blackjax.adaptation.meta.verdict

blackjax.adaptation.meta.verdict#

Post-run verdict extractors for the meta-adaptation controller.

Functions#

extract_meta_verdict() — single-chain verdict from

MetaAdaptationCoreState.

extract_multi_chain_verdict() — multi-chain verdict from

MultiChainMetaAdaptationCoreState.

Functions#

extract_meta_verdict(...)

Build a MetaAdaptationVerdict from the final core state.

extract_multi_chain_verdict(...)

Build a MetaAdaptationVerdict from a multi-chain final core state.

Module Contents#

extract_meta_verdict(final_state: blackjax.adaptation.meta._state.MetaAdaptationCoreState, max_grad_budget: int, num_warmup_steps: int, adaptation_info: Any = None) → blackjax.adaptation.meta._state.MetaAdaptationVerdict[source]#

Build a MetaAdaptationVerdict from the final core state.

Call after warmup.run() completes. Pass adaptation_info (the second return of warmup.run()) for true gradient counts.

budget_returned_steps is ADVISORY: the scan runs its full length in v1; this field shows where a stopping host would have saved steps.

extract_multi_chain_verdict(final_state: blackjax.adaptation.meta._state.MultiChainMetaAdaptationCoreState, max_grad_budget: int, num_warmup_steps: int, adaptation_info: Any = None, *, pooled_draws_by_window: Any = None) → blackjax.adaptation.meta._state.MetaAdaptationVerdict[source]#

Build a MetaAdaptationVerdict from a multi-chain final core state.

Drop-in counterpart of extract_meta_verdict() for the MultiChainMetaAdaptationCoreState produced by build_multi_chain_meta_core().

Parameters:

  • final_state – Final MultiChainMetaAdaptationCoreState after warmup.run().

  • max_grad_budget – Same semantics as extract_meta_verdict().

  • num_warmup_steps – Same semantics as extract_meta_verdict().

  • adaptation_info – Same semantics as extract_meta_verdict().

  • pooled_draws_by_window – Optional per-window pooled draw array exposed for nested-R̂ diagnostics (shape (n_chains, n_per_window, d) per window). Not used internally — passed through as flags["pooled_draws_by_window"] for the evaluation layer.

Returns:

Verdict with multi-chain–specific flags: n_chains, chain_collinearity, unimodality_gate, deferred_to_ensemble, and mode_coverage="multi_chain_certified" when all gates passed.

Return type:

MetaAdaptationVerdict