Active Assembly Theory and Practice

Assembly traces can support explicit probabilistic inference without being treated as recovered program meaning. Active Assembly Theory and Practice makes that distinction executable: transparent x86-64 fixtures are decoded with Capstone, traced under bounded Unicorn emulation, and mapped to typed observations that distinguish missing, unseen, and unsupported evidence. Two Generalized Notation Notation (GNN) specifications define passive-phase and prospective sensor models whose tensors drive real pymdp execution. A native categorical Julia/RxInfer companion consumes the passive model's same parameters (Friedman 2026; Heins et al. 2022; Bagaev et al. 2023). The sensor Agent selects an action before fixture evidence is acquired. Independent finite calculations check beliefs and policy scores; native smoothing uses the same parameters and first-observation convention. Every inference comparator and fitted diagnostic also consumes an upstream-parsed numeric GNN specification, with source-bound runtime instances and rendered model graphs.

A corpus of 10 authored fixtures supports complete-fixture exclusion diagnostics. At the fixed informative likelihood, the memoryless decoder outperforms the temporal filter; temporal advantages appear under synthetic observation ambiguity. Designed category ambiguity remains a visible failure case. The separate Gaussian and categorical Bethe free-energy diagnostics are not compared numerically with policy expected free energy.

The quotient benchmark contains 64 transformation-family identifiers and 4096 rows, with 16 families held out from model fitting. The co-primary portfolio includes an interpretable classifier, a temporal probabilistic model, a fixed one-layer tanh recurrent classifier, and surface baselines; the temporal endpoint reaches balanced accuracy 0.775 and macro-F1 0.772 on this frozen bounded benchmark. Held-out identifiers share authored template lineages with the fitting data. Family-resampling intervals and conditional randomization diagnostics therefore describe this finite inventory; the retrospective protocol correction withholds superiority claims. A complementary pinned static-only corpus supplies 5 licensed source projects and 100 source-contract rows, including 40 lexical witnesses, 40 contract-directed near misses, and 20 identity controls. The corpus establishes provenance and coverage; it does not execute arbitrary external binaries or support application-effectiveness claims.

Selected normalized machine-state relations are checked by Python, native execution, and a bounded Lean fragment. Exact finite sensing certificates and source-bound reader artifacts make assumptions and failures inspectable. Independent physical replication remains unearned. The contribution is a reproducible substrate for testing assembly-to-inference mechanisms, with explicit limits on semantics, statistical generalization, and external validity.