bayeswire

The reproducible, agent-verifiable Bayesian workflow — one spec, a reference engine, a fast engine that must agree with it, and a harness that refuses to let either you or the agent skip a step.

Quickstart

You write a model; bayescycle provisions the sampling engine and the plotting tool for you.

Install the workflow harness from PyPI (it pins bayeswire, so nothing else to install yet):

uv tool install bayescycle

Write model.py:

from bayeswire import Data, Observed, Param, model
from bayeswire.constraints import Positive
from bayeswire.distributions import Normal, Truncated

@model
class LinearRegression:
    alpha = Param(Normal(0.0, 1.0))
    beta = Param(Normal(0.0, 1.0))
    sigma = Param(Truncated(Normal(0.0, 1.0), lower=0.0), constraint=Positive())

    x = Data.vector()
    mu = alpha + beta * x
    y = Observed(Normal(mu, sigma))

and data.json:

{"x": [-1.0, -0.5, 0.0, 0.5, 1.0], "y": [-1.6, -0.3, 0.4, 1.1, 2.2]}

Sample, then plot:

bayescycle sample model.py --data data.json -o run/
bayescycle plot trace run/

sample downloads and sha256-verifies the pinned Bayesite engine release into a local cache the first time it runs; later runs reuse it. plot reaches bayesite-viz through uvx at a pinned PyPI version, so there is nothing else to pip install. See packages/bayescycle for the rest of the CLI: prior-predictive checks, simulation/recovery, SBC, and diagnostics.

The toolchain

Four of the five packages below are published from this same monorepo at lockstep versions; only bayesite lives elsewhere.

Every workflow step is a command that reads files and writes files; every run is seeded and replayable; run directories are append-only; provenance records what was actually done. The customer is an agent and the human auditing it; the product is trustworthy process, not a sampler.

bayesjax and bayescycle depend on bayeswire as sibling workspace members at the exact lockstep version; bayesite-viz/bayesite-idata pin it as a version-matched test fixture. bayesite vendors and conformance-tests against the corpus; its copy of the spec is generated and hash-checked, never edited. See the root AGENTS.md and docs/releasing.md.

This repository

The model declaration language and wire format for the bayes* toolchain.

bayeswire owns the language: a declarative Python eDSL for Bayesian models, the resolved-metadata IR it compiles to, the bayeswire_ir v1 wire format that IR serializes as, the dimension-label sidecar format, the normative spec for all of it, and the golden fixture corpus every producer and consumer conformance-tests against.

It contains no inference, no distribution math, no plotting, and no workflow orchestration, and it imports nothing outside the Python standard library.

from bayeswire import Data, Observed, Param, model
from bayeswire.constraints import Positive
from bayeswire.distributions import Normal, Truncated
from bayeswire.ir import canonical_bytes, meta_to_dict
from bayeswire.model import model_meta

@model
class LinearRegression:
    alpha = Param(Normal(0.0, 1.0))
    beta = Param(Normal(0.0, 1.0))
    sigma = Param(Truncated(Normal(0.0, 1.0), lower=0.0), constraint=Positive())

    x = Data.vector()
    mu = alpha + beta * x
    y = Observed(Normal(mu, sigma))

meta = model_meta(LinearRegression)
document = meta_to_dict(meta)          # {"bayeswire_ir": 1, "model": ...}
wire_bytes = canonical_bytes(meta)     # sha256(wire_bytes) is the model hash

Decoding is code-free: meta_from_dict(document) reconstructs resolved metadata without executing any user code, and bindable_from_meta(meta, dimensions=...) returns a pure metadata class that backends bind and sample.

Closed model composition

Use Submodel to reuse a complete, already-validated model under an explicit namespace rather than through Python inheritance:

from bayeswire import Data, Observed, Param, Submodel, model
from bayeswire.constraints import Positive
from bayeswire.distributions import HalfNormal, Normal

@model
class GroupEffects:
    n_groups = Data.scalar()
    mu = Param(Normal(0.0, 1.0))
    sigma = Param(HalfNormal(0.5), constraint=Positive())
    z = Param(Normal(0.0, 1.0), size=n_groups)
    theta = mu + sigma * z
    measurements = Observed(Normal(theta, 1.0))

@model
class Study:
    effects = Submodel(GroupEffects)
    y = Observed(Normal(effects.theta, 1.0))

Composition is closed: GroupEffects owns all of its inputs, so binding Study requires effects.n_groups, effects.measurements, and y. The child's parameters, data, expressions, free values, and likelihood factors are flattened into ordinary resolved metadata with opaque dotted names such as effects.mu; there is no hierarchical IR node or backend recursion. Two Submodel(GroupEffects) declarations create independent namespaces. Child Observed declarations contribute their likelihood factors but, like ordinary Observed declarations, are not expression values.

Layout

• src/bayeswire/ — the package: model, distributions, constraints, math, ir
• ../../spec/ (repo root, toolchain-normative) — the normative wire spec: ir-format-v1.md, generated ir-v1-tags.md, dimension-sidecar-v1.md, data-document-v1.md, model-data-fingerprint-v1.md
• src/bayeswire/corpus/ — golden IR documents, canonical hashes, canonical data documents, fingerprint test vectors, and JAX-oracle evaluation fixtures, shipped as package data so Python consumers conformance-test against the installed pin (see its README)
• scripts/regenerate_corpus.py — regenerates corpus documents, hashes, data documents, fingerprints, and the generated tag spec after a deliberate format change
• ../../docs/releasing.md (repo root) — the release procedure: lockstep version bump, tag, publish, and the two surviving cross-repo edges with bayesite

Development

uv sync
uv run ruff format --check .
uv run ruff check .
uv run ty check
uv run pytest

The test suite includes a no-JAX walk (every module must import with jax and blackjax blocked) and produce-conformance (reference declarations must reproduce the corpus byte-for-byte). See AGENTS.md for the working discipline and docs/invariants.md for the invariants.