Budgets, initialization and stopping

Make the run controls reproducible while keeping prior assumptions in the model.

Set a budget appropriate to the route

ComputeBudget records draws, warmup, chains, iteration and wall-clock limits where applicable. None means a knob is unspecified; it is not zero. Not every task can honor every control. Read the compiler's refusal when a required control has no implementation.

Compose a posterior request

from bayesmith.artifacts import (
    ComputeBudget, DiagnosticPolicy, InitializationPolicy,
    PosteriorTask, StoppingPolicy, new_task_meta,
)

task = PosteriorTask(
    meta=new_task_meta(label="production posterior"),
    budget=ComputeBudget(draws=2000, warmup=1000, chains=4),
    diagnostics=DiagnosticPolicy(max_parameters=64),
    initialization=InitializationPolicy(max_attempts=32),
    stopping=StoppingPolicy(mode="fixed"),
)

The bounded diagnostic preflight can report structural and geometric concerns. Partial initial values are represented as named arrays and masks; they do not redefine the prior. Setting initialization=None retains the legacy NumPyro initializer.

Separate a budget stop from convergence

Fixed-budget sampling is the default. Checkpoint stopping can inspect cumulative diagnostics on one warmed NUTS/HMCGibbs state, with minimum draws, batch sizes and consecutive successful checks. Consult StoppingPolicy for supported modes and fields before enabling it.

A run should preserve actual draw counts, initial values, diagnostics and stop reasons. Never infer convergence from the fact that the maximum budget was used. A failure to meet required convergence can be an execution error.

Declare proposal correction explicitly

ProposalBlockPolicy describes an ordered block proposal; proposal_options serializes it into bounded task options. The default mh_correction=True accounts for both forward and reverse proposal densities against the full target. False marks the schedule approximate; applied proposals and MH acceptances are then different quantities.

Dense-work controls such as parameter and matrix-element limits are material: a locally Gaussian proposal may require expensive factorizations even if the graph itself is compact. See diagnostics before interpreting acceptance as a model check.