself), operator_retune (lognormal, 10 min mean), degraded (the beam continues at reduced performance: the event is counted, never unavailable). A class may carry its own recovery distribution: {"dist": "fixed" | "exponential" | "lognormal" | "empirical", "value_s" | "mean_s", "sigma_ln", "samples_s"} — empirical draws from a list of measured recovery times.

The class split routes a block's events over classes: the campaign builds it from leg A's recovered-by fractions per scenario class (non-critical → auto_rephase, recovered → operator_retune, every strategy failed → downtime; a critical case that was never attempted counts as unknown and enters no fraction) and applies it to every block whose scenario_class names that class, so the trip budget reflects what the beam dynamics said about each fault. availability.json records the split by class, the split as applied per block, and the blocks that received none; a block table without scenario_class values runs with its own fault_class only.

availability.variants names the SRF-trip rate variants run side by side — srf_fdr (the design MTBF of an SRF trip) and srf_sns (an operating-machine rate) by default — each as an extra auto_rephase block.

Outputs

Every unit starts in its steady state (up with probability MTBF/(MTBF+MTTR), otherwise inside a residual outage), so a finite horizon carries no everything-starts-up bias.

availability.json (per variant: availability mean / p05 / p95 and the closed-form value, trips per year by class, downtime hours and event counts per block, the trip-duration histogram in ESS-style bins — ≤ 10 s, 1 min, 5 min, 20 min, 1 h, 4 h, longer — and the analytic sensitivities dA per e-fold of every block's MTBF and MTTR), the CSVs availability.csv, trip_histogram.csv, sensitivity.csv, and the figures availability_hist.png and trip_budget.png.

The closed forms linac_gen.reliability.availability.analytic_availability (series product with the k-of-n binomial) are what the self-test and the unit tests pin the Monte Carlo against.