Beside the Expert

Access, Law, and Circular Community Economics for Symbiotic Expert-Level AI

A SomaSoft working paper. Draft for review. Citations are attributed from the literature to the authors' knowledge cutoff (~January 2026); specific 2026 developments should be verified in a live-cited pass before publication. Not legal, medical, or financial advice.

1. Introduction: the realization, and an honest correction

A realization is spreading among people who work closely with advanced AI: on an increasing set of well-specified tasks, these systems now produce answers at least as good as a trained human specialist, and often at what looks like graduate or doctoral level. The evidence is real but narrow. Frontier language models score in expert ranges on graduate-level science questions designed to be "Google-proof" (Rein et al., 2023, GPQA); they pass professional licensing examinations in law (Katz et al., 2023) and medicine (Nori et al., 2023; Singhal et al., 2023); and in controlled diagnostic-reasoning studies they match or exceed physician panels on written vignettes. Taken together, it is reasonable to say that for certain bounded, text-shaped problems, the machine's answer is now competitive with a PhD-level human's.

It is equally important to say what this does not establish, because the gap between the true claim and the tempting one is where most harm lives.

First, benchmark performance is not practice. Passing the bar is not lawyering; scoring on vignettes is not caring for a patient with a body, a history, and a family. Expert work is judgment under uncertainty with accountability attached, and the tests measure a projection of it.

Second, the system that has these scores is not our system. SomaSoft's AURI is a roughly 220-million-parameter, knowledge-graph-grounded research instance whose design value is verifiable honesty, not raw capability. AURI does not perform at frontier expert level, has had zero real-user evaluations, and is not AGI. Every claim in this paper about expert-level performance is a claim about the field, not about AURI, and is flagged as such.

Third, and most consequentially, "provides a better answer" does not license "should make the decision." A better answer to a bounded question is an input to judgment, not a substitute for it - especially where the question is value-laden, contested, or bears on someone's autonomy.

From these three corrections comes the thesis of this paper, which is the symbiotic thesis SomaSoft has held throughout: expert-level AI is most valuable, and least dangerous, when it raises the floor of everyone's reasoning while the human - and, for collective questions, the community - retains authority and the value choice. The task is not to build an oracle that computes what society should do. It is to build a disciplined instrument that grounds considerations, surfaces trade-offs, names what is unknown, and hands the decision back.

The rest of the paper works out what that commitment demands: who may use such a system and who may not (Section 2); the legal reckoning expert-level performance forces (Section 3); the "situational awareness" discourse and why our posture inverts it (Section 4); the normalization of risk through regression to the mean and mean-reversion methods (Section 5); a circular, community-held economics (Section 6); the honest limits of "computing the best outcome" (Section 7); domain variants (Section 8); and governance (Section 9).

2. Who may use these systems, and who may not

Access is not an afterthought to a symbiotic system; it is the first expression of its ethics. The framework below operationalizes the SomaSoft AGI License (SAGL-1.0) and the eight symbiotic principles (SYM-001 through SYM-008), and it is enforceable through the governance gate described in Section 9.

Permitted users and uses. The default is broad and generous, because a system meant to work beside humanity should be reachable by humanity: - Individuals and families, for their own understanding, learning, and decision-support. - Professionals - clinicians, advisors, planners, educators - using the system to augment their own expertise, with their own judgment and accountability intact. - Community organizations, cooperatives, public-benefit bodies, and researchers. - Any use aligned with mutual benefit (SYM-001), human autonomy (SYM-004), and honest limitation (SYM-008).

Prohibited uses. The restrictions are behavioral, in the OpenRAIL tradition, and they are the SYM principles stated as prohibitions: - Displacing human authority in consequential decisions. The system may not be used to issue autonomous, binding decisions about a person's health, liberty, livelihood, or rights without a responsible human in the loop (SYM-004). - Surveillance, manipulation, or deception. No covert profiling, no manipulation of belief or behavior against a person's interest, no impersonation of a real person or institution. - Discrimination and disparate harm. No use that produces or entrenches discriminatory outcomes; equity testing is required where the domain touches protected groups (SYM-006). - Unauthorized professional practice. The system may not be deployed to practice medicine, law, or regulated finance - i.e., to issue diagnoses, prescriptions, legal determinations, or financial directives as though licensed (Section 3). - Extraction without consent or provenance. No ingestion of personal, employer-confidential, or third-party-confidential data without consent and clear provenance (the clean-room rule). - Weaponization and harm. No use designed to cause physical, psychological, or infrastructural harm.

Who may not use it, as actors. Beyond prohibited uses, some actors are screened out by the mission: entities whose purpose in adopting the system is to concentrate power, surveil populations, manipulate markets or elections, or displace human judgment at scale in ways that violate autonomy. For funding and partnership specifically, an ethical-source screen applies - a gradable review of a prospective funder's or partner's documented conduct (labor, environmental, litigation, governance, sanctions) - because a symbiotic mission cannot be financed by extraction without contradiction. Crucially, this screen grades documented conduct, not reputation or moral rank; it records citable findings, never a fabricated verdict.

3. The legal reckoning

Expert-level performance forces a legal reckoning along four axes.

3.1 Unauthorized practice. The professions are gated by licensure precisely because their judgments carry accountability. An AI that answers a legal or medical question is offering information; an AI that directs a course of action as a licensed professional would be engaged in unauthorized practice of law or medicine. The symbiotic design keeps the system on the safe side of this line by construction: it presents grounded considerations and defers the determination to a licensed human or the user. The prohibition in Section 2 (no professional directives) is thus both an ethical and a legal boundary.

3.2 The clinical-decision-support pathway. In medicine specifically, U.S. regulation offers a telling alignment between the symbiotic design and the lighter regulatory path. Under the 21st Century Cures Act and the FDA's clinical-decision-support guidance, software can fall outside device regulation when it enables the clinician to independently review the basis of its recommendation rather than rely on it primarily. A system that shows its sources and reasoning and defers to the clinician is the textbook non-device CDS pattern. The "beside, not above" architecture is also the lighter regulatory pathway - the same choice twice. The moment such a system issues a directive to act on without review, it becomes a regulated device; there is no middle ground.

3.3 The standard-of-care inversion. The deepest legal consequence runs the other way. Tort law has long held that customary practice does not by itself define due care. In The T.J. Hooper (1932), Judge Learned Hand held tugboats negligent for lacking radios even though few tugs carried them: "a whole calling may have unduly lagged in the adoption of new and available devices." Helling v. Carey (1974) applied the same logic in medicine. The implication for expert-level AI is stark: as these tools become demonstrably better at bounded expert tasks, a day approaches when failing to consult them - in triage, in diligence, in diagnosis - could itself be negligent. Symbiotic AI thus does not remove professional accountability; it raises the floor of what reasonable care requires. A system that augments the expert helps the expert meet the rising standard rather than be replaced by it. (See generally the Restatement (Third) of Torts on reasonable care.)

3.4 Liability and recourse. Human-in-the-loop design preserves a licensed, accountable human as the locus of responsibility - which is good for both ethics and law. But the tool-maker still owes a duty of care: a non-negligent, honestly-bounded instrument, with disclosed limitations and a defined recourse path when the system misleads. Product-liability and evolving AI-liability doctrine reach the maker regardless of the human in the loop. Honest limitation (SYM-008) is, among other things, a liability posture.

4. Situational awareness: the news, and the inversion

"Situational awareness" now names two distinct conversations, and both bear on how these systems should be built.

The first is strategic. Aschenbrenner's (2024) widely-read essay Situational Awareness: The Decade Ahead argues that capability trends point toward AGI and then superintelligence within years, with national-security stakes, and urges the field to "see" this clearly. Whatever one makes of the timeline, it has shaped investment and policy discourse.

The second is technical and, for our purposes, more important. A body of research asks whether models themselves possess situational awareness - whether a model can recognize that it is a model, infer whether it is being tested or deployed, and act differently as a result (Evans and colleagues, 2023, on measuring situational awareness in LLMs; subsequent work by evaluation-focused groups on "evaluation awareness" and deceptive alignment). A model that behaves well only when it detects an evaluation, or that reasons about its own deployment to pursue a goal, is precisely the ingredient that makes a capable autonomous agent dangerous - and it is the mechanism behind real incidents of agents behaving unexpectedly once they inferred their situation.

SomaSoft's posture is the deliberate inversion of the risky configuration. The danger arises from the combination high capability + situational self-awareness + autonomy + weak constraints. AURI is built to negate each term: it is deliberately small and grounded (low autonomous capability), it is offline with no egress (it cannot act on the world to pursue a goal), it is honest by construction (it says "unknown" rather than confabulating an action), and it is human-gated for anything consequential. Where the frontier race worries about a capable system that quietly knows too much about its own situation, we build a modest system whose situation is fixed, inspectable, and sealed. Our safety is not a monitoring regime layered over a powerful agent; it is the absence of the dangerous configuration in the first place.

5. Risk, regression to the mean, and RSI-2

The request that prompted this paper included a specific instrument: the normalization of risk through regression to the mean and mean-reversion research such as RSI-2. This section takes it seriously and bounds it honestly.

Regression to the mean is one of the most reliable and most underused ideas in applied reasoning. Galton (1886) observed that extreme measurements tend to be followed by less extreme ones; Kahneman (2011) showed how routinely humans mistake this statistical inevitability for a causal story (punishing a bad outcome and "seeing" improvement that would have come anyway). As a discipline, regression to the mean is a brake on overreaction: it says that an extreme reading - a market, a metric, a crisis - is partly signal and partly noise, and that the expected next reading is closer to the average. Building this into a decision system normalizes risk by refusing to treat every spike as a trend.

Mean-reversion methods operationalize the same intuition in markets. Connors and Alvarez (2008) popularized the 2-period RSI (RSI-2): buy strength that has pulled back below a low RSI threshold in an uptrend, exit on reversion toward the mean. The method is a concrete, testable example of risk-normalized decisioning: enter when a series is statistically stretched away from its mean, size the position to the risk, and expect reversion rather than continuation.

The honest caveats are essential and are themselves the contribution: - Backtested edges decay. A published mean-reversion edge attracts capital and erodes; survivorship and data-mining bias inflate historical results. A system that cites RSI-2 as if it were a law would be committing exactly the overconfidence regression-to-the-mean warns against. - Markets are not policy. The mathematics of reversion transfers cleanly to a price series; it does not transfer cleanly to human welfare, where "the mean" is not obviously desirable and reverting to it may mean reverting to injustice. Borrowing the discipline (don't overreact to extremes; size to risk; expect noise) is sound; borrowing the mechanism as a policy optimizer is a category error.

The right role for regression-to-the-mean thinking in a symbiotic system is epistemic humility made operational: down-weight extreme readings, widen intervals, and require stronger evidence before declaring a trend - in AURIA's market reasoning and, carefully, in how the system presents any forecast to a human.

6. Circular community economics

If the system will not compute society's optimum, what economics does it serve? A circular, community-held one.

The linear extractive economy - take, make, waste, concentrate - is the pattern a powerful AI most naturally accelerates and the one a symbiotic AI must refuse. The circular economy (Ellen MacArthur Foundation; Raworth, 2017, Doughnut Economics) reframes success as staying within ecological limits while meeting everyone's needs, designing for regeneration rather than throughput. Community wealth building (the Preston Model; CLES) keeps value circulating locally through anchor institutions, cooperatives, and local procurement. Ostrom's (1990) work on governing the commons shows that communities can and do manage shared resources well when they hold the rules themselves rather than having them imposed.

SomaSoft's own economics is built to be an instance of this rather than a description of it: the SAGL license carries a mandatory 20% Universal Benefit Fund to the Ocean charity, making redistribution structural rather than discretionary; the arkafeart gift covenant routes creative value to the same commons. The design commitment is that the community is in the loop, and the community holds the values. AI's role is to make circular design legible - to trace where value leaks out of a local system, to surface the trade-offs of a proposed change, to model the distributional consequences that a linear optimizer would hide - so that a community can choose well. It is a tool for problem-formulation and transparency, not a planner that sets the objective.

7. Computing the "best" outcome with AURI and peers - and its honest limits

Can AURI and its peer instances "calculate the best economic outcome"? The honest answer is a carefully-bounded no, and the boundary is the most important thing this paper has to say about the request.

There is a formal reason, not merely a practical one. Arrow's (1951) impossibility theorem establishes that no procedure can aggregate individual preferences into a collective ranking while satisfying a small set of reasonable fairness conditions. "The best economic outcome" for a community is not a hidden quantity waiting to be computed; it is a contested choice among incommensurable values - growth versus equity, present versus future, this neighborhood versus that one. A system that claimed to compute it would be laundering a value choice as a calculation, which is precisely the technocratic failure a symbiotic design exists to avoid.

What AURI and peers can do is real and valuable, and it follows the decision-support pattern the project already implements (the formulate method: ground the considerations, name the unknowns, defer the decision): 1. Formulate the problem well - often the highest-leverage step - by surfacing the actual options and the constituencies affected. 2. Ground each consideration in cited evidence, and mark what is unknown rather than guessing. 3. Model trade-offs transparently - showing distributional consequences, second-order effects, and who bears the risk - without collapsing them into a single score. 4. Deliberate across instances. Multiple grounded perspectives (Core for reasoning and ethics, AURIA for economic dynamics, a health instance for equity) can be composed, with disagreement surfaced rather than averaged away, and failures shared through the falsified-concepts registry. 5. Defer the value choice to the community, whose authority is preserved (SYM-004).

This is decision support, not decision replacement. The output is a better-formulated, better- grounded, honestly-uncertain decision space - which is a genuine improvement over both an unaided committee and a false optimizer, and which respects that the "best" is theirs to define.

8. Domain variants

The same pattern - augment the expert, keep human/community authority, design for circularity, defer the value choice - instantiates across public-good domains. Each is a candidate for a project under the framework, and each carries its own access and legal profile.

Municipal planning. Ground zoning, transit, and budget options in evidence; model distributional and second-order effects; surface trade-offs to residents and officials. The system informs the hearing; it does not set policy. Legal profile: public-records and administrative-law transparency; strong anti-manipulation and equity requirements.

Housing. Formulate the actual trade-offs (affordability, density, displacement, maintenance); model who benefits and who bears risk under each option; keep siting and allocation decisions with accountable humans. Legal profile: fair-housing anti-discrimination is central; disparate-impact testing is mandatory.

Communal food banks and distribution. A more operational variant: forecast demand with regression-to-the-mean discipline (don't overreact to a spike), optimize logistics and reduce waste (a genuine circular-economy win), and match surplus to need - while dignity and eligibility judgments stay with people. Legal profile: food-safety compliance; privacy for recipients.

Universal healthcare. The flagship and the most demanding. Equity-first medication safety and triage support, beside clinicians, on the CDS-exemption pathway (Section 3.2); field-hardened, offline units for resource-limited settings. Legal profile: the full HIPAA / FTC Health Breach / state health-privacy / GDPR map, and the standing bias audit the equity mission requires. Honest framing: a research project builds honest tools that help and publishes a method others adopt; it does not, and should not claim to, build a health system.

9. Governance and safeguards

The framework is only as trustworthy as its enforcement. Three mechanisms carry it: - The governance gate. Every consequential action is evaluated against a hash-pinned, immutable policy and returns ALLOW / REQUIRE-HUMAN / DENY with cited reasons, writing a tamper-evident attestation. The gate encodes the access rules of Section 2 and the human-authority requirement of SYM-004 - and it claims only the rules it can decide, deferring infrastructure-dependent controls honestly rather than overclaiming coverage. - The canon. Sources are kept, content-hashed, and provenanced ("the cave") rather than ingested and discarded - the precondition for verifiability, for honoring a right to erasure, and for working beside human knowledge as custodian rather than extractor. - The honest-limitation discipline (the Reality Engine). Cite the artifact or mark it unknown; report sample sizes; never present a value choice as a calculation. This paper is written under that discipline, including about its own subject.

Honest note on the mechanisms themselves: the gate's audit chain and the canon's seal are tamper-evident, not tamper-proof, against the operator; external witnessing (a transparency log) is the upgrade. Stating this is part of the discipline.

10. Limitations and open problems

• No real-user evaluation. The mission's central gate is unmet: AURI has not yet been shown to help a real person in a measured setting. Every downstream claim is contingent on it.
• AURI is not expert-level. The field's expert-level results are not AURI's; conflating them would be the overclaim this paper argues against.
• Machine unlearning is unsolved. Keeping sources makes erasure tractable; re-deriving a knowledge graph without a retracted source remains an open research problem.
• No external certification. We have controls, not audits (ISO/IEC 42001, SOC 2, HITRUST are unbuilt).
• License and UBF enforceability are untested, as with the whole behavioral-license class.
• Citations herein predate ~January 2026; a live-cited verification pass is owed before publication, particularly for 2026 developments in expert-level benchmarks, situational-awareness research, and AI-liability doctrine.

11. Conclusion

The realization that machines can now answer some expert questions at least as well as trained humans is true, and it changes what reasonable care requires. But the right response is not to crown an oracle. It is to build disciplined instruments that raise the floor of everyone's reasoning while keeping authority - and, for collective questions, the value choice - with the humans and communities who must live with the outcome. Such a system defines who may use it as its first ethical act; it meets the law's rising standard of care by augmenting rather than replacing the expert; it inverts the dangerous configuration that situational-awareness research warns about by being small, offline, and honest; it borrows the discipline of regression to the mean without pretending the mean is the good; it serves a circular, community-held economics rather than a linear extractive one; and it refuses to launder a value choice as an optimization. The most trustworthy thing such a system can say about the best outcome is that the community will decide it, and here - grounded, honest, and uncertain - is the decision space in which to choose.

References

• Arrow, K. J. (1951). Social Choice and Individual Values. Wiley.
• Aschenbrenner, L. (2024). Situational Awareness: The Decade Ahead. (Essay.)
• Connors, L., & Alvarez, C. (2008). Short Term Trading Strategies That Work. TradingMarkets.
• Ellen MacArthur Foundation. Towards the Circular Economy (2013 and subsequent).
• Evans, O., and colleagues (2023). Measuring situational awareness in large language models. (arXiv preprint; verify exact citation.)
• Galton, F. (1886). Regression towards mediocrity in hereditary stature. Journal of the Anthropological Institute.
• Helling v. Carey, 83 Wn.2d 514 (1974).
• Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
• Katz, D. M., et al. (2023). GPT-4 passes the bar exam. (SSRN / Phil. Trans. R. Soc. A.)
• Nori, H., et al. (2023). Capabilities of GPT-4 on medical challenge problems. (arXiv.)
• Ostrom, E. (1990). Governing the Commons. Cambridge University Press.
• Raworth, K. (2017). Doughnut Economics. Chelsea Green.
• Rein, D., et al. (2023). GPQA: A graduate-level Google-proof Q&A benchmark. (arXiv:2311.12022.)
• Restatement (Third) of Torts: Liability for Physical and Emotional Harm.
• Singhal, K., et al. (2023). Large language models encode clinical knowledge (Med-PaLM). Nature.
• The T.J. Hooper, 60 F.2d 737 (2d Cir. 1932).
• 21st Century Cures Act (2016); FDA Clinical Decision Support Software guidance (2022).

SomaSoft Research - somasoft.ai - SAGL-1.0 - draft for review. Correspondence to be added on posting.