[docxology]

What changed

• gnn/gaussian_agent.gnn.md — Generalized Notation Notation (GNN) specification of the continuous linear-Gaussian agent implemented in active_infer_ants/agent.py: identity dynamics F and readout H, process noise Q = I / prior_precision, sensory noise R = I / sensory_precision, the preferred-state prior, and the precision-weighted posterior update that infer() performs. ModelParameters records num_dims, prior_precision, sensory_precision and learning_rate.
• gnn/upstream.json — the pinned GNN revision: v5.0.3, 8f293d9ba810e6eda5c09842d9c7ed9ffa5f92c1.
• gnn_conformance.py — extracts the specification from that pinned commit via git archive (never from the GNN checkout's working tree) and checks the extracted model against the package agent: dimensions, identity F/H, Q/R/prior, and that infer() equals the declared Gaussian posterior with the action set to the posterior mean. Continuous models need GNN's runtime dependencies, so the harness resolves a GNN-capable interpreter (GNN_PYTHON, then the checkout's .venv) and reports a clear unavailable state when there is none.
• scripts/validate_gnn_spec.py — CLI gate: exit 0 pass, 1 conformance failure, 2 GNN unavailable.
• tests/test_gnn_conformance.py — the structural half (pin record, all required sections, declared dimensions against the agent) always runs; the pinned-parser half skips when the GNN checkout is absent.
• README.md, SPEC.md, AGENTS.md — document the specification, the gate, and the related repositories (Generalized_Notation_Notation, ants, Active_Skillference).

Verification on this branch

Command	Result
Compile sweep over tracked sources (python -m py_compile, the CI step)	clean
python -m pytest -q	41 passed
python -m pytest tests/test_gnn_conformance.py -q	10 passed
ruff check gnn_conformance.py scripts/validate_gnn_spec.py tests/test_gnn_conformance.py	pass
python scripts/validate_gnn_spec.py	exit 0 — model_kind continuous, 3/3 dimensions, 4/4 checks

The gate reports the extracted specification as continuous with num_states = num_observations = 3 and checks: dimensions match the package agent; F and H are identity; Q, R, prior_mean and prior_cov match the agent's Gaussian model; infer() performs the declared posterior update. Syntax was also checked with CPython 3.10 and 3.11 to match the CI matrix.