We publish the record and the receipts. NuSy's ideas live in a public preprint trail going back to 2025, and every quantitative claim we make in public resolves to an on-disk evaluation you can check. Below: the program, and — because our engine is now closed — exactly how you verify us anyway.
The program
Our research is a continuous, dated public record — timestamped preprints on Zenodo, not a retrospective story.
All papers → — every preprint with its DOI, its headline result, and, where one exists, a plain-English explainer of what it found and what it does not claim.
The arc runs from the early cognitive-architecture work to the current V20 program (a governed, proof-carrying reasoning runtime — the Noesis being).
Foundations (2025 → early 2026). The cognitive architecture and its substrate:
- The NuSy brain — a neurosymbolic architecture (#101), and training neurosymbolic systems (#102).
- Reasoning performance (#103) and learning while reasoning / continuous learning (#104).
- Graph-native episodic memory (#107) and fractal knowledge loading (#112).
- Neurosymbolic perception (#108).
- The Y0–Y6 cognitive-layer architecture (#117), honest failure (#118), and the Yurtle knowledge-representation format.
Certification & the cross-domain yardstick (2026). The AGI Certification Framework (ACF) (#122) — a graph-based framework for measuring general capability across dimensions. ACF is open (below).
The current program (V20). The active direction: a governed reasoning runtime — four planes
over governed state, an authority axis over memory (Candidate < Evidence < VerifiedProof), and
refusal as a first-class outcome. The near-term papers (the "Graph Governance" position paper and the
causal-reasoning-behind-a-proof-gate work) are in preparation; we publish them when the evidence is
on disk, not when a trend peaks.
Every claim wears a label. Across this site we mark each claim Proven (measured at a stated scope, regenerable from a test), Nascent (real but early or narrow — growing, never "comprehensive"), or Demo (a showcase, not a deployed integration). The preprint record above is the published, timestamped trail; our claims ledger is the internal gate that decides which numbers reach a page at all — it is the discipline behind what we say, not a place to look numbers up. The reader-facing route to checking us is How to Verify an AI's Claims.
How you verify us — even though the engine is closed
We went maximum-closed: the reasoning core, the never-launder gate, and the proof engine are commercial IP. That changes how you check us — from "read the code" to "check the receipts" — but not whether you can. Verification here is not a slogan; it's a set of things you can actually take us up on today.
- Signed evaluation artifacts. Every public number resolves to an on-disk eval. Ask, and we hand you the artifact behind a figure — the number, its as-of date, and the run-command + commit that produced it — so you can see exactly how it was measured and, under the audit path below, confirm it against the system. (This is the standing discipline behind the claims ledger.)
- Open specifications — check the formats and the method, not the engine. We publish the formats and methods as implementation-free, openly-licensed specs (CC-BY-4.0): the conformance-bundle shape, the evidence-envelope type, the kanban/format specs, and the two methods — Hypothesis-Driven Development and the honesty-dashboard / claims-ledger discipline. Each carries a Conformance section you can falsify by construction. (See How to Verify an AI's Claims for the index and the method.)
- NDA audit access. For the parts that stay closed, a scoped, NDA-covered audit path is available to serious partners — a look behind the boundary under terms, rather than a black box or an open repo.
The point of maximum-closed done honestly: a company that publishes the ruler it measures itself with, and hands you the receipts on request, is one you can check without reading its source.
Open, and installable today
Not everything is closed. What's open is genuinely open — installable, cited, licensed:
- ACF — the AGI Certification Framework.
pip install acf-framework(v1.1.1, live on PyPI). Archived on Zenodo under the concept DOI 10.5281/zenodo.21651775, which always resolves to the latest deposited release. The cross-domain yardstick, open on purpose: whoever owns the measure shapes the definitions. - noesis-ship. The NATS-based multi-agent communication platform (Python core + pluggable adapters).
- The Yurtle lineage. The markdown-carrying-RDF format and its tooling — historic, paper-cited, public.
(Our earlier engine mirrors were archived when we went closed — we say so plainly on the open-source page rather than leave stale "install this" instructions pointing at retired code.)
Verification, not trust. The preprint trail is the public record; our internal claims ledger gates every number to its evidence before it ships; the open specs let you check the method itself. As of 2026-08-17. Every figure we publish traces to an on-disk evaluation (our guardrail: no number ships that doesn't).