PaperlayerView exact revision on arXiv

Source-linked AI summary

Constitutive Priors for Machine Intelligence: A Legitimacy Theory of the Artificial Physical World

Jiang Jiang, Yifu Sun, Qi Shen

arXiv:2608.15147v1cs.AIcs.MA

TL;DR

Physical AI lacks grounded meaning and faces a cold-start problem because symbolic systems do not directly connect symbols to the physical world. This paper proposes a four-world account and a legitimacy criterion for prior frameworks, concluding that archived, intentionally constituted physical domains can support binding priors whose deviations count as faults.

  • Problem

    Physical AI lacks grounded meaning because language models relate symbols to symbols rather than directly to the physical world.

  • Method

    The paper derives four task-relative cognitive worlds and tests prior-framework legitimacy through intentional constitution, readable archives, and direction of fit.

  • Results

    Prior frameworks are legitimate when their object world is intentionally constituted and supported by a readable generative archive, with deviations treated as world-level faults.

  • Takeaways & Limitations

    The framework applies within institutional physical domains that maintain generative archives and does not claim validity beyond that boundary.

  • Takeaways & Limitations

    The framework’s deployment claims are confined to the physical domain and stop applying where institutions no longer keep readable archives.

Abstract

from arXiv · show

Machine intelligence has conquered the symbolic world but stalled at the physical one. The stall is structural: physical AI faces a cold-start deadlock -- no intelligence without data, no data without deployed intelligence. Our thesis: the deadlock is real but unevenly distributed, and the exception has a name: the artificial physical world. Buildings, industrial facilities, and infrastructure are intentionally constituted and documented: designed artifacts ship with readable archives that precede and constitute their instances; here, norms are promulgated before instances, not averaged from them. Four contributions. (i) From a four-world ontology we derive a legitimacy criterion for constitutive prior frameworks: prior extraction is legitimate if and only if the object domain is intentionally constituted and has left a readable archive; the criterion is testable through direction of fit -- deviation from a constitutive norm is a violation in the world, not a revision of the model. (ii) We establish a layering lower bound: any such framework has at least four layers -- syntax, concept, knowledge, instance -- because four construction goals pair into mutually incompatible carriers. (iii) We register deployment claims across five industrial domains and a 32-class failure-mode vocabulary. (iv) We stake the framework on five falsifiable predictions, the central one checkable on the public engineering record: if it fails, the framework fails. Semi-formal arguments back these claims (Appendix A): a Gold-type boundary on rule coverage in archiveless worlds, a decidability result for failure reduction over closed concept layers, and a boundary theorem for certificate-anchored calculi. Large language models find an honored place here -- as readers of the archive, not as the archive. First of three companion works; the companions take up the questions deliberately left open.

1. Introduction: An Overlooked Asymmetry

The introduction identifies a domain-dependent asymmetry: framework-first approaches failed in symbolic AI, while engineered physical artifacts succeed because they are designed and documented before deployment. It therefore asks when prior-framework extraction is legitimate and previews a four-world analysis, a legitimacy criterion, necessary layering, and falsifiable predictions.

  • Motivation: Framework-first approaches failed in rule-based NLP and expert systems, whereas engineered artifacts routinely use drawings, control logic, and design specifications before operation.The contrast motivates investigating why framework-first construction works in artificial physical domains but failed in earlier symbolic systems.
  • Research question: The research question asks when an AI system may legitimately acquire a prior framework before accumulating large-scale observational data.The question concerns conditions under which a framework constrains instances as a promulgated norm, making failures violations rather than statistical deviations.
  • Contributions: The paper derives four task-relative worlds—phenomenal, basic physical, artificial physical, and artificial symbolic—from data interface, constraint structure, and design rules.The partition is instrumental for analyzing how learners acquire knowledge of the world, not a metaphysical claim.
  • Contributions: Prior extraction is legitimate if and only if a domain is intentionally constituted and its constitutive process leaves a readable generative archive.Under these conditions, the extracted framework binds instances as promulgated norms rather than statistical extrapolations.
  • Contributions: A qualifying prior framework has a layering lower bound of ≥4: syntax, concept, knowledge, and instance layers, because four construction goals require incompatible carriers.The four goals are stability, openness, abstraction, and concreteness.
  • Contributions: The framework is committed to five falsifiable predictions, including continued breakthrough order, a data-scarce design-domain victory map, concept-layer invariance, a VLA ceiling, and ontology-layer irreplaceability.The paper explicitly states that failure of the key predictions would make the framework wrong.

2. Re-reading the History of AI from the Learner’s Point of View

AI’s history looks different from the learner’s point of view: rule-based methods failed where no prior design rules existed, while statistical methods succeeded first in worlds with accessible data and constrained regularities. Physical AI exposes unresolved grounding, data scarcity, and normativity gaps because prediction does not determine correctness.

  • Rule-based AI: Rule-based NLP and expert systems failed because exceptions proliferated and articulated rules captured far less knowledge than experts actually used.The bottleneck was knowledge, not inference, and grammar rules became descriptions of descriptions when the object world lacked prior design rules.
  • Statistical AI: Statistical AI abandoned hand-written rule content, not priors: inductive bias moved into architectures, while generative archives were ingested as data.AlphaFold 2’s multiple sequence alignments exemplify an archive serving as input rather than norm.
  • Breakthrough order: Breakthroughs arrived earliest in closed-rule games, then code and language systems, followed by image generation, while robotics still lags.The sequence correlates with data-interface accessibility and constraint structure, rather than occurring randomly.
  • Physical-world cracks: Physical AI faces three cracks: unresolved symbol grounding, scarce and privatized instance data, and confusion between predicting dynamics and defining correct states.Large language models relate symbols to symbols, physical data lacks a free corpus, and world models estimate P(s_t+1 | s_t,a) without establishing what ought to happen.
  • Three-variable lens: The learner’s framework distinguishes data interface, constraint structure, and prior readable design rules; their combination explains AI’s historical failures, breakthroughs, and physical-world obstacles.The next section uses these variables to derive a four-cell partition of the cognitive world.

3. Deriving the Four Worlds: Partitioning the World from the Learner’s Point of View

The four worlds partition reality from the learner’s point of view, according to data interface, constraint structure, and design rules rather than metaphysics. They form a compositional stack in which the artificial physical world uniquely combines physical instances with readable, prior design norms.

  • Methodological scope: The partition asks which differences among worlds matter to an AI learner, not what the world is metaphysically made of.Its legitimacy is assessed by explanatory and predictive power in learning theory.
  • Partition variables: Three learner-facing variables determine the occupied cells: data interface, constraint structure, and whether design rules exist before instances.These variables concern how data is presented, how compressible regularities are, and whether prior rules constitute and regulate instances.
  • Artificial physical world: The artificial physical world comprises physical systems whose instances are purposively designed, governed by prior specifications, and documented in readable generative archives.Examples include buildings, machines, lamps, and air conditioners; the archive specifies both what an instance is and what counts as failure.
  • Composition and stack: The four worlds compose a complete intelligence stack, developing from phenomenal experience through basic physics and artifacts to symbols.They are floors rather than elective parallel ontologies, and none is dispensable.
  • Bidirectional structure: The artificial physical world is the only bidirectionally permeable floor: engineering moves downward from symbols, while cognition moves upward from experience, and it ships with its own source code.This makes it simultaneously a physical object and a symbolic norm; text-only language models receive only second-hand shadows and cannot substitute for direct world contact.

4. A Legitimacy Criterion for Prior Frameworks: When May the Framework Be Erected First, and Learning Done After?

A prior framework is legitimate only where symbolic norms intentionally constitute their objects and remain available in a readable archive. This boundary distinguishes designed, institutionally maintained artifacts from language, organisms, and inaccessible historical technologies, with direction of fit making deviations defects in the artifact rather than revisions to the framework.

  • The Legitimacy Criterion: Designed artifacts differ from language because drawings, specifications, standards, and control logic precede and regulate what counts as properly built.Grammar compresses prior linguistic usage, whereas engineering symbols prescribe artifact generation and conformity.
  • The Legitimacy Criterion: Temporal precedence is insufficient: genes precede organisms, but only constitutive norms define their objects and make deviations defects rather than variations.The discriminating variable is what the symbolic structure does, not merely when it exists.
  • The Legitimacy Criterion: Legitimacy requires an intentionally constituted object whose constitutive knowledge survives in a readable archive.Intentional constitution alone is insufficient when the knowledge is lost; the framework therefore targets artifacts that are archived and institutionally maintained.
  • The Legitimacy Criterion: Engineering archives are unusually accessible because they are written for readers and institutionally standardize names, symbols, and classifications across people and time.This makes drawings, BIM models, CAD programs, data sheets, control-logic books, codes, and standards a comparatively cheap learning interface.
  • Scope Boundary: The criterion stops applying when institutions stop maintaining archives, and the paper confines its deployment claims to high-utility, institutionally maintained physical artifacts.Consumer goods and domains without durable institutional records fall outside the stated claim boundary.
  • Direction of Fit: Direction of fit is world-fits-rule: when a building deviates from its drawings, the building is defective, whereas descriptive rules are revised when they conflict with the world.This supplies the operational basis for treating constitutive frameworks as authoritative over their instances.

5. Layering Necessity: What a Legitimate Prior Framework Must Look Like

Because the artificial physical world is intentionally constituted and archivally documented, a legitimate prior framework can derive its purposes, functions, and norms from design archives. Its four construction goals require mutually incompatible carriers, establishing a four-layer lower bound rather than an exactly four-layer architecture.

  • Legitimacy and construction goals: Legitimacy follows because the artificial physical world is intentionally constituted and its constitution leaves archives written to be read.The framework may take purposes, functions, and norms directly from constitutive norms promulgated by design archives.
  • Legitimacy and construction goals: The framework must satisfy four goals—stability, openness, abstraction, and concreteness—derived from the world picture and direction-of-fit structure.Abstraction enables learning to apply across future instances, while concreteness anchors concepts to specific physical objects.
  • Layering lower bound: Pairing the four goals across change rate and abstraction yields four pairs whose mutually incompatible carriers require a layered framework.The incompatibilities concern revision discipline and denotation: stable versus continuously ingesting carriers, and abstract types versus concrete clauses.
  • Layering lower bound: The required carriers are syntax for stable abstraction and knowledge for stable concreteness, with the full framework organized around all four goal pairs.The syntax layer provides stable grammar, while the knowledge layer codifies concrete failure modes, procedures, and operating criteria.
  • Layering lower bound: The proposition establishes at least four layers, while implementations may subdivide layers without compressing the theoretical skeleton.Potential fifth goals such as safety, interpretability, economy, and real-timeliness do not add a new content dimension or layer.

6. Falsifiable Predictions: Putting the Theory on the Table

The section stakes the framework as a research programme through five ordered, falsifiable predictions spanning breakthrough order, domain-specific priors, concept-layer invariance, VLA limitations, and ontology-layer irreplaceability.

  • P1 (breakthrough order): P1 predicts that the next scalable AI breakthrough will occur in the artificial physical world, before the open phenomenal world.It is falsified if general world-model systems trained mainly on open phenomenal video achieve scalable physical-operation success before archive-driven routes.
  • P2 (the prior’s map of victory and defeat): P2 predicts prior-framework methods will beat pure posterior methods in data-scarce, archive-readable design domains, while prior extraction loses in archiveless domains.Falsification includes posterior methods matching cold-start performance in archive-complete domains or hand-written content priors reviving in archiveless domains.
  • P3 (concept-layer invariance): P3 predicts that qualified frameworks retain an industry-invariant concept layer while marginal modeling costs decrease when onboarding new industries.The test uses public engineering records: legitimate purposes decompose into energy, matter, information, mechanical-work transfer, and spatial relations, while failures fit a closed typology.
  • P4 (the VLA ceiling): P4 predicts VLAs without semantic grounding will demonstrate progress yet remain below endorsed institutional deployment at scale.It is falsified if VLA-class systems achieve endorsed deployment in facility operations or industrial inspection without a constitutive semantic layer.
  • P5 (the irreplaceability of the ontology layer): P5 predicts that stronger LLM capabilities will increase, rather than eliminate, institutional demand for ontology-based AI retaining a prior cognitive framework.Replacement would require general-purpose LLM agents to deliver endorsed institutional decisions directly while bypassing ontology-based systems, beginning at the criterion’s exposed middle band.

Appendix note: the concept-layer boundary of the organizational domain (a structural corollary, not

The organizational domain has a triple structure: mission work lacks a shared cross-customer concept layer, while convergent functions reuse one. The paper closes by tying five predictions to distinct framework components as a Lakatos-style wager.

  • Organizational domain: At the mission layer, no cross-customer shared concept layer is expected; per-customer manual distillation remains the permanent cost.The passage characterizes mission work as the divergent zone and identifies forward-deployed engineering as the example mechanism.
  • Organizational domain: At the convergent layer, finance, compliance, and administration already share a reusable concept layer, exemplified by the ERP paradigm.The passage contrasts these functions with mission work and describes their shared layer as longstanding and reusable.
  • Closing: Lakatos’s wager: Each of the five predictions is linked to damage in a specific framework component: constraint axis, criterion, layering lower bound, normativity argument, or semantic anchorage–LLM complementarity.The closing frames these predictions as a Lakatos-style wager to be tested against how the world develops, rather than accepted merely for argumentative elegance.

7. Objections and Replies · 8. Theoretical Positioning and Related Work

The paper records a self-overturned criterion and answers six objections by distinguishing promulgated, archived norms from descriptive regularities, tacit practice, and learned representations. It positions the framework as a legitimacy test that changes eligible learning tasks from induction toward archive-based conformance checking, while delimiting its coverage.

  • 7. Objections and Replies: The initial temporal-priority criterion was rejected because genes precede proteins, yielding the stronger requirement of intentional constitution and a readable archive.The paper presents this self-overturning as evidence of methodological honesty.
  • 7. Objections and Replies: LLMs can recite specifications but statistically flatten “shall” statements, lack mechanisms for normative constraint, and contain representations that are not promulgated concepts.The paper therefore assigns LLMs the role of archive readers rather than normative archives.
  • 7. Objections and Replies: Agent systems themselves erect human-written system prompts, tool specifications, orchestration rules, and memory files that constrain model behavior as archived priors.This industrial scaffolding is presented as counter-evidence to the claim that models already contain the relevant prior framework.
  • 7. Objections and Replies: 32 physics-consistency indicators tested five frontier general-purpose models and found widespread failures in building-environment simulation tasks.The indicators covered temperature convergence, temperature-difference stability, sequencing, load response, and setpoint–flow feedback.
  • 7. Objections and Replies: AlphaFold shows that evolutionary data can provide structure without intentional constitution, leaving purpose and failure normativity inaccessible.The objection demonstrates that the archive conjunct may occur by happenstance while the intentional-constitution conjunct remains absent.
  • 7. Objections and Replies: Repurposing does not defeat the criterion because institutional purpose drift produces new design events, approvals, reviews, and reissued archives.The framework accepts that purely tacit knowledge lies outside its recorded-in-design scope.
  • 7. Objections and Replies: System-level accidents may be intrinsically unpredictable, but many apparent emergences remain detectable chains of small, observable, and explicable errors.The reply separates prior prediction from event-time detection and handling.
  • 7. Objections and Replies: The framework claims classification rather than a learning bound: where intentional constitution and readable archives hold, concepts are read from archives and deviations become faults in the world.It distinguishes descriptive physical laws and PINN constraints from promulgated norms, and presents a decidable boundary rather than a sample-complexity theorem.

R.1 The KR Tradition and Description Logics: Shared Conceptualization, or Constitutive Promulgation? · R.2 Normative Systems: The Oldest Formal Home of Promulgation · R.2’ Bayesian Priors: A Terminological Demarcation

These sections distinguish constitutive promulgation from KR’s neutral conceptualization, normative systems’ institutional norms, and Bayesian priors’ quantitative beliefs. They position the framework as an archive-grounded account that complements description-logics and normative-systems formalisms while sharply demarcating Bayesian terminology.

  • R.1 The KR Tradition and Description Logics: Shared Conceptualization, or Constitutive Promulgation?: KR’s ontology definition treats law promulgated to the world and summaries extracted from it equivalently, leaving epistemic status unspecified.The framework’s criterion asks whether domains are intentionally constituted and leave readable constitutive archives.
  • R.1 The KR Tradition and Description Logics: Shared Conceptualization, or Constitutive Promulgation?: OntoClean and formal ontology discipline concept meaning, while KL-ONE emphasizes subsumption structure without addressing why a lattice may precede experience.The passage contrasts meta-theoretic constraints and structural descriptions with the unresolved source of prior legitimacy.
  • R.1 The KR Tradition and Description Logics: Shared Conceptualization, or Constitutive Promulgation?: CYC’s failure is attributed to choosing common sense, an archiveless domain, whereas the framework’s diagnosis does not target circumscription itself.Circumscription addresses defeasible conclusions from incomplete common-sense axiom systems, a different problem.
  • R.1 The KR Tradition and Description Logics: Shared Conceptualization, or Constitutive Promulgation?: Description logics map closely onto the four-layer skeleton: signature, TBox, rule base, and ABox extended with streaming data.The implementation is realized in OWL/RDF, with inheritance and instantiation relations corresponding to the framework’s vertical edges.
  • R.1 The KR Tradition and Description Logics: Shared Conceptualization, or Constitutive Promulgation?: The traditions complement one another: description logics relate expressiveness to reasoning complexity, while this framework justifies four separations for systems that grow yet remain unified.The passage presents DL decidability results as mathematical support for the enforced layers.
  • R.2 Normative Systems: The Oldest Formal Home of Promulgation: Normative systems formalize promulgation through deontic logic and life cycles of creation, promulgation, compliance, violation, and sanction, rather than statistical averaging.The tradition is identified as the framework’s closest theoretical ally.
  • R.2 Normative Systems: The Oldest Formal Home of Promulgation: Normative MAS constitutive norms create institutional facts, whereas this framework constitutes normal states of the artificial physical world.Its criterion specifies when prior promulgative authority holds: domains must be intentionally constituted and documented with readable archives; normative machinery supplies deontic, lifecycle, and enforcement formalisms.
  • R.2’ Bayesian Priors: A Terminological Demarcation: Bayesian priors are quantitative belief distributions supplied by modeler beliefs, updated by Bayes’s rule, and designed to wash out as evidence accumulates.Their direction of fit runs from world to belief, unlike constitutive promulgation.

R.3 Industrial Ontologies: Engineering Forced into the Semantic Layer by Pain · R.3’ Safety Boundaries: ODD and SOTIF · R.4 Neuro-Symbolic AI: Our Closest Fellow Traveler in the Learning Community, and the Foundation It Lacks

The paper situates its framework among industrial ontologies, automotive safety boundaries, and neuro-symbolic AI, arguing that engineered systems already rely on promulgated semantics and normative limits. It distinguishes neuro-symbolic coupling from the prior question of where authoritative symbols come from and what legitimizes their constraints.

  • R.3 Industrial Ontologies: Engineering Forced into the Semantic Layer by Pain: Industrial engineering independently demonstrates that layered, promulgated semantics is a practical necessity rather than merely a philosophical preference.The paper presents three industrial signals as evidence that engineering practice has converged on semantic layering.
  • R.3 Industrial Ontologies: Engineering Forced into the Semantic Layer by Pain: Palantir’s Ontology combines semantic elements—objects, properties, and links—with kinetic elements—actions, functions, and dynamic security—in an operational closed loop.The paper identifies this as a deployed industrial ontology with a promulgated semantic layer at its core.
  • R.3 Industrial Ontologies: Engineering Forced into the Semantic Layer by Pain: Building automation created semantic layers such as Brick Schema and Project Haystack to integrate mutually incompatible control data.Brick Schema defines building entities and relations, while Project Haystack maintains a community-governed ontology of equipment and data points.
  • R.3’ Safety Boundaries: ODD and SOTIF: Automotive safety regulation uses ODD and SOTIF to define a promulgated normative boundary whose violation triggers a prescribed minimal-risk response rather than a model update.An operating instance either satisfies the declared domain or violates it, making the direction of fit normative by law.
  • R.4 Neuro-Symbolic AI: Our Closest Fellow Traveler in the Learning Community, and the Foundation It Lacks: Neuro-symbolic AI argues that statistical learning alone is insufficient for robust, interpretable intelligence and therefore combines neural learning with symbolic representation and logical reasoning.The paper calls this community its closest fellow traveler within learning science.
  • R.4 Neuro-Symbolic AI: Our Closest Fellow Traveler in the Learning Community, and the Foundation It Lacks: Neuro-symbolic research addresses how symbols and networks couple, whereas this paper addresses where symbols originate and by what right they constrain networks.The paper treats coupling style as a free design choice and locates promulgative authority in intentional constitution with a readable archive.

R.5 Agent Scaffolding: The Return of Promulgation in the Wild

The evolution of LLM engineering shows promulgation returning through increasingly explicit scaffolding, as posterior capability proved insufficient. Prompts, tool specifications, memory stores, and orchestration rules are human-written norms whose direction of fit is world-fits-document.

  • R.5 Agent Scaffolding: The Return of Promulgation in the Wild: LLM engineering progressed from prompt engineering to system prompts, tool specifications, persistent memory stores, and multi-agent orchestration rules.The progression added roles, protocols, and division of labor as increasingly explicit scaffolding.
  • R.5 Agent Scaffolding: The Return of Promulgation in the Wild: These artifacts are norms written by humans to constrain model behavior, not descriptions distilled from model behavior.They occupy the most extreme cell of the archive gradient because machines are their first readers.
  • R.5 Agent Scaffolding: The Return of Promulgation in the Wild: Their direction of fit is world-fits-document: a system prompt’s value lies in promulgating how the model shall act.This criterion distinguishes behavioral prescriptions from descriptive accounts of model behavior.

R.6 A Certified Agent Calculus: An External Inventory · R.7 Control Group: PINNs, and the Two Fates of Explicitness

The external inventory presents COIN as an independent confirmation of the four-layer lower bound while locating its certification boundary in intentionally constituted, documented systems and pre-promulgated symbolic structure. PINNs serve as a control group: both use explicit structure, but PINNs describe observed behavior, whereas constitutive semantics distinguish physical impossibility from normative fault.

  • R.6 A Certified Agent Calculus: An External Inventory: COIN formalizes agentic workflows as typed free-monad task spaces in a Grothendieck topos.The framework’s inventory identifies workflow signatures and monad composition laws with syntax, and typed task declarations with concept.
  • R.6 A Certified Agent Calculus: An External Inventory: COIN exhibits all four framework layers: syntax, concept, knowledge, and instance.The passage explicitly presents COIN as confirming the four-layer lower bound from an independent mathematical tradition.
  • R.6 A Certified Agent Calculus: An External Inventory: COIN’s instance layer relies on external human-run systems assumed to realize schemas and satisfy contracts.Examples include databases, session services, and organizational planning involving budgets and compliance requirements.
  • R.6 A Certified Agent Calculus: An External Inventory: COIN’s certification calculus is bounded by domains that are intentionally constituted and documented, with organizational cases showing grayness in norms and documentation.This boundary is described for external systems on which COIN’s instance layer docks.
  • R.6 A Certified Agent Calculus: An External Inventory: COIN presupposes promulgated symbolic structure: typed tasks draw on existing schemas, certificates rely on declared assumptions, and applicability requires explicit context.The passage states that the calculus falls silent where neither promulgated norms nor readable structure is available.
  • R.7 Control Group: PINNs, and the Two Fates of Explicitness: PINNs are a control group because they share explicit structure, skepticism toward black boxes, and respect for domain laws.Their similarity makes the distinction between descriptive and constitutive explicitness especially important.
  • R.7 Control Group: PINNs, and the Two Fates of Explicitness: PINNs’ explicit content is descriptive, extracted from observations of the world’s behavior, rather than promulgated.The comparison holds explicitness constant and flips whether the structure is promulgated.
  • R.7 Control Group: PINNs, and the Two Fates of Explicitness: Physics engines supply constraint boundaries, while constitutive semantics supply normative coordinates for normality and fault.The relation is complementary rather than competitive, and the decisive question is whether deviation violates the world or revises the model.

R.8 Closing: Four Communities, One Control

Four independent communities, joined by a parallel mathematical construction, converged on an explicit, promulgated, layered semantic layer. A control group isolates the decisive variable, while the routes to the physical world are measured using the criterion as instrument.

  • Convergence: Four communities independently converged on an explicit, promulgated, layered semantic layer despite different problems, objects, and agendas.The communities are engineering, learning science, normative-systems theory, and the LLM industry.
  • Convergence: A parallel mathematical construction independently converged on the same explicit, promulgated, layered semantic structure.The mathematical construction is identified as R.6.
  • Control: One control group isolates the decisive variable in the convergence analysis.The passage identifies this control group as R.7 but does not state the variable’s full description.
  • Measurement: The routes to the physical world are measured using the criterion as the instrument.This measurement is presented in Table 2.

9. Conclusion: The Artificial Physical World — Low-Hanging Fruit on the Road to AGI

The conclusion identifies design as the crack in the data–intelligence deadlock: the artificial physical world is intentionally constituted and documented, enabling framework-first learning through readable archives. It positions this world as the most tractable physical entry point while maintaining that complete intelligence requires all four worlds and leaving embodiment and framework carriers to companion papers.

  • Core conclusion: Design breaks the otherwise universal data–intelligence deadlock because valuable artificial physical artifacts are intentionally constituted and documented.The conclusion calls this exception a “crack” created by civilization’s designed artifacts.
  • Core conclusion: Only the artificial physical world permits a framework to precede its instances, requiring at least four layers: syntax, concept, knowledge, and instance.The syntax layer promulgates norms, while the concept layer preserves stability; the passage also identifies knowledge and instance layers.
  • Core conclusion: The other three worlds remain indispensable floors of complete intelligence, but only the designed world supports layer-by-layer and clause-by-clause review of understanding.The conclusion explicitly withholds a proposed path to AGI while requiring no missing floor if AGI is to stand.
  • Assurance and evidence: Assurance cases are buildable exactly where worlds are intentionally constituted and readably archived, supplying the ontological ground that engineering assurance methods lack.The criterion determines where claims about deployed systems can be pursued to explicit evidence.
  • Assurance and evidence: The designed world uniquely exposes its source code through drawings, making archive literacy the most economical route for machines entering the physical world.The conclusion characterizes the artificial physical world as a purpose-first domain whose drawings are openly available.
  • Open questions: Two companion papers address the intelligent body and the carrier for the framework’s knowledge and instance layers, while all three share the axiom that the artificial physical world is purpose-first.The present paper deliberately leaves both questions open.

Appendix A. Semi-Formal Statements of the Criterion and the Lower-Bound Propositions

Appendix A formalizes the legitimacy criterion, direction of fit, and layering lower bound as checkable predicates and propositions, while marking the limits of semi-formal treatment. It also states computational results distinguishing archiveless coverage, closed-layer failure reduction, and archive readability.

  • A.1 Basic Conventions: An object domain D comprises instances I, norms N, and a machine-readable archive A, with constitution requiring norms to precede and determine instance identity.Constitution is written N ◁ I and requires compliance with norms to distinguish an instance as that instance rather than another.
  • A.1 Basic Conventions: A deviation is promulgative when it is judged a defect of the world, but descriptive when it is judged a defect of the model.Promulgated direction is world-to-word; described direction is word-to-world.
  • A.2 The Legitimacy Criterion (Formal Skeleton of §4.2): The legitimacy criterion requires intentional constitution and a readable archive, yielding promulgative constraints, loud violation verdicts, and zero-shot operability before instance data.Under the stated assumptions, deviation is locatable and reportable, and the framework is fixed before the first instance arrives.
  • A.2 The Legitimacy Criterion (Formal Skeleton of §4.2): Among four worlds, only the artificial physical world satisfies intentional constitution and readable archival conditions simultaneously.The phenomenal and basic physical worlds fail constitution, while the artificial symbolic world does not bind physical instances.
  • A.3 The Layering Lower-Bound Proposition (Formal Skeleton of §5.4): The layering proposition establishes a lower bound of four carriers because the four construction goals form contradictory pairs across time and denotation.The result forbids compression below four layers but does not establish four-layer sufficiency; implementations may subdivide within layers.
  • A.3 The Layering Lower-Bound Proposition (Formal Skeleton of §5.4): Inter-layer relations are exclusively abstraction and instantiation, with layer-by-layer collapse defining both the reduction path and the explanation path.Knowledge instances instantiate concepts, concepts instantiate syntax word classes, and instance objects instantiate knowledge entries.
  • A.5 A Computational Triad: The Complexity of Coverage, Reduction, and Extraction: For an open, nonconstituted, undocumented world, no finite or recursively enumerable rule set can be confirmed to cover its normative structure from positive examples alone.The appendix frames this as a Gold-type computability boundary rather than merely an engineering shortfall.
  • A.5 A Computational Triad: The Complexity of Coverage, Reduction, and Extraction: A finite closed concept layer makes failure localization decidable in polynomial time, while extraction ranges from O(n) parsing for purpose-written archives to NP-hard causal discovery for happenstance-readable archives.The reduction depth is bounded by d = 4, whereas no archive returns extraction to the open-world coverage problem.

A.4 A Boundary Theorem: The Applicability Domain of Certificate-Anchored Calculi

Certificate-anchored calculi apply legitimately only where tasks have typed contexts extracted from and traceable to promulgated readable archives. Without intentional constitution or a readable archive, their soundness holds at most vacuously because certificate chains lack anchored declarations.

  • Archive anchoring: A typed task context is legitimately grounded if every type object and interface declaration comes from a promulgated readable archive and traces to one of its clauses.This is the archive-anchoring definition for task T in world W.
  • Boundary of certificate chains: If a world is not intentionally constituted or lacks a readable archive, no task there has a legitimately grounded typed context.The boundary follows directly from the archive-anchoring requirement.
  • Boundary of certificate chains: In such worlds, a certificate calculus’s soundness theorem remains true only vacuously because its typed-context and anchored-certificate antecedents cannot be legitimately satisfied.Certificates require anchored declarations, so absent an archive the terminal link of every certificate chain is unavailable.
  • Applicability domain: The proposition preserves the calculus’s mathematics while mapping which worlds can legitimately supply its antecedents: intentionally constituted and documented worlds.The legitimacy criterion identifies the applicability boundary rather than invalidating the conditional theorem.

Appendix B. The Failure-Mode Type Vocabulary (Excerpt) … Appendix C. Framework Demonstrations

Appendix B presents an enumerable failure-mode vocabulary organized into distinct engineering categories, including physical-system, medium-state, transfer, and environmental mechanisms. The excerpt also specifies transfer types and illustrates how deviations from designed conditions define failure classes.

  • Appendix B. The Failure-Mode Type Vocabulary (Excerpt): The concept-layer excerpt reduces an industry’s object system to six object classes and presents the vocabulary as engineering-inventory evidence that types are enumerable.The excerpt adds implementation facts not carried in the main text and unfolds the framework’s four-layer shape elsewhere.
  • B.1 Physical-System Failure Mechanisms (10 classes): 10 classes define physical-system failure mechanisms as interruptions, degradation, or destabilization caused by structural, functional, flow, or control anomalies.The category covers modes by which systems, equipment, or components lose designed function.
  • B.2 Medium-State Failure Mechanisms (9 classes): 9 classes define medium-state failures as composition, property, or phase deviations that reduce a system’s functional support capacity.The failures concern state changes in a system’s internal medium.
  • B.3 System Transfer Factors (4 classes): 4 transfer classes divide cross-system supply anomalies by transferred physical category: energy, matter, information, or mechanical.The earlier transfer-type × transfer-state scheme is collapsed by treating interruption as an extreme deviation and carrying severity in the four-grade failureDegreeProfile.
  • B.3 System Transfer Factors (4 classes): The transfer vocabulary distinguishes matter, energy, information, and mechanical transmission by their carriers, meanings, or transmitted forces and loads.Examples include chilled-water flow, supply voltage, control commands, and transmission torque.
  • B.4 Environmental Factors (9 classes): 9 environmental-factor classes capture physical interference at equipment boundaries that exceeds design tolerances.Examples include temperature, humidity, mechanical loads, electromagnetic fields, spatial constraints, medium intrusion, foreign objects, chemical exposure, and radiation.
  • B.4 Environmental Factors (9 classes): Environmental examples include foreign-object intrusion, alongside anomalies involving temperature, humidity, mechanical conditions, electromagnetic exposure, space, media, chemicals, and radiation.These conditions alter operating boundaries or exceed equipment tolerance and design protection capability.

Preamble: The Existence Claim · The Demonstration Register: Selection, Disclosure Status, and Commitments

The paper claims that constitutive prior frameworks already operate across five industrial domains and presents a publicly testable cross-domain invariance commitment. Its four-case demonstration register distinguishes reasoning capability from industrial coverage while openly disclosing that the reports remain incomplete and withheld.

  • Preamble: The Existence Claim: Prior cognitive frameworks are claimed to operate across five major domains, including civil buildings, manufacturing, advanced installations, municipal infrastructure, and data centers.The passage identifies examples such as shopping malls, semiconductors, a nuclear-fusion device, district heat networks, and data centers.
  • Preamble: The Existence Claim: Across industries, the concept layer allegedly migrates unchanged, with new industries adding blocks rather than changing the framework’s shape.The passage states that marginal modeling cost decreases as the number of industries increases.
  • Preamble: The Existence Claim: New-industry onboarding has reportedly required instantiating known failure types rather than inventing new ones.This claim is presented as part of the framework’s cross-domain invariance proposition.
  • Preamble: The Existence Claim: The framework’s cross-domain invariance is a publicly testable falsifiable proposition: modifying the concept layer during onboarding would count against it.The passage identifies this commitment as Section 6, P3, though the supplied text truncates the proposition’s final wording.
  • The Demonstration Register: Selection, Disclosure Status, and Commitments: The demonstrations form a four-case register divided into two tiers addressing whether a legitimate prior framework can reason and related questions.The first tier includes C.1, the Curiosity Mars rover drill-feed-mechanism anomaly on Sol 1536 in December 2016.
  • The Demonstration Register: Selection, Disclosure Status, and Commitments: C.1 is both a reasoning demonstration and a limiting case involving a single physically inaccessible instance with no failure train.Its event record comes from NASA’s public mission materials.
  • The Demonstration Register: Selection, Disclosure Status, and Commitments: The four case reports are incomplete in this version, while the register and C.1’s event record are already fixed and public.The reports are withheld as a set pending de-sensitization and patent applications.
  • The Demonstration Register: Selection, Disclosure Status, and Commitments: The register is presented as checkable rather than promissory because its selection rationale, public C.1 record, and theory-bearing claims are disclosed.The stated claims concern the four-layer skeleton’s reasoning chain and Appendix B’s concept-layer coverage of new industries.
Loading 2608.15147v1…