Overview
COGANT (Codebase-to-GNN Translation) deterministically converts software repositories into structured Active Inference artifacts expressed in the Active Inference Institute's Generalized Notation Notation (GNN). It is an evidence compiler: it propagates reviewable program facts through a finite fixpoint rule pipeline and emits graph, matrix, provenance, visualization, and round-trip artifacts with confidence and provenance, rather than a single opaque embedding. A reverse synthesizer reconstructs a runnable Python package from an emitted GNN bundle, closing a forward-reverse-forward evaluation loop. This is the first public release (v0.6.0). Source: https://github.com/ActiveInferenceInstitute/COGANT
Overview source: Curated paper metadata.
Methods and contributions
Read the source for the full argument, qualifications, and evidence.
Findings and contributions
- On the v0.6.0 roundtrip ledger, all 25 targets are role-preserved, but only 1 of 25 meets strict structural isomorphism.
- The author cautions that fixtures are in-sample, with no held-out split or confidence intervals, so scores upper-bound rather than estimate out-of-sample performance.
- The fixpoint ablation shows a single pass suffices on every shipped fixture, with the K=10 cap serving as a safety valve.
- Rule-family ablation indicates structural rules drive HIDDEN_STATE while semantic rules drive OBSERVATION/ACTION/POLICY/PREFERENCE roles.
- The author states the passing test suite, coverage and type-check gates support reliability and reproducibility but do not by themselves establish semantic adequacy.
Methods
- Program-graph IR with confidence and provenance on nodes and edges — Repositories are parsed (primarily via Python's standard-library ast) into a program graph IR whose nodes and edges carry confidence and provenance.
- Fixpoint translation engine with 22 declarative rules in five families — Rules (structural, semantic, control, behavioural, resilience) map nodes to 7 Active Inference mapping kinds, with priority-and-score conflict resolution.
- A/B/C/D matrix derivation and GNN (Generalized Notation Notation) export — Derives likelihood, transition, preference and prior matrices from the compiled state space and program-graph edges, normalized for the upstream GNN validator.
- Forward-reverse-forward roundtrip over a 25-target regression corpus — A reverse synthesizer rebuilds a Python package from each GNN bundle; role_preservation_score and strict isomorphism are recorded per target.
- Rule-family and fixpoint-iteration ablations on packaged fixtures — Removes each rule family and varies the iteration cap K in {1, 2, 5, 10}, recording changes in SemanticMapping counts on the shipped fixtures.
Summary sources: Paper metadata and evidence · Extracted source text.
Document downloads
PDFs are hosted here; supporting sources may use pinned archives.
- COGANT-0.6.0.pdf PDF · 3.49 MiB · extracted-text source
Citation
Citation metadata follows the unified bibliography.
Catalog details and resources
Related works
Ranked by recorded methods, topics, contributors and associated software, with contributor roles and domain-only matches labelled. Metadata overlap is not a citation or a scientific equivalence.