Adapt Bayesian¶

adapt_bayesian — mvp.adapt_bayesian

Cluster: ML & Optimisation | Type: component | MCP Tools: 48

Overview¶

Bayesian probabilistic modelling block that computes posterior distributions for Gaussian, Bernoulli, and Poisson models using analytic conjugate priors — no MCMC required. When PyMC is installed it can be extended to full HMC/NUTS sampling; without it the block falls back to pure-Python closed-form conjugate updates with Box-Muller posterior sampling and credible interval calculation. Results include the posterior mean, standard deviation, and a configurable credible interval.

When to use:

• Estimating the posterior mean of a measurement process from observed samples
• Computing credible intervals for A/B test conversion rates (Bernoulli model)
• Modelling event-rate uncertainty in count data (Poisson model)
• Integrating uncertainty estimates into a goal-engine planning loop

Example:

from mvp.adapt_bayesian import AdaptBayesianBlock, BayesianInput

block = AdaptBayesianBlock(name="bayes")
result = block.infer(BayesianInput(
    data=[0.82, 0.79, 0.85, 0.88, 0.76],
    model_type="gaussian",
    prior_mean=0.8,
    prior_std=0.1,
    credible_interval=0.95,
))
# result.value.posterior_mean → ~0.82; result.value.credible_interval_high → upper bound

Works well with: adapt_sklearn, align_evals, adapt_learning

Public API¶

BayesianPriorRecommendation¶

Validated prior + likelihood recommendation decision record.

Methods:

distribution() -> str¶

to_dict() -> dict[str, Any]¶

The advisory payload as elicitation.handle_suggest_prior returns it.

to_metadata() -> dict[str, Any]¶

BayesianDecisionError(ValueError)¶

The LLM did not produce a usable, validated prior recommendation.

LLMBayesianRuntime¶

Provider-neutral prior + likelihood recommendation runtime over G6's LLM caller.

Constructor:

Methods:

recommend(data: list[float], context: str = '') -> Result[dict]¶

BayesianPriorPlanner¶

Runtime-first facade with the deterministic floor as honest fallback.

Constructor:

Methods:

recommend(data: list[float], context: str = '') -> BayesianPriorRecommendation¶

AdaptBayesianBlock(AIBlock[BayesianInput, BayesianOutput, None])¶

Bayesian probabilistic modelling.

Methods:

infer(data: BayesianInput) -> Result[BayesianOutput]¶

BayesianInput(BaseModel)¶

Input to AdaptBayesianBlock.

BayesianOutput(BaseModel)¶

Output from AdaptBayesianBlock.

Methods:

credible_interval() -> tuple[float, float]¶

Backward-compatible combined interval used by older integrations.

AdaptBayesianMCPBlock(AIBlock[MCPBayesInput, MCPBayesOutput, dict])¶

Full-featured Bayesian inference block with SQLite persistence.

Methods:

close() -> None¶

Release resources (DB connection).

infer(data: MCPBayesInput) -> Result[MCPBayesOutput]¶

MCPBayesInput(BaseModel)¶

MCPBayesOutput(BaseModel)¶

BayesianStore¶

Sync SQLite store with 6 tables for Bayesian metadata.

Constructor:

Methods:

close() -> None¶

Close the database connection (idempotent).

upsert_model(name: str, spec: dict, backend: str) -> int¶

get_model(name: str) -> dict[str, Any] | None¶

list_models(limit: int = 50) -> list[dict[str, Any]]¶

delete_model(name: str) -> bool¶

delete_model_cascade(name: str) -> tuple[bool, list[str]]¶

Delete a model and every row that references it, atomically.

count_model_traces(name: str) -> int¶

upsert_trace(name: str, model_name: str, n_samples: int, n_chains: int, path: str, backend: str, summary: dict) -> int¶

get_trace(name: str) -> dict[str, Any] | None¶

list_traces(limit: int = 50) -> list[dict[str, Any]]¶

delete_trace(name: str) -> dict[str, Any] | None¶

upsert_observation(name: str, data: list, shape: list, dtype: str, x_data: list | None = None) -> int¶

get_observation(name: str) -> dict[str, Any] | None¶

list_observations(limit: int = 50) -> list[dict[str, Any]]¶

delete_observation(name: str) -> bool¶

add_experiment(model_name: str, obs_name: str, waic: float, loo: float, backend: str, notes: str) -> int¶

list_experiments(limit: int = 50) -> list[dict[str, Any]]¶

count_all() -> dict[str, int]¶

upsert_ema(name: str, alpha: float, value: float, count: int) -> int¶

get_ema(name: str) -> dict[str, Any] | None¶

delete_ema(name: str) -> bool¶

create_ab_experiment(name: str, variants: list[str]) -> int¶

get_ab_variants(name: str) -> list[dict[str, Any]]¶

record_ab_outcome(name: str, variant: str, success: bool) -> None¶

get_conjugate_state(name: str) -> dict[str, Any] | None¶

upsert_conjugate_state(name: str, conj_type: str, params: dict, n_obs: int) -> int¶

delete_conjugate_state(name: str) -> bool¶

delete_ab_experiment(name: str) -> bool¶

Functions¶

agentic_planner_enabled(default_enabled: bool = True) -> bool¶

Decide whether the agentic bayesian prior-planner should be used.

validate_prior_recommendation(rec: dict) -> dict¶

Reject any prior recommendation outside the canonical shape / bounds.

recommend_prior_floor(data: list[float], context: str = '') -> BayesianPriorRecommendation¶

Deterministic heuristic prior selector (the honest floor).

MCP Tools¶