NACE Deep Dive

NACE (Non-Axiomatic Causal Explorer) is a pure-Python research agent by Patrick Hammer that extends Berick Cook's AIRIS reinforcement-learning apparatus to harder settings — partial observability, nondeterminism, nonstationarity, and external change — and adds Non-Axiomatic Logic (NAL) frequency/confidence evidence tracking. It is a real, maintained implementation, not a paper sketch. In short: an implementation-backed, AIRIS-derived, NAL-adjacent causal-learning agent — not a NARS reasoner, not an AIRIS clone, and only partially connected to Hyperon (through a MeTTa bridge).

Last verified: 2026-06-04 (at repo HEAD 361fddf, 2025-05-06).

NACE is a real implementation, not a paper sketch — pure-Python at HEAD 361fddf (2025-05-06): the tree lists 13 .py files + 4 .metta files, with no src/ directory and no C++ sources (the Qt support noted in recent commits is via PyQt bindings in gui.py, not C++ Qt). The repo extends Cook's AIRIS apparatus with partial observability, nondeterminism, nonstationarity, external changes, and NAL frequency/confidence evidence tracking. It is distinct from AIRIS (which it extends) and from MeTTa-NARS (which it bridges to, but is not).

Core Mechanisms

The implementation is centered in nace.py:

• Causal-rule learning. Cognitive schematics of the form (precondition, operation) => consequence, learned through direct interaction with the environment (README.md:5-9).
• AIRIS confidence inheritance. nace.py:49 returns airis_score from _Plan; nace.py:189 annotates the score as "AIRIS confidence"; nace.py:198 describes planning that searches for the highest reward, "or absent that, biggest AIRIS uncertainty". This is direct use of an AIRIS-derived signal, not a paraphrase.
• Curiosity-driven exploration. nace.py:69-85 uses airis_score to choose between exploration and exploitation, including a printed CURIOUS behavior label at nace.py:84; nace.py:71-72 applies thresholds (airis_score >= 0.9) and revisit conditions (airis_score == 1.0) for goal-directed action selection.
• NAL frequency/confidence truth-values. Hypothesis truth values use NAL (frequency, confidence) pairs (README.md:5-9 plus the planner). These are paradigm-distinct from PLN's (strength, count) — the same distinction noted for AIRIS.
• AIRIS comparison anchors. nace.py:403, :408, :442 contain explicit comparison-to-classical-AIRIS code paths — confirming NACE positions itself as an extension and comparison study, not a re-implementation.
• MeTTa bridge surface. bridge.py plus spaces.metta / knowledge.metta / tasks.metta / input.metta connect NACE to a MeTTa runtime via the ONA (OpenNARS for Applications) bridge surface — one of the world-interaction paths, not the whole engine.

Implementation Surface

How NACE Relates to Its Neighbors

• vs AIRIS: Inheritance and extension, not identity. NACE explicitly extends AIRIS to partial-observability, nonstationary, and external-change settings, and adds NAL evidence revision. The paradigm distinctions established for AIRIS carry forward: NACE's curiosity is not ECAN attention, its confidence is not a PLN truth value, and its state graph is not an AtomSpace.
• vs MeTTa-NARS: Shared author (patham9) and shared NAL truth-value semantics (frequency/confidence), but different paradigms: NACE is a causal-rule learning agent that uses NAL evidence, whereas MeTTa-NARS is a faithful NAL-1..NAL-5 reasoner ported to MeTTa. NACE bridges to MeTTa-NARS / ONA via bridge.py; it is not itself a NARS implementation.
• vs PLN: Paradigm-distinct. NAL frequency/confidence is not PLN strength/count. Treat the cross-link as a boundary, not an alternative implementation.
• vs ECAN: Paradigm-distinct. NACE's curiosity drives exploration locally via airis_score; ECAN attention is a global economic STI/LTI dynamic. Not the same mechanism.
• vs MOSES, AI-DSL, Concept Blending, Pattern Mining: Adjacent but non-integrated; no cross-imports at HEAD.

What This Card Is Not

• Not a NARS implementation. See MeTTa-NARS for that. NACE uses NAL evidence values but does not implement the NAL-N inference rules.
• Not an AIRIS clone. See AIRIS. NACE extends AIRIS to a strictly larger problem class; the original AIRIS apparatus is referenced and compared, not re-implemented.
• Not C++. The repo is 100% Python at HEAD. Any "C++" / src/main.cpp / NACE.cpp claim is mistaken — it originates from a pre-review reviewer error, corrected on direct inspection of the repository.
• Not part of the Hyperon-core stack. NACE is a standalone Hammer-lineage research engine. Its MeTTa connection is the bridge.py ONA bridge surface, not deep AtomSpace/PLN integration.

Status and Sources

Implementation status: Active, maintained research line. The recent commit theme is Qt/Wayland support (a PyQt GUI). NACE demonstrates AIRIS-derived control on grid-world environments with Non-Axiomatic evidence tracking.

Primary sources:

• Hammer, P. (n.d.). NACE: Non-Axiomatic Causal Explorer. README + repository at patham9/NACE.
• Cook, B. & Hammer, P. (2024). AIRIS: Autonomous Intelligent Reinforcement Inferred Symbolism. KTH/DiVA diva2:1890142. The source paper for the AIRIS apparatus that NACE extends.
• Wang, P. (2013). Non-Axiomatic Logic: A Model of Intelligent Reasoning. World Scientific. The source for the NAL frequency/confidence semantics NACE uses for evidence tracking.

Provenance: the verdicts on this card come from a 2026-05-07 source-code review; the extraction archive is at scripts/archive/non_clustered_haa_pilot/source5_NACE_AIDSL_MOSES_MeTTaNARS/.