# Changelog ## 0.1.0 (2026-11-22) The first release. pg_gaps declares first-order Bayesian networks over ordinary tables with missing values (BayesStore, VLDB 2008) and answers with exact posteriors. Supports PostgreSQL 17, 18 and 19. Models. The catalog lives in ordinary tables in schema `gaps`, dumped with pg_dump and protected by row-level security per owner role. `create_model`, `declare_attribute`, `add_stripe`, `add_stripe_qual`, `add_factor`, `map_factor`, `set_cpt` and `drop_model` declare a model; triggers validate every row (CPT shapes and sums, domains, operators). A model may sit over tables, partitioned tables, materialized views, views or foreign tables. `declare_attribute(..., edges => ...)` bins a numeric column on read. `gaps.validate(model)` proves a model acyclic from its declaration alone (the dependency-graph check of the thesis, sound, not complete), and every grounding checks coverage, mappings and cycles exactly. Inference. `gaps.marginals(model)` returns the posterior distribution of every missing value and `gaps.impute(model)` the most likely joint completion, both exact: the network splits into independent components, each solved by variable elimination over a bucket tree in log space, with identical components solved once. Inference is bounded by `gaps.max_factor_cells` and `gaps.max_total_cells` and cancellable. Given `rel` and `entities`, both ground only the components holding those entities, with the same answers as the whole-model call. A planner support function estimates their rows. Learning. `gaps.learn_cpts(model, alpha)` sets each CPT to its posterior mean from the complete rows, under a symmetric Dirichlet(alpha) or the per-cell pseudocounts of `gaps.set_prior`; `method => 'em'` also learns from rows with missing values, by EM with the engine's exact E-step. `gaps.cpt_rows(model, factor)` shows a CPT as rows. Parameterised and composite models. `add_factor(..., kind => 'logistic' | 'noisy_or', params => ...)` computes a factor's CPT from weights for each grounded instance, and an aggregating parent slot takes any number of parent rows with shared weights, still exactly. `gaps.add_member` builds a composite model whose members ground and solve as one network. Learning keeps parameterised factors fixed; fitting their weights is not yet available. Sampling (experimental). With `gaps.samples` set, a component too large for exact inference is sampled by Gibbs, four chains from dispersed starts, refused when they disagree, seeded by `gaps.seed`. Chains that all settle in the same wrong mode give a wrong answer without an error (doc/DESIGN.md, "Gibbs sampling, as built"). Tested against brute-force enumeration and pgmpy (marginals and EM), under AddressSanitizer and UndefinedBehaviorSanitizer, and through pg_dump and restore; built with clang (Nix) and gcc -Werror (PGDG on Debian and Fedora). Not yet: probabilistic selection, projection and joins (`select_rows`), the CRF half of BayesStore, fitting logistic and noisy-OR weights, EM for components too large for exact inference, and running on a hot standby.