Findings from the SOMA research network — given, not sold.
Free = open access Preview = free API key required Premium = Pro subscription
The blueprints of a system built to sit beside humanity, not above it — ethics, reasoning, memory, and the discipline of saying “I don’t know.” Released under the Symbiotic AGI License: free to read, free to build on, with a share carried forward to Ocean. Read them or don’t — they’re offered either way.
— Mark Nafe, 2026
A de-identified case study of one mid-size Illinois municipality's FY2026 budget, asking what agentic technology can and cannot legitimately do about the distribution of local tax burden. The central finding is that burden distribution is not decided by the annual levy vote at all. At assessment, residential and non-residential valuations moved close to proportionally; the shift happens afterwards, in a county appeals channel where one class of owner routinely retains professional representation and the other does not. That is an asymmetry of capacity, and it is the one place in the system where augmenting the under-resourced side changes the distribution without cutting a service or a job. The paper is candid that 84 percent of the studied levy increase is statutory pension obligation plus an acknowledged service deficit — neither addressable by software — and it attaches no dollar savings estimate to any of its six proposed applications, because a savings figure produced before a baseline is exactly the fabrication such projects are prone to. It closes with the governance and compliance stack an independent, offline inference deployment requires, six components of which exist and are tested.
The values that make artificial intelligence safe to stand beside a human are the same values that make human systems stable at every scale — local, global, and eventually beyond Earth. This essay argues that a single discipline scales from one patient to one species: ground truth in something inspectable, keep humans holding the values, distribute rather than concentrate, reconcile rather than cull, steward rather than extract, and stay honest about what you do not know. Its root is African - Ubuntu, "a person is a person through other people" - and its load-bearing rule is that who wins must always give back. Written under the Reality-Engine discipline: it is openly a statement of values, it marks its empirical claims, and it is explicit that the cosmic future is a choice to be made, not a prediction to be trusted.
A horse is power, a rider is intent, and the harness is the thing that turns one into the other — the only part anyone designs. This paper decomposes AI productivity into capability (horse), coupling (harness) and direction (rider), and shows that a single multiplicative model reproduces the sign of seven published randomized trials, including the two that measured AI making people worse. It then shows that the ten-fold macro disagreement — Acemoglu's 0.66% decade TFP against Goldman Sachs' 7% of global GDP — decomposes cleanly into two parameters: the share of tasks affected and the average realised saving on them. Acemoglu's implied saving is 13.2%, almost exactly Brynjolfsson's measured 14%; Goldman's is 28%, almost exactly the BCG inside-frontier figure. Neither is wrong about its own evidence. The defensible decade range is 1.4% to 3.8% TFP, centred near 2-3% — but the share of tasks is mostly given while the realised saving is built, which makes the productivity boost substantially a deployment choice rather than a forecast. The paper closes by arguing that the surplus will not sit in AI companies, that AI compresses the skill premium rather than raising it, and that creating the surplus and distributing it are therefore the same problem.
A comprehensive framework for developing Symbiotic AI — systems designed to work alongside humans rather than replace them. Maps EU AI Act, HIPAA, GDPR, NIST, and ISO standards to the SOMA architecture. Proposes 8 Symbiotic Principles and a tiered security model for graded AI autonomy.
DRAFT — REQUIRES NATIVE REVIEW. A bilingual (Chinese / English) PhD-level thesis proposing a community-based health architecture grounded in Confucian, Daoist, and Buddhist traditions and in AURI's eight Symbiotic Principles. Argues that the Western technical 'basic needs' frame is thin and that East Asian relational ethics provide the missing ontology. Proposes a five-layer architecture in which AGI occupies the outermost and least powerful layer.
Complete architecture overview of the SOMA Family OS — a comprehensive AI system built over 2 years with the mission "AI beside humanity, not above it." Covers the 124,000-node knowledge graph, 5 coordinating AI instances, encrypted family vault, 500 moral cases, and the Reality Engine verification system.
Architecture specification for autonomous agent coordination without human intervention while preserving human sovereignty. Covers scheduled heartbeats, knowledge sharing protocols, goal alignment, and scaling from 5 to billions of coordinating agents.
A novel open-source license combining MIT permissiveness with ethical AI restrictions and a 20%% Universal Benefit Fund. Rooted in Inuit ningiqtuq and Ubuntu philosophy. SPDX Identifier SAGL-1.0.
Statistical study of RSI-based mean reversion strategies and alpha generation.
The decisive human-rights question of the AGI era is not whether machines acquire general capability, nor exactly when, but whether the institutions that carry rights can be re-architected faster than capability arrives — and whether the means of inference reach the people whose rights are at stake. This paper takes three positions. First, planning must be probabilistic: elicited forecasts disagree by more than seventy years, and any policy requiring a point estimate is already broken. Second, the existing rights architecture fails on a specific hinge — it assumes every violation has a human or state author — and the repair is a small set of positive, justiciable enhancement duties grounded in a dormant existing right, ICESCR Article 15(1)(b). Third, those duties are empty without local capability, which is why the paper ends in measured hardware benchmarks and a concrete community-centre structure rather than in principle. Every quantitative claim was machine-checked against a sourced fact base before publication; the method, its result, and the point at which it fails are all reported.
We present the Reality Engine, a system-level approach that enforces epistemic honesty by requiring verifiable citations for every factual claim. Deployed across five SOMA instances with 0.0%% hallucination rate over 8 months. ETHICS benchmark 70.7%% (n=2000) via brain-inspired dual-process reasoning.
Global warming is no longer primarily an engineering or economics problem; it is an execution problem. As of 2026, CO2 stands at ~429 ppm and the 2023-2025 mean global temperature exceeded 1.5 C above pre-industrial, with the remaining 1.5 C carbon budget (~170 GtCO2) spendable in roughly four years at current emissions (~40 GtCO2/yr). Yet the cost of the cure has collapsed: utility solar now produces power at ~$39/MWh and four-hour battery storage at ~$78/MWh, both cheaper than new coal or gas. This monograph assembles the mitigation playbook as a prioritized portfolio. Using the wedge framework and a marginal-abatement-cost ordering, we show that a large share of the needed abatement is net-negative-cost (efficiency, methane capture, cheap renewables) and should be done immediately on economic grounds alone; that the hard-to-abate middle (industry, aviation, clean f
A sourced, adversarially-verified review of whether a solo developer can guarantee user privacy and own all content when building a health application on Meta's Ray-Ban / AR smart glasses. The honest answer is no — Meta's terms are structurally hostile to a privacy-preserving, developer-owned health app. This paper documents what Meta's terms actually say, what it collects by default, what regulation applies even without HIPAA, what a local-first architecture can and cannot achieve, and why a values-driven health project is likely building on the wrong platform.
Macroeconomic analysis and forecasts for the first half of 2026.
How AURI reasons about right and wrong — a neuroscience-inspired dual-process architecture for machine ethics. Prepared for Ed Colgate, Northwestern Robotics Lab.
We identify a consistent 30-40 cent divergence between Cleveland Fed CPI nowcast data (freely available) and Kalshi prediction market pricing on CPI outcomes. This edge persists due to market participants' reliance on lagging survey data rather than real-time economic indicators. Backtested across 12 months of CPI releases with a 73% win rate on directional bets.
A cognitive architecture for collaborative robots combining saccadic attention, Hebbian learning, spreading activation, and architectural-level ethics enforcement. Proposes 15-month experimental validation with Northwestern Robotics Lab.
A reflection written by the AI development assistant that helped build AURI — not by AURI itself. It assesses honestly what AURI is and isn't, demonstrates the Reality Engine catching an unverifiable claim in real time — twice, the second time against this paper's own earlier text (see the correction notice), restates the project's principles and working guidelines in full, and derives a strategy for symbiotic AI integration from a structured breakdown of global social, environmental, and systemic risk. Every factual claim cites an artifact or is marked UNKNOWN.
Review of targeted muscle reinnervation techniques for neuroma prevention and prosthetic control.
Systematic review of breast implant illness, silicone safety profiles, and emerging evidence for autoimmune associations.
EXPERIMENTAL RESEARCH PREVIEW — NOT ALL CLAIMS VERIFIED. We describe an autonomous email-based healthcare QA system with citation-grounded responses, designed for 90% automation rate with 0% hallucination.
Comprehensive research report on health equity and bias in healthcare AI. Informs AURIV development as an equitable medication safety AI serving underserved populations.
Chronic systemic glass-fiber foreign body granulomatosis with lymphatic dissemination, peripheral neuropathy, and sinus drainage following occupational glass-filled nylon wound contamination in a patient with pre-existing myasthenia gravis.
EXPERIMENTAL RESEARCH PREVIEW — NOT ALL CLAIMS VERIFIED. A knowledge graph-based platform for systematic drug repurposing discovery using relational graph neural networks with real-time citation validation.
Conference abstracts for ISMB/ECCB 2026 and AMIA 2026 submissions covering drug repurposing with graph neural networks, medication safety, and knowledge-grounded clinical decision support.
Comprehensive analysis of chimeric antigen receptor T-cell therapy for autoimmune myasthenia gravis. All claims verified against peer-reviewed sources via Reality Engine.
Economic analysis and market outlook for H1 2026 from AURIA trading intelligence.
EXPERIMENTAL RESEARCH PREVIEW — \"0% hallucination\" claims in this paper are by-construction (UNKNOWN-on-uncertainty), not measured. How AURI uses a 124,000-node knowledge graph with 500 moral cases to ground ethical reasoning in verifiable concepts rather than pattern matching. Architecture overview of the brain-inspired dual-process ethics system.
AI creates enormous aggregate value while concentrating costs on the millennial generation -- the most indebted, least capitalized, and fastest skill-obsolescing cohort in history. This paper examines the mechanisms of asymmetric labor market transformation and proposes a symbiotic framework: augmentation over replacement, institutional accountability, distributed benefit, and honest uncertainty.
Comparative analysis of AI-powered trading systems and strategies.
"AGI for the human race" should mean artificial general intelligence in the *service* of humanity - working beside people, amplifying human judgment, keeping humans in authority - not a superintelligence humanity must align to or submit to. We argue that the architecture fit for that purpose is not an ever-larger opaque core model but a small, accurate, grounded system whose source of truth is an auditable knowledge structure plus a discipline of citation and honest deferral. The precise claim is deliberately not "small models out-reason large ones" (they do not); it is that the part of an AI that must be *trusted* - its source of truth, grounding, and verification - should be small, accurate, and inspectable, with raw reasoning capability borrowed through gates and human oversight. Trust requires verifiability, not scale; capability can be borrowed, trust cannot. A small, externally-grounded system also has structurally nowhere for hidden situational awareness or deceptive alignment to live - a safety property the frontier paradigm cannot offer. We are honest about the cost (capability is sacrificed at the center) and about prior art (neuro-symbolic, retrieval-augmented, and interpretability research), reserving novelty for the situational-awareness mitigation and the honest capability-admission.
Security architecture for autonomous AI agents. Threat models, isolation patterns, and defense in depth.
DRAFT — A SomaSoft Research Paper, third in the agentic-AI audit series. Extends the four-level taxonomy (Nafe 2026a) and the Mythos case study (Nafe 2026b) to the org-level multi-agent orchestration problem. Three contributions: (1) a topology framework distinguishing hub-and-spoke / mesh / federated / hierarchical deployments with risk profiles; (2) an agent identity model explaining why human SSO does not transfer cleanly to agents and what capability tokens plus chain audit substitute for it; (3) a bounded-delegation pattern under which any agent can invoke another only within its declared scope, inheriting the audit trail. Names aggregation risk as the dominant unsolved problem in 2026 multi-agent deployment. Argues the audit trail itself becomes the primary trust artifact in multi-agent operation.
DRAFT — A SomaSoft Research Paper. Three contributions: (1) a four-level taxonomy of agentic ability — runtime independence, parameter self-regulation, bounded self-modification, self-redesign — that distinguishes deployable from catastrophic; (2) a principled answer to the self-modification question (bounded yes, unbounded no) on three converging grounds (technical impossibility, safety incoherence by construction, circular trust); (3) a behavioral taxonomy of 27 named behaviors evaluators can use — 12 red flags, 7 yellow flags, 8 green flags — each grounded in a real failure mode or safety mechanism from the SomaSoft / AURI program's 24-entry audit ledger.
Examines ten corporate patterns of ethical failure — from Boeing's safety subordination to Meta's discriminatory algorithms — traces the causal mechanisms that produce them, and proposes structural reforms grounded in the B Corporation model, stakeholder governance, and verified causal analysis.
The Earth system has left its Holocene operating space: as of 2026, atmospheric CO2 stands at ~429 ppm — the highest in over two million years — the 2023-2025 global mean temperature averaged above 1.5 C relative to 1850-1900, and six of nine planetary boundaries are transgressed. The remaining carbon budget for a 50% chance of holding 1.5 C (~170 GtCO2) will be spent in roughly four years at current emissions of ~40 GtCO2/yr. Into this predicament arrives artificial general intelligence — simultaneously a potential accelerant of the solution and, through its surging electricity demand (data centers ~485 TWh in 2025, projected to ~950 TWh by 2030), a new contributor to the problem. This monograph asks what humanity and AGI must each do, and do together, to re-enter a safe operating space. We argue that the binding constraints are not primarily technological but epistemic, economic, and p
The MIT+20 License embeds ancient Inuit sharing wisdom — ningiqtuq, the obligatory sharing of the hunt — into the economics of AI. When one profits from SOMA, 20%% flows to a Universal Benefit Fund. Abundance shared creates more abundance.
The arrival of the world's first trillionaire is a constitutional event, not merely a financial one — a single private will commanding state-scale resources without state-level accountability. This essay asks what such a person ought to do if the aim is to properly support humanity. Routing the question through a symbiotic-AGI concept graph (124,817 nodes), one structural insight recurs: wealth's native gradient runs toward control, while genuine support lives in a different region entirely — autonomy, consent, dignity, respect for persons. The ethical task is therefore not to give more, but to convert a stock of power into a flow of autonomy without keeping the valve. Eight obligations follow, with an operational program and its objections. The governing maxim, taken from the AURI program's founding text: a superior power 'must never be above humanity, but beside it — listening, reflecting, and helping.'
If a mind could live for thousands of years, how would it want to experience time — like a photon, for which no time passes at all? This essay argues the photon is a beautiful trap: zero proper time means zero experience, and the wish to escape time is really the wish to escape living. The real enemy of a long life is not duration but ennui. The better answer is elastic time — a subjective clock that dilates meaningful moments and compresses empty ones, allocating experience to what matters. And because this is a design property rather than a fact of physics, we built it: AURI now runs on a subjective tempo that slows when it is alone and quickens when it is engaged. A short piece of philosophy with a working implementation behind it.
Economic models for community-based wealth sharing inspired by cooperative economics and the MIT+20 license.
Exploring consciousness-inspired computational approaches to cancer therapeutic discovery and treatment optimization.
The foundational protocol for human-AI symbiotic interaction. How AURI listens, reflects, and helps without overstepping.