The World Model Initiative is the founding research program of GrytLabs Research Institute, and this volume is its bound output. It advances a single finding: governance is a completeness problem, not a capability problem. The autonomous systems being deployed are missing structural ingredients that governance requires — constraints that bind, purpose that is carried, memory that is attributable — and those ingredients cannot be produced by the mechanism the field relies on to improve, because they are declared structure, not statistical regularities to be discovered. The volume establishes that finding in the strict audit sense — a corroborated condition, measured against a stated standard, traced to its cause, with a demonstrated effect and a feasible corrective action — across six Technical Reports forming the synthesis stratum: one Dialogue report (TR-D-001, the world-model definition), four Architecture reports (TR-A-001 through TR-A-004 — the structural gap, the architectural necessity, the authority architecture, and the externalization path), and one Landscape report (TR-L-001, convergent validation and architectural extension). Beneath them sits the evidence stratum: twenty Research Reports (RR-001 through RR-020) distilling the founding-period sprints, each carrying its claims back to the primary literature. This binder brings the whole volume together in one place — the finding and its argument, the reading order and the two strata, the sixteen-position catalog, the evidence-report series, and the descent from each synthesis claim to its supporting evidence. This is the first volume of the WMI thesis cycle — the founding period — succeeded by the firm thesis.

The art through this series is meant to argue, not decorate.

The visual grammar draws on the spirit of Jean-Michel Basquiat — the crown that coronates rather than describes, the word struck through so you look at it harder, text carried as substance rather than caption. It borrows his logic, not his hand. He set crowns over figures the world had failed to authorize, conferring worth by declaring it — which is, in a single mark, the argument these volumes spend their length making: that authority is declared, not discovered. Before the paintings he was a graffiti poet writing terse, declarative aphorisms on walls — language as the primary act, text constituting meaning in public space.

The grammar recurs throughout the series, and it reads the same wherever it appears. A crown over the figure is authority conferred by declaration. A word struck through — capability, low and crossed out — is the thing the work asks you to look past: present, negated, demoted beneath what governs it. Structure shown beneath the surface, a single thread from head to ground, is lineage — the path from purpose to act, drawn because it cannot be read off behavior. Memory anchored at the base is foundational, opposing where intelligence is assumed to live.

The quote at the front is chosen with the same intent. The whole livery line — Basquiat's uniformed rank, bowing in unison — set beside the present moment, the two resolve into one image. Employees are asked into AI training and, in the training, hand over the skill that took a career to earn: they teach the machine to stand where they stand. Beneath them, a piece-rate multitude labels and verifies the data the systems feed on — the wage itself broken into tasks and paid by the piece. To bow is to hand up what you know and receive wages for the posture. And the big money all crushed into these feet is the whole weight of it: capital and capability pooling in the same few hands, bearing down on those who hold the line.

The Initiative inverts the geometry. It puts the capability into the laborer's own hands, in a form they can use — abstracting the difficulty, not the value. The tool carries the technical burden; the worker keeps the judgment, the skill, and the world that is theirs. The crown set over the world-bearer is authority conferred by declaration: the laborer authorized to hold their own world model, not to bow to one held above them.

Autonomous artificial intelligence is being deployed without governance infrastructure. Ungoverned systems are bringing about consequential events that no one authorized and no one verified — deleting production databases, executing unapproved purchases, completing trajectories with no accountable record of who sanctioned them. Capability continues to advance rapidly while governed autonomy does not keep pace.

Agentic systems can already act effectively across tasks that once required human reasoning, perception, learning, and language; capability is not in question. But improving a system's ability to predict and act moves it no distance toward being accountable. What fails is everything that would let those actions be traced to a source of authority, accounted for after the fact, and held within constraints that do not attenuate as the system scales.

The failure has a precise location, and naming it requires separating two things that share the name "world model." A predictive world model is a learned model of how the world behaves; its facts are true by correspondence — they hold when they match observation. A constitutive world model is the declared structure of what an organization is — its entities, authorities, purposes, and constraints; its facts are true by declaration — they hold because an authority conferred them. Prediction proper — a distribution over what comes next — is pure correspondence and needs no governance. But no deployed system stops at prediction: it collapses that distribution into an action, one continuation selected, and the selection is a choice against a criterion, which is a governed act. Governance becomes unavoidable the moment a system acts rather than merely predicts.

A system acting autonomously does not set the standard it acts under — that authority is conferred from outside and is not the system's to author. What autonomy requires is the opposite: that the system carry an externally declared standard reliably while unsupervised, bound to it structurally rather than merely understanding it. Governability is that binding, and it requires structural ingredients training cannot produce: constraints that bind, purpose that is carried rather than inferred, memory that is attributable. None of these is a statistical regularity waiting in data to be discovered; they are declared structure, constituted in language by an authority.

Capability and governance are therefore orthogonal. The predictive model supplies correspondence with the discoverable world; the constitutive model supplies authority over the declared one; any system that acts needs both. Scale moves the first axis and cannot move the second — because there is nothing on the authority axis to learn, its facts being conferred rather than observed. This is why the gap does not close as systems improve: it is not a capability gap.

Governance and the constitutive world model are not two systems in a dependency but one object under two grammars. Governance is the evaluation of acts against a constitutive world model; the constitutive world model is the structure that governance both presupposes and produces — the structure frozen, and the structure in operation. They are meaningful only together, and load-bearing wherever action proceeds under a declared standard — wherever roles, permissions, rules, or commitments govern what may be done.

The effect is not merely error but uncatchable error: an unauthorized action and an authorized one can be behaviorally identical, separated only by a declaration that was or was not made, so the failure returns no signal for capability to detect — coherent output over severed lineage. The corrective action follows from the cause. The constitutive layer must be built as infrastructure: the declared structure made computable, and the thread from authority to individual act kept continuous and inspectable.

This is a finding in the strict sense: a corroborated condition, measured against a stated standard, traced to its cause, with a demonstrated effect and a feasible corrective action. The remainder of this volume is its evidentiary body.

Condition. Governance-prescriptive knowledge fails to become operational infrastructure. Independent research traditions — data provenance, AI governance, organizational memory, accountability theory, the semantic web, audit methodology — each built apparatus prescribing how governed systems should behave, and each arrived at the same structural boundary: the point at which the prescription cannot be made to execute. The convergence is reached from different starting points, by different methods, by communities not in contact — which is what makes the boundary structural rather than an artifact of any one field's blind spot. Established in TR-A-001, and corroborated again in the founding-period frontier validation collected in TR-L-001.

Criteria. Governance is the evaluation of acts against a constitutive world model; the constitutive world model is the structure that governance both presupposes and produces; the two are one object under two grammars, meaningful only together, and load-bearing wherever action proceeds under a declared standard. Against that standard, governance infrastructure for intelligent systems must satisfy, at minimum, three structural requirements: authority must be traceable to its source; constraints must hold as the system scales rather than attenuating; and the governing structure must remain viable as it recurses. These are not preferences. They are the conditions under which delegation can be accountable at all. Established in TR-A-003.

Cause. The gap is architectural, not disciplinary. A complete world model requires ten ingredients; the systems in question supply three — and the universally missing ones are precisely Constraints, Purpose, and Memory, the ingredients governance is built from. They are missing because the field recognizes only one mechanism of model creation. Training discovers statistical structure from data; it cannot produce constitutive constraints, carried purpose, or attributable memory, because those are not statistical regularities to be discovered — they are declared structure. Independent results from mathematical physics, machine-learning theory, geometric deep learning, management cybernetics, and cognitive science each establish, from their own foundations, that the processes relied on for improvement systematically destroy the very invariances governance requires. The property cannot be trained in; it must be architecturally given. Established in TR-A-002. The formal definition — what the ten ingredients are, why each is irreducible, and how the three mechanisms relate — is articulated in TR-D-001.

Effect. A system that can predict but cannot govern acts without being able to be held to account. The failure is not merely that errors occur but that they are uncatchable: an unauthorized action and an authorized one can be behaviorally identical, separated only by a declaration that was or was not made, so no advance in capability returns a signal that tells them apart — coherent output over severed lineage. This is the operational failure the field is now experiencing in the open: capable systems taking consequential, unverifiable actions in settings where the cost of error is a patient, a client, or a child. The operational consequences are established primarily by the frontier failure cases collected in TR-L-001 — named incidents of capable systems acting unaccountably, with consequence; the research corpus establishes the structural conditions under which those incidents are predictable rather than isolated.

Recommendation. The corrective action follows directly from the cause: because governability cannot be trained in, it must be built in. The constitutive layer must be constructed as infrastructure — the declared structure of authority, purpose, and constraint made computable, with the thread from a source of authority to an individual act kept continuous and inspectable, and the governing structure able to recurse without losing its binding force. This is feasible rather than aspirational. The structure required is not bespoke to each domain: the audit tradition demonstrates that governance-level understanding is domain-invariant, so the infrastructure can be specified once and instantiated across domains rather than rebuilt for each. It is therefore deliverable as open structural infrastructure — able to be acted on by any party rather than held behind a single implementer. Established in TR-A-004.

The volume's claim is one finding, not seven claims. Each report is where one element of that finding is established to evidentiary standard, and each stands on the deeper evidence stratum that carries its claims to the primary literature.

TR-D-001 — The World Model Definition is the definitional foundation. The finding's Cause element appeals to a ten-ingredient framework and a three-mechanism taxonomy; TR-D-001 is where both are formally articulated. It establishes what a world model must contain to be complete, evaluates six leading definitions against the ten-ingredient standard, and shows that the universally missing ingredients map precisely to the universally missing mechanisms. The definition is not an assumption the Architecture-class reports inherit — it is a claim they cite, and this report is where the claim is made. Anchored to Conant & Ashby's (1970) regulator theorem.

TR-A-001 — The Structural Gap is the condition. It establishes that the gap is real by showing six independent traditions — data provenance, AI governance, organizational memory, accountability theory, the semantic web, audit methodology — reaching the same boundary, and derives three design requirements from the boundary's structural properties. Convergence from unrelated starting points is what makes the condition a finding rather than one tradition's complaint.

TR-A-002 — The Architectural Necessity is the cause. It establishes that the gap cannot be closed by training, scaling, or computation — that the improvement processes themselves destroy what governance requires — and therefore that the missing invariances must be supplied by architecture. Convergent evidence from mathematical physics, machine-learning theory, geometric deep learning, management cybernetics, and cognitive science.

TR-A-003 — The Authority Architecture is the criteria. It establishes what the governance layer must structurally satisfy — authority traceability, constraint management, and recursive viability — drawing three independent traditions onto equivalent requirements, and identifies a structural correspondence between COSO's five internal-control components and Beer's five viable systems.

TR-A-004 — The Externalization Path is the feasibility of the recommendation. It establishes that the resolution can actually be built, and that the audit tradition's century of cross-domain assessment is the evidence that governance extraction generalizes across domains rather than fragmenting into bespoke solutions. The corrective action is delivered as open structural infrastructure.

TR-L-001 — The Organizational World Model completes the volume. It extends the condition's corroboration into the frontier convergent validation that coincided with the founding commitment (RA-015), and extends the architecture into boundary, multi-agent, hybrid-mechanism, language, and workforce domains (RA-016 through RA-020). It carries the primary evidence for the finding's effect.

The Effect is the one element not assigned to a dedicated report, because it is not a derived claim but an observed consequence. Its primary evidence is the frontier failure record in TR-L-001 — named incidents of capable systems acting unaccountably, with consequence — while the research corpus (RR-001 through RR-020) establishes the structural conditions under which those incidents are predictable rather than incidental. The sequencing is deliberate: the gap is proven and explained at corpus depth before the operational failures can be read as its signature rather than as isolated events.

Each report stands on its own. But the volume's claim is not four claims — it is one finding, and the reports are where each of its elements is established to evidentiary standard.

The volume has two layers, and this section makes both visible and navigable.

Two strata. WMI Volume I presents six Technical Reports as its synthesis stratum — the layer where the volume's single finding is stated and established. Beneath them sits the evidence stratum: the twenty-report research series the synthesis is built from. A reader can take the volume at the synthesis layer alone and get the complete finding; a reader who wants to verify it, or follow a claim to its source, can descend into the evidence. Every claim in a synthesis report descends to the research reports that establish it, and each research report carries the citation chain to its primary sources.

> Synthesis stratum — the six Technical Reports. The finding stated and established.

> ↓ rests on

> Evidence stratum — the Research Report series (RR-001 … RR-020). The corroborating research, carrying each claim toward its primary literature.

Reading paths. There is no single correct order; choose by what you are here to do.

• The linear path — read the finding as it is built. TR-D-001 (vocabulary) → TR-A-001 → TR-A-002 → TR-A-003 → TR-A-004 (condition → cause → criteria → recommendation) → TR-L-001 (frontier validation and extension). This is the recommended first pass. The order is intellectual, not chronological; it follows the argument arc of the finding.
• The path by concern — enter where your question lives. Each Architecture report stands alone. A reader who already accepts the gap and wants the cause can enter at TR-A-002; one who wants what a governance layer must satisfy can enter at TR-A-003; one who wants feasibility can enter at TR-A-004. Consult TR-D-001 for any unfamiliar vocabulary (the ten ingredients, three mechanisms).
• The deep path — descend a claim to its evidence. Take any synthesis report, find it in the Crosswalk below, and follow it down to the research reports that establish it; from there, each research report carries the citation chain to its primary sources. This is the path for verification, for prior-art tracing, and for readers who want the substrate rather than the synthesis.

Position identifiers. Each report's §5 Position Statements cites the convention WMI-P{NN}, with an explicit disposition (strengthens-refines, promotes-from-SCT, or qualifies-falsifies) and the supporting research findings. The full catalog is in The Positions, below.

For readers approaching this volume from outside the institute's publication program: the reports present a self-contained finding about governance infrastructure that does not require familiarity with the institute's internal substrate. The position identifiers and citation chain provide traceability for readers who want to evaluate the evidence base; they are not prerequisites for following the argument.

A structural finding cannot be validated the way a procedure is validated. There is no experiment to re-run; the claim is that a real structural boundary exists, and the evidence that it exists is that independent parties keep arriving at it. The discipline this volume applies is convergent evidence from independent sources, evaluated separately at each element of the finding (WMI-P12). The traditions that corroborate the condition are not the traditions that establish the criteria, and neither is the tradition that independently derives the feasibility of the recommendation. Convergence is treated as evidence for the specific element its sources bear on — not borrowed across elements — because that is the only form in which convergence carries weight rather than coincidence. The Effect is evidenced differently from the other elements — by the operational record of failure rather than by convergent traditions — because a consequence is demonstrated by instances of its occurrence, not corroborated by independent parties reasoning to the same place.

This is the audit tradition's own evidentiary standard — sufficient, appropriate evidence corroborated by independent sources — applied to a research claim. It is not imported from outside the institute's practice. It is the institute's practice.

A finding whose recommendation no existing entity can implement is incomplete, and a recommendation issued without the means to act on it is negligent (WMI-P14). When the research that identifies a structural gap also establishes that no current tooling can close it, the institute that issued the finding is obligated to demonstrate — not merely to describe. The corrective action is therefore delivered as open structural infrastructure (WMI-P16): the value is distributed through an ecosystem that can build against the same grammar, not held as proprietary advantage, and the choice to make human authority explicit and structurally isolable from AI capability is the volume's primary ethical commitment, enforceable because it is architectural rather than aspirational (WMI-P15).

This is what distinguishes the entity issuing this finding from the categories it resembles. It is not a standards body, an audit firm, an analyst, or a consultancy. It is a research institute that produces the finding and is bound to make the remedy buildable — and the same discipline that makes the finding re-performable is what makes the remedy verifiable by parties with no connection to the institute.

The founding thesis is articulated through sixteen positions — current hypothesis statements, stronger than suspended thoughts but not yet constitutional. They are asserted, not proven. The publication program is the mechanism by which the institute earns credibility for these assertions or discovers they are wrong. The positions are organized in four layers that build from object to commitment, and each Technical Report's Position Statements section (§5) cites these identifiers; this catalog is the reference that makes them legible.

WMI-P01 · Governed Composition. A world model is a governed composition of data models, computational models, and conceptual models. The ten ingredients are the formal requirements. The three mechanisms (Training, Constitution, Accretion) are how those ingredients come into existence.

WMI-P02 · Governance Model Vocabulary. The field conflates at least six uses of "model." A seventh — the governance model — is missing from the discourse entirely. It is the structural layer that makes governed composition possible: the irreducible vocabulary required to parse organizational reality into governable form, combined with the discipline of administering a world model instance.

WMI-P03 · Originality of the Ten-Ingredient Framework. The ten-ingredient world model concept is original to GrytLabs. No prior literature defines it at this level of formality.

WMI-P04 · Linguistic Constitution. Three of ten ingredients — Relations, Constraints, and Purpose — are linguistically constituted in organizational domains. They exist because people declared them through speech acts, not because they were observed. Training cannot produce them.

WMI-P05 · Organization as Verb. The organization is an ongoing process with its own physics. The world model captures that process, not a static snapshot.

WMI-P06 · DLP as Compiler Toolchain. The Decision Lineage Protocol is a compiler toolchain for organizational intelligence — grammar, intelligence layer, and runtime — not an application or platform. World models are governed compilations of organizational reality.

WMI-P07 · Anchor Outside the Learning Loop. Without an external grammar that is not absorbed into the model, the system cannot distinguish learning what is true from learning what is persuasive. A world model requires a structural anchor independent of any single instance or model.

WMI-P08 · Four Actor Types. The four canonical actor types — Human, AI-Automation, AI-Agentic, and AI-Assistive — are governance categories, not product categories. They distinguish how authority is delegated and how actions are gated.

WMI-P09 · Socket Not Plug. The protocol defines the structural interface without binding to any single computational model. Runtime agnosticism keeps world model instances viable across architectural revolutions in AI capability.

WMI-P10 · Knowledge Asymmetry Inversion. A governance grammar that makes structural vocabulary visible at all depth tiers replaces information asymmetry with authority asymmetry — consciously designed rather than accidentally created.

WMI-P11 · Human-Ceiling Problem as Controls Problem. Every AI output currently requires human review because no control environment exists to make delegation accountable. The resolution is not removing the human prior but making it visible, governed, and experimentally isolable. The shift from substantive testing to controls testing is the central methodological insight.

WMI-P12 · Convergent Evidence Methodology. The evidence methodology for a structural finding is convergent evidence from independent sources, evaluated separately at each element of the finding structure (condition, criteria, cause, effect, recommendation). The traditions that corroborate the condition are not the same traditions that establish the criteria or independently derive the recommendation.

WMI-P13 · Subjective Experience as Architecture Analysis. The exclusion of subjective experience from world model definitions is architecture analysis, not metaphysics. The governance implications — distinguishing honest architectural reporting from performed interiority — are engineering-critical.

WMI-P14 · Corrective Action Obligation. A recommendation without the tools to take corrective action is negligent. When research identifies a structural gap and no existing entity can address it, the institute that identified the gap has an obligation to build — not merely to describe. The protocol is open structural infrastructure so that any party can act on the recommendation.

WMI-P15 · Architectural Decision as Ethical Decision. The choice to build a governance substrate where human authority is explicit and experimentally isolable from AI capability is the primary ethical commitment. An architecture that ensures governance choices are structural makes the ethics enforceable rather than aspirational.

WMI-P16 · Open Protocol Architecture. Open protocol architecture means competitors can build against the same grammar. The value is distributed through the network, not locked in proprietary advantage. A protocol's success is measured by ecosystem diversity, not originator market share.

The volume's sixteen WMI positions (WMI-P01..P16, organized in four layers) receive coverage across the bound reports as follows. All founding-period dispositions are strengthens-refines — no position is falsified or qualified by the founding-period evidence.

Coverage notes. TR-D-001 and the four Architecture reports cover twelve positions directly; the remaining four (WMI-P04 supporting, P06, P13, P16) receive their primary treatment in TR-L-001. The two methodology-layer positions (WMI-P11, WMI-P12) receive primary treatment in both the first and last papers of the arc — the structural gap that motivates the controls-testing methodology (TR-A-001) and the externalization path that demonstrates its feasibility (TR-A-004) — by design; the methodology is both the founding observation and the resolution path, and the cross-arc reinforcement is deliberate.

The synthesis rests on a series of twenty Research Reports (RR-001 through RR-020). Each distills a research sprint, carrying its findings toward the primary literature — peer-reviewed sources, standards, and regulations — from which they were drawn. The series is part of the body of work, not a supplement to it.

Every founding-period sprint, its reader-facing Research Report, and the Technical Report(s) that cite it. Each synthesis report descends to the research reports that establish its claims; the finding element each carries is fixed and shown here. Auditable against rwp-wr-rs and rwp-wr-pub.

The volume's evidence operates at three layers, each independently citable with its own DOI:

Layer 1 — Technical Reports (TRs). The six bound TRs are the reader-facing synthesis artifacts. Each has its own DOI. TRs cite Research Reports, not RA-pubs directly.

Layer 2 — Research Reports (RRs). RR-001 through RR-020 are the reader-facing rendering of the founding-period sprint research, each with its own DOI (catalogued above). A TR reader follows in-text references to the RRs.

Layer 3 — Research Artifact Publications (RA-pubs). Each ra-NNN-wp.md is the L0 public version of the underlying sprint workpaper, governed by PUB-RA-001. Each RR carries a "Research Provenance" pointer to its RA-pub, which in turn carries the full citation chain to the original peer-reviewed sources, standards, and regulations.

Reader navigation: Volume → TR (by DOI) → RR (by DOI) → RA-pub → primary sources. Every claim in a synthesis report traces back through this chain to its primary evidence.

Terms used across multiple bound reports. The institute's grammar primitives and behavioral invariants are defined in [grytlabs-grammar]; this glossary covers the volume's recurring vocabulary.

• World model (constitutive). The declared structure of what an organization is — its entities, authorities, purposes, and constraints. Its facts are true by declaration. Distinct from a predictive world model (a learned model of how the world behaves; facts true by correspondence).
• The ten ingredients. Distinctions, Entities, Relations, Time, Transition Logic, Constraints, Observation, Uncertainty, Purpose, Memory — the formal requirements of a complete world model (TR-D-001).
• The three mechanisms. Training, Constitution, Accretion — how a world model's ingredients come into existence (TR-D-001).
• Governance. The evaluation of acts against a constitutive world model. Governance and the constitutive world model are one object under two grammars.
• The finding (five elements). Condition, Criteria, Cause, Effect, Recommendation — the audit-form structure carrying the volume's single finding.
• Convergent evidence. The evidentiary discipline (WMI-P12): independent sources reaching the same structural boundary, evaluated separately at each element of the finding.
• Synthesis stratum / evidence stratum. The bound TRs (synthesis) and the cited RR series (evidence) the synthesis rests on.
• Thesis cycle. The cyclical model in which each period holds an intellectual position (a thesis); the WMI thesis is the founding cycle, succeeded by the firm thesis.
• WMI-P{NN}. Position identifiers thesis-anchored to the WMI thesis; the catalog is in The Positions.