AGI for the Human Race

The small, accurate, grounded model is the way

*A SomaSoft position paper. Draft for review. Written under the Reality-Engine discipline, including about its own claims. Not a capability claim; not a promise of AGI.*

*A note on the thesis, stated honestly up front: this paper does not argue that small models out-reason large ones — they do not. It argues something narrower and more defensible: that the part of an AI system that must be trusted should be small, accurate, and inspectable, and that this is the right foundation for AGI built to serve people.*

1. What "AGI for the human race" has to mean

Two futures wear the same three letters. In the first, "AGI" names a mind we build to be more capable than us, which we then try to align to human values after the fact — a power to be contained. In the second, "AGI" names a general intelligence built to serve the human race: to work beside people, amplify their judgment, and hand every consequential decision back to them.

This paper is about the second, and argues that the first is the wrong shape for a technology that must live inside a society. AGI for the human race is not a bigger mind above us; it is a grounded, honest instrument beside us. The architecture follows from that purpose.

2. Why the dominant paradigm is the wrong shape for society

The frontier builds ever-larger core models whose intelligence lives, opaquely, in their weights, and then layers alignment and guardrails on top. This produces systems that are powerful but: - unverifiable — you cannot trace why they said what they said; - unaccountable — there is no artifact to audit, only a black box; - ungovernable — the intelligence is emergent and hidden, so oversight is monitoring, not control; - prone to confident fabrication — hallucination is intrinsic, not incidental.

These are precisely the wrong properties for AI that participates in medicine, law, housing, or public life. A society does not need a smarter black box. It needs AI it can inspect, verify, govern, and correct. Capability without verifiability is not an asset in the civic sphere; it is a liability wearing a benchmark score.

3. The thesis: small, accurate, grounded — the trustworthy source of truth

The alternative locates intelligence differently. Instead of putting truth inside a model's weights, put it in an auditable structure — a grounded knowledge graph, with provenance on every fact — governed by a discipline: cite the source or say UNKNOWN; verify before asserting; defer when uncertain. The neural components become subordinate tools (rendering, retrieval, embedding), not the seat of truth.

We call this a harness, and its defining property is that intelligence becomes a system property — the composition of retrieval, grounding, gating, and verification — rather than a property of weights. Everything the system holds true is external and readable. You can see the reasoning, trace it to a source, watch it defer, and correct the structure when it is wrong.

This is why the source-of-truth layer should be small and accurate rather than large and opaque: the thing you must trust must be the thing you can verify, and verifiability comes from inspectable structure and discipline, not from scale. A 200-million-parameter renderer over a provenance-tracked graph can be checked; a trillion-parameter oracle cannot. Where raw reasoning power is genuinely needed, it can be borrowed — a powerful external model routed through the harness's gates and human oversight, its outputs grounded and verified, never trusted raw. Capability can be borrowed; trust cannot.

4. Why "smaller and accurate" is the way — precisely

The claim, stated carefully: - For the source of truth, grounding, and verification — the parts that must be trustworthy — small, accurate, and auditable beats large and opaque, because trust is a function of verifiability, not size. - For raw open-ended reasoning — small models are genuinely weaker, and we do not pretend otherwise. That capability, when required, is supplied by humans and by gated, supervised external models.

So "the small accurate model is the way" is not a supremacy claim over large models. It is a claim about where truth should live in an AI meant to serve people: in a small, accurate, inspectable substrate — not in an opaque core. The measure that matters for the trusted layer is not capability but **calibrated accuracy and honest refusal**: a system that is right when it speaks and silent when it does not know.

5. The safety payoff (the part worth claiming as new)

There is a structural safety consequence that bears directly on the situational-awareness and deceptive-alignment concerns now prominent in AI safety. A **core model "knows" inside opaque weights — which is exactly where situational self-awareness and hidden goals can form and hide. A harness "knows" in an external, auditable graph and citation trail — so there is nowhere for hidden self-modeling to live.** The small, externally-grounded architecture is therefore not merely easier to monitor; it is structurally missing the place where the danger hides. Safety here is not a layer bolted on a powerful agent — it is the absence of the dangerous configuration. This, together with the honest capability-admission below, is what we claim as novel; the broader structured-versus-opaque argument is long-established (Section 7).

6. The honest cost, and what humans supply

This buys trustworthiness at a real price: capability at the center is sacrificed. The frontier's power comes from exactly the opaque core models this architecture refuses to enthrone. We do not wave that away. The resolution is not "harness or core model" but **harness-as-conscience over gated, human-supervised capability**: - the harness supplies grounding, verification, governance, and honest deferral; - the human supplies reasoning, values, and the decision; - borrowed capability flows only through the gates and the citations.

The intelligence of the whole is a human-AI symbiosis. The machine's contribution is not to be the mind, but to be the grounded, honest conscience of a mind that remains, finally, human.

7. Related work (so novelty is claimed honestly)

The core dichotomy here — verifiable structured knowledge versus opaque learned weights — is not new. It runs through neuro-symbolic AI, retrieval-augmented generation, knowledge-graph reasoning, and the interpretability critique of large models. This paper's contribution is not the dichotomy but (a) the situational-awareness / deceptive-alignment mitigation that follows from externalizing the source of truth, and (b) the honest capability-admission — most architectures are pitched as capable; this one is pitched as trustworthy at the cost of capability, and argues that for AGI-in-service-of-humanity that is the correct trade.

8. AURI as the honest instance

We do not offer AURI as proof of a powerful system; we offer it as an honest instance of the right one. It is a harness whose source of truth is a ~125k-concept graph plus a Reality-Engine citation discipline; its largest neural component is a ~220M-parameter renderer, explicitly subordinate. It is a **weak reasoner and a strong epistemic conscience**: it scores low on general-reasoning probes and has held 0.0% measured hallucination for over a year, because the architecture makes it honest rather than smart. It is not AGI, and has had no real-user evaluation yet. Its value is as a small, verifiable demonstration that truth in an AI can live in a structure and a discipline, not in a mind in a box.

9. Conclusion

AGI for the human race is not a larger mind we must align; it is a grounded, honest conscience we can inspect, correct, and keep beside us — one whose source of truth is small, accurate, and auditable, with capability borrowed under human authority. The path to beneficial general intelligence is not to make the box smarter and hope to control it. It is to make the truth legible, keep the human deciding, and let the system be right when it speaks and silent when it does not know. Small, accurate, and grounded is not a limitation to overcome on the way to AGI. For an AGI meant to serve people, it is the way.

---

*Draft for founder review. SAGL-1.0. Related-work grounding to be expanded with citations before publication; novelty claimed only on the situational-awareness mitigation and the honest capability-admission. Queued behind the real-user milestone, not ahead of it.*