PRIORITY FILING - INITIAL DRAFT
A fully revised and formatted version will follow.
Gregory Caldwell Beier
Susarb LLC, Cambridge, Massachusetts
March 19, 2026
From the Prolegomena to the Unification of Economics, Computer Science, Mathematics, and Physics by Gregory Caldwell Beier (forthcoming 2026).
© 2026 Gregory Caldwell Beier. Licensed under CC BY-NC-ND 4.0.
This paper defines the Judgment Kernel as the minimal information architecture satisfying three simultaneous conditions: incentive-compatible elicitation of human judgment (the market surface), geometric fidelity in AI training data (the AI surface), and coupling of both surfaces through a single data stream. Three propositions establish that such an architecture exists as a canonical minimal implementation (Proposition 106), that it is self-improving through the autocatalytic loop between human disclosure and AI processing (Proposition 107), and that it preserves the structure of irreducible human judgment in a format that machine intelligence can process without destroying what makes the judgment valuable (Proposition 108).
Keywords: judgment kernel, human-AI interface, disclosure architecture, geometric fidelity, incentive compatibility, dual surface, Beier Extropy
Define the Judgment Kernel as the minimal information architecture that simultaneously satisfies three conditions:
(JK-1) It elicits honest, structured human judgment about constraint environments through incentive-compatible mechanisms (the market surface).
(JK-2) It produces training data whose geometric signature in weight space faithfully encodes the conservation laws of the domain it describes (the AI surface).
(JK-3) The two surfaces are coupled through a single data stream: the same disclosure that fixes prices trains the AI, and the same AI that processes disclosures improves the information environment for subsequent human judgment (the coupling condition).
Then:
(i) An architecture satisfying JK-1, JK-2, and JK-3 simultaneously exists. It is the five-field disclosure format scored by proper scoring rules with track record accumulation.
(ii) Any architecture satisfying all three conditions must contain, at minimum: a structured elicitation format with explicit output fields (to satisfy JK-1 via Propositions 5-7), verified identity with outcome scoring (to satisfy JK-1 via Proposition 7 and JK-2 via NR-1's requirement that training data respect conservation), and a scaling assessment field (to satisfy JK-3 via Proposition PF-2, which proves interaction-effect information is generated only when explicitly elicited).
(iii) The Judgment Kernel is therefore the intersection of three independently derived requirements - market incentive compatibility, AI geometric fidelity, and cross-surface coupling - and the five-field scored disclosure format sits at that intersection.
Proof. (i) Existence. The five-field format with proper scoring satisfies JK-1 by Propositions 5-10 (incentive compatibility of the disclosure mechanism). It satisfies JK-2 by the conjunction of BE-3 (the disclosed signal is Blackwell-superior to the price signal, so AI trained on it has access to strictly more information about the constraint state) and NR-1/NR-3 (when the training data respects conservation laws, the trained model's Fisher geometry faithfully encodes those laws as near-zero eigenvalues). It satisfies JK-3 by the Dual-Surface Proposition (Beier, 2026a) (I(V;D|P) > 0 guarantees both surfaces improve simultaneously) and MI-1 (the human-AI exchange is superadditive when both operate on the shared disclosure stream). Existence is established by construction.
(ii) Necessity of components. Suppose an architecture satisfies JK-1 without structured output fields. Then disclosed judgments are unstructured narrative, which by PF-1 carries strictly less information about the constraint state than the five-field format. The architecture fails to maximize the information available on the market surface, violating the minimality condition (a strictly better architecture exists). Therefore structured output fields are necessary for any minimal architecture satisfying JK-1.
Suppose the architecture satisfies JK-2 without verified identity and outcome scoring. Then the training data is unattributed and unscored. Unattributed data cannot be weighted by track record quality, so the AI trains on an unweighted mixture of accurate and inaccurate judgments. By the metric tensor result (BT-2), the track record weight is the metric that distinguishes signal from noise in the disclosure space. Without it, the AI's training signal has lower effective information content. The architecture fails to maximize geometric fidelity, violating minimality. Therefore verified identity with outcome scoring is necessary.
Suppose the architecture satisfies JK-3 without a scaling assessment field. Then by PF-2, the disclosed judgments carry approximately zero information about interaction effects at aggregate scale. The AI surface receives no interaction-effect data. The coupling between surfaces is incomplete: market participants receive individual-transaction analysis but the AI cannot learn scaling dynamics because the training data doesn't contain them. The coupling condition JK-3 requires that the AI improve the information environment for subsequent human judgment, which requires the AI to detect patterns across disclosures including interaction patterns. Without scaling data, the AI's cross-disclosure analysis is limited to individual-asset patterns, missing the interaction effects that produce the most consequential coordination failures. The architecture fails to satisfy the coupling condition fully. Therefore the scaling field is necessary.
(iii) The Judgment Kernel is the architecture at the intersection of these three requirements. Each requirement independently constrains the design space. Their intersection is the five-field scored disclosure format - a canonical minimal architecture that satisfies all three simultaneously. QED.
Let JK(t) denote the Judgment Kernel operating at time t, producing disclosure quality Q_D(t), AI geometric fidelity G_AI(t), and coupling strength rho(t). Under the conditions of Propositions LA-1 (learning acceleration), MI-1 (superadditivity), and IC-1/2/3 (inexhaustibility of theta_C):
d/dt [Q_D(t) + G_AI(t) + rho(t)] > 0
The Judgment Kernel's performance on all three surfaces improves simultaneously over time, without external intervention, once the architecture is operating above critical mass N*.
Proof. The improvement on each surface drives improvement on the others through the coupling condition JK-3.
Surface 1 (Market): Disclosure quality Q_D(t) improves because scored feedback improves individual calibration (Proposition 6), cross-learning from other disclosures improves collective calibration (Proposition CL-1), and the performative property generates reasoning that would not otherwise exist (Proposition PF-1). The improvement rate is itself increasing (Proposition LA-1).
Surface 2 (AI): Geometric fidelity G_AI(t) improves because the training corpus grows in volume (more disclosures accumulate over time), in quality (the disclosures are produced by agents whose calibration is improving by Surface 1), and in track record discrimination (longer track records provide sharper metric tensor weights by TB-2). By NR-1/NR-3, higher-quality training data produces tighter conservation signatures in the Fisher geometry.
Surface 3 (Coupling): Coupling strength rho(t) increases because the AI's improved geometric fidelity enables better cross-disclosure pattern detection (identifying interaction effects across domains that no individual discloser can see), which enriches the information environment for subsequent human disclosure (the AI surfaces patterns the human then incorporates into their next assessment), which improves the training data for the next round of AI learning. This is the mutual ignition loop (MI-1, MI-2).
The three improvements are mutually reinforcing: better disclosures train better AI, better AI improves the disclosure environment, better environment produces better disclosures. The fuel for the improvement is theta_C - the forward-looking constraint information that the world keeps generating (IC-1/2/3). Since theta_C > 0 permanently, the fuel never runs out. Since the improvement is self-reinforcing above N* (MI-3), no external intervention is required to sustain it.
The improvement is bounded (LA-2): Q_D approaches Q_max (irreducible constraint uncertainty), G_AI approaches the physical limits of the training data's information content, and rho approaches 1 (perfect coupling). The trajectory is S-shaped - accelerating initially, decelerating as the bounds are approached. But the direction is monotonically upward on all three surfaces simultaneously. QED.
Let H denote the space of computationally irreducible human judgments (the domain where humans have comparative advantage by Proposition EX-1). Let M denote the space of computationally reducible machine operations (the domain where AI has comparative advantage). The Judgment Kernel JK is a map:
JK: H -> M
that projects irreducible human judgment into structured data that machines can process, aggregate, route, and learn from, while preserving the information content of the projection (by BT-1, the total information is framework-invariant).
The Judgment Kernel satisfies three properties that no other known architecture satisfies simultaneously:
(i) Fidelity: the projection preserves the constraint information in the human judgment. By the Beier Tensor formalization (BT-1/BT-2/BT-3), the five-field format carries the judgment's information content through the coordination system without compressing it to a scalar.
(ii) Incentive compatibility: the human is compensated for performing the projection through track record capital that accrues regardless of whether the price has already adjusted (Proposition 9, micro-founded by J4 and secured by IC-1/2/3). The compensation is permanent because theta_C is permanent.
(iii) Bidirectionality: the projection is not one-way. The AI processes the projected judgments and returns enriched information to the human (pattern detection across disclosures, anomaly identification, mosaic interpolation), improving the human's subsequent judgment. The bridge carries traffic in both directions.
The Judgment Kernel is therefore the cognitive architecture for wise machines - machines whose intelligence is grounded in, trained on, and accountable to the irreducible judgments of human beings operating under real stakes with verified identities and scored outcomes.
Proof. (i) Fidelity follows from BT-1 (coordinate invariance of information content) and BE-3 (Blackwell improvement: the projected signal is strictly more informative than the scalar price). The five-field format preserves K dimensions of the judgment where the price preserves one. The equivocation H(J|P,D) is strictly less than H(J|P). The projection is still lossy (no finite format captures the full irreducible judgment), but E_B = I(J; D|P) > 0 bits of the original judgment are preserved that the scalar price would have destroyed.
(ii) Incentive compatibility follows from the complete chain: Propositions 5-7 (truth-telling under proper scoring with commitment), TB-1/TB-2 (the threshold tightens over time, making the mechanism self-reinforcing), Proposition 9 (the return is independent of price efficiency), and IC-1/2/3 (the return is permanent because theta_C is permanent). The human is not performing charity by disclosing. The human is building an asset - track record capital - that compounds over time and cannot be replicated by any machine (by EX-1, the judgments that build the track record are in the computationally irreducible domain where humans have permanent comparative advantage).
(iii) Bidirectionality follows from MI-1 (superadditivity: the combined system exceeds the sum of parts) and the coupling condition JK-3. The AI receives the projected human judgment, processes it (cross-referencing, pattern detection, anomaly identification, mosaic interpolation across thousands of disclosures), and returns the processed output to the information environment the human operates in. The human's next disclosure is informed by the AI's processing of previous disclosures. The bridge carries irreducible judgment from human to machine AND processed patterns from machine to human. Neither direction is sufficient alone. Both together produce the mutual ignition (MI-2, MI-3) that makes the system self-sustaining.
The Judgment Kernel is the architecture that makes this bridge possible. Without it, human judgment remains invisible to machines (compressed into prices that destroy the reasoning). Without it, machine intelligence remains ungrounded (trained on compressed residue rather than structured judgment). With it, the two forms of intelligence - irreducible human wisdom and scalable machine processing - meet at a structured interface where each contributes what the other cannot, measured by the same proper scoring rules, improving through the same feedback loop, serving the same constraint environment.
This is what "architecting superintelligence's judgment kernel" means. Not building the superintelligence. Building the architecture through which human judgment and machine intelligence combine into something wiser than either alone - grounded in physics, accountable to outcomes, and improving over time because the world keeps generating the problems that only the combination can solve. QED.