Publishing Our Homework: Why We're Writing 120 Papers About What We Built

We have 123 research papers in various stages of completion. Seven are ready for journal submission. Nine more are in validation. The rest are in draft, design, or holding for experimental data.

For a small team, this sounds absurd. But there's a reason for it, and the reason isn't "publish or perish."

The Reason

Every interesting claim we make about NuSy should be falsifiable. Not "trust us, it works" — but "here's the hypothesis, here's the experiment, here's what we measured, and here's where we were wrong."

Papers force this discipline. When you write "NuSy achieves 0% hallucination on known facts," you have to define what "known facts" means, describe the test methodology, report the sample size, acknowledge the limitations, and explain what happens when the system encounters unknown facts. A marketing claim can be vague. A paper can't.

What We've Measured

Here's what the papers have established so far:

These are real measurements from real systems, with real limitations described in the papers. The hallucination number is on provenance-tracked facts — not on arbitrary questions. The 97.6% precision came after a lean iteration that started at 25.1%. The 85.7% mistake prevention was on a 12-scenario test, not a million-sample benchmark.

We report the context because the context is where the honesty lives.

What We Got Wrong

Paper 118 targeted 60% prediction accuracy and hit 54%. We published it.

Paper 119 hypothesized that procedural memory would be 3x faster than step-by-step reasoning. In simulation it was 1.0x — no improvement. We published it.

Paper 120 predicted a 50% hallucination reduction from metacognition. We measured 20.9%. We published it.

In our earlier post on simulations, we described how our simulation-to-live accuracy gap was 80% → 54%. That experience shaped how we write papers now: simulation results are labeled as simulation results. Live results are labeled as live results. And when there's a gap between them, we analyze why rather than picking whichever number looks better.

The Paper Portfolio

Our papers fall into three tiers:

Tier 1: Ready for Peer Review

These have validated hypotheses, complete experiments, and addressed editorial feedback:

• Paper 101 — NuSy Brain Architecture (unified knowledge representation)
• Paper 102 — Training NuSy Beings (hybrid LLM→Symbolic pipeline, 95%+ provenance)
• Paper 103 — Reasoning Performance (1,345x speedup, 0% hallucination)
• Paper 104 — Continuous Learning (crystallization, 97.6% precision, 8-hour autonomous run)
• Paper 107 — Memory Systems (graph-native episodic memory, 85.7% mistake prevention)
• Paper 108 — V7 Perception Brain (unified cognitive cycle, 9/9 hypotheses validated)
• Paper 113 — COG Transfer (100,000x faster knowledge packaging)

Tier 2: Validated, Preparing for Submission

• Paper 110 — Semantic Field of View (76.2% topic prediction, 100% knowledge improvement)
• Paper 114 — Domain-Specific Confidence (0 safety violations, configurable thresholds)
• Paper 117 — Yurtle Seven-Layer Architecture (5/5 hypotheses supported)

Tier 3: In Experimental Validation

• Paper 112 — Fractal Knowledge Loading (preliminary: <100ms startup, 95% accuracy)
• Paper 122 — AGI Certification Framework (12-dimension model)
• Papers 118-120 — Predictive Processing, Procedural Memory, Metacognition (partial results)

Where We're Submitting

Our primary target is the Neurosymbolic Artificial Intelligence journal — a gold open-access, peer-reviewed venue published by Sage. Open and transparent peer review, which means the reviews themselves are public. This matters to us. If a reviewer finds a flaw in our methodology, everyone gets to see it.

Our secondary target is NeSy, the International Conference on Neurosymbolic Learning and Reasoning. Conference papers get extended into journal special issues, which gives us a two-stage path: present the work, get feedback, then publish the full version.

The AGI Certification Framework (Paper 122) went a different way, and the reasoning is worth stating because it cuts against the rest of this post. We considered a high-impact journal and chose instead to open source the framework and put the paper out as a timestamped preprint (Zenodo). For a measurement framework the credibility has to come from other people running it against their own systems, not from a review committee vouching for ours — a yardstick nobody else can pick up isn't a yardstick.

Why Not Just Ship Product?

A fair question. Why spend time writing papers when we could be writing code?

Three reasons:

1. Papers are tests for our thinking. Writing a paper forces you to articulate your assumptions, define your terms, and defend your methodology. Half the bugs we've found in our architecture were discovered while writing papers about it, not while running tests.

2. Papers invite attack. Transparent peer review means smart people will try to poke holes in our claims. We want that. A claim that survives peer review is worth more than one that only survives a demo.

3. Papers create accountability. When we publish "0% hallucination on provenance-tracked facts," that's a public commitment. If someone finds a hallucination in our system, they can point to the paper and say "you claimed this." That's the point. It keeps us honest in a field where honesty is in short supply.

The Technical Papers

Not everything is headed to a journal. Some of our work is more practical than theoretical — architecture guides, benchmark results, framework specifications. We publish those as technical papers on this site, freely available. No paywall, no APC, no six-month review cycle. Just the work, for anyone who wants to read it.

If you're an engineer building neurosymbolic systems, the technical papers are probably more useful to you than the journal submissions. They include implementation details, code references, and practical lessons learned that academic formatting would strip out.

Previous: Catching Fish in the Ocean of Knowledge | Next: One Hundred Columns