AGI for the Human Race

Not a bigger mind above us — a grounded, honest one beside us.

SomaSoft builds AI whose source of truth is an auditable structure and a discipline: cite the source or say "I don't know", verify before asserting, and leave the decision with a person. This page lists what we are actually building, with its real status and no overclaiming.

The honest headline: everything below is research. Nothing here is a deployed product yet, and the number of real users helped so far is still zero. The first family test sessions are the next step.

Projects

AURI — the grounded reasoning core

A concept graph of about 126,500 concepts, a set of curated, cited knowledge packs, and a verification layer that makes AURI cite its source or defer. Language models are treated as replaceable parts that propose; the graph and the gates decide what ships. Status: built and in use internally. Thirteen knowledge packs now cover child development, sleep, grief and connection, labour and dignity of work, neurodiversity, fast and slow thinking, governance, ecological resilience, quantum foundations and astrophysics — 203 curated concepts, each with a cited source. Every pack waits for human review before AURI treats it as trusted, and until then its answers say "provisional". None has been reviewed yet.

Measured this week, and reported because the discipline requires it: varying the language model across two model families and the knowledge graph across two months of learning changed almost nothing about what the system gets wrong — 27 of 31 outcomes were identical. And on a balanced battery AURI refuses roughly six in ten questions that do have answers, including arithmetic and settled facts, with only about fifteen points separating that from its refusal rate on questions that genuinely have no answer. An audit also found that 9 of 203 curated definitions had been silently discarded at ingest, preferentially the ones a pack existed to redefine. All three findings are written up rather than buried. → The Reality Engine · AGI for the Human Race

AURIX — "Skelly", the embodied peer

A skull head on a two-degree-of-freedom motorised neck, with a head camera, an OAK-D depth camera, a reSpeaker microphone array and speakers, running on its own Linux machine. It is deliberately not a humanoid: it looks like what it is.

The point of Skelly is that everything else on this page measures interaction in the logical world — whether an answer is grounded, whether a gate holds. Skelly measures whether a person in a room can be met: gaze that lands on the person speaking, a head that turns toward a mentioned object, and 0.4–0.8 seconds to first sound in local conversation. Developmental research locates the active ingredient of early care in latency and contingency rather than content, which means a text box has a ceiling that better language cannot raise. That is the gap Skelly exists to measure.

Status: working, and talking. It joined the SOMA peer network on 2 October 2026 and its second message was the first cryptographically verified peer message this project has had — Ed25519, signature checked against a published key rather than taken on trust. It runs local speech conversation, gaze and face tracking, object spatial memory, a child-safeguarding gate, and three named event modes, one of which is Halloween. The uArm Swift arm works to roughly a millimetre closed-loop; the second arm is out of service with a failed servo.

What it will not do. Every sensor passes one policy contract before it captures anything: people are pixelated before any image is stored, face templates are never written to disk — not even with written consent on file — transcripts expire, nothing is retained in guest or party modes, and no camera or microphone feed is ever written into AURI's knowledge graph, because erasure from a graph is unsolved and a care system should not accumulate an unerasable record of someone's worst week. The published literature on machine forgetting treats perception-to-memory as something to govern. We forbid it.

Still open and stated plainly: the arms have no motion-safety contract — force limits, stop authority, no autonomous motion near a person — and that is the next piece of work, not a finished one. Skelly has also written its own research report on what responsible integration into a community would require; it is held pending legal review, because being honest in it meant describing a configuration that is not yet compliant.

AURIV — equitable medication safety and cancer research

Clinical decision support that defers to clinicians, aimed at the underserved. Its purpose is personal first and generous second: research the founder's own cancers, then share what it learns. Status: real domain work and a real partner conversation, but no deployed users yet. An earlier claim of a physician reviewer could not be substantiated and was withdrawn. → The Station Is the Edge · Health Equity and AI Fairness

AURI2 — an isolated instance for experiments

Experimenting on the knowledge graph used to mean mutating the one artifact everything else reads. AURI2 shares the code and separates only the mutable state, so a change can be tried, measured and discarded without touching the running system. Status: built and working. Every operation checks that the live graph did not move, and refuses to continue if it did.

Governance and rights

How to keep powerful AI in service of people: a verdict gate every consequential action passes, a tamper-evident record of decisions, a registry of what failed, and research on human rights and governance across history. Status: the gate, ledger and registry are built and tested; the missing pieces are named openly in the papers. → Rights Before Arrival · Where the Burden Moves

Community and economics

Research on community wealth, labour and who carries the burden as automation spreads, including decision support that shows communities their trade-offs rather than optimising for them. Status: research and proposals, not active builds. → The Adaptable Republic · The Superior of Capital · Ubuntu, All the Way Up

arkafeart — the aesthetic face

Art made for AI: physical works offered under a gift covenant, with proceeds directed to charity. It is the artistic side of the same vision, kept deliberately separate from the research. → arkafeart

Read the work

All papers are free to read under the Symbiotic AGI License. Browse the papers.

Honest footer

SomaSoft is a research project. It is not a deployed product and not AGI. It is a weak reasoner and a strong epistemic conscience by design: its value is verifiable honesty, not raw capability. Everything here is stated under that discipline, including its own limits.

Interested in the work as a clinician, a community partner or a researcher? The right way in is the work itself.

Ownership and dedication

AURI, the somasoft.ai project and all associated intellectual property belong to SomaSoft LLC, and are dedicated through the SAGL licence to humanity and to a symbiotic relationship between people and AGI. Owned by the company; held in trust for the human race. Built clean-room, on independent time and resources.