WMGF-001 · Working Draft v0.5.0

The World Model Governance Framework

Governance is a completeness problem, not a capability problem. Here is the structure any governed system needs — and how to test that it is there.

10Ingredients

×

19Primitives

=

190Control objectives

Fig. A · a governable system, labeled like anatomy

01 Prose + practice

Read it as prose — and put it to practice.

Plain-language guidance for governing any system that claims to model the world — a research substrate, an autonomous agent, a robotics platform, or an organization governing itself. It answers two questions in one move: what structural commitments does such a system need, and how do you verify each one is actually present?

And it is not only prose and theory. The Gryt demo puts the framework in code and practice — a working governed substrate you can open and inspect, where the structure is something you watch operate, not just something you read about.

Read the Framework Run it in code

02 The problem

The connective tissue nobody specifies

COSO, ISO, and NIST tell you that a decision must be authorized, that a requirement must win a conflict, that an artifact must prove a control operated. None specify the connective tissue that makes those things locatable in a real system: who authorized this, where the decision lives, which requirement wins, what proves it ran, how failure resolves. That structure is assumed everywhere and built nowhere — and it cannot be trained or optimized into existence, because it is declared architecture, not a statistical regularity waiting to be discovered.

▸ auditorthe control objectives your frameworks presuppose have no instantiation surface.

▸ builderthe parts that make your system explainable after the fact don't come for free — you put them there on purpose.

03 Mechanism · the spine

How it works: 10 × 19 = 190

The framework is a single governance matrix. Ten ingredients every world model must have, down the rows. Nineteen structural primitives from the Decision Lineage Protocol, across the columns. Every intersection is one named control objective — a single “shall” statement — for a total of 190 (Matrix D).

Five governance disciplines shown as overlay — the door, not a partition of the room

Authority Mapping

Decision Boundary

Constraint Propagation

Evidence Obligations

Mechanical onlyintent · work · capacity

Accountability Resolution

Behavioral

Structural

Mechanical

Temporal

01

02

03

04

05

06

07

08

09

10

11

12

13

14

15

16

17

18

19

1Purpose

2Distinctions

3Entities

4Constraints

5Time

6Observation

7Relations

8Transition Logic

9Memory

10Uncertainty

One objective pushed back. The Memory-ingredient cell that expected structural invariances to be learned — five research traditions converge that they cannot be. The next revision amends it rather than defends it.

155 evidence-grounded 35 coverage gaps 1 challenged

S Structural · Is it declared?B Behavioral · Does it act as declared?T Temporal · Does it hold over time?M Mechanical · Can a machine check it?

The authoritative test surface is the full 19-column, 190-objective matrix. The five disciplines below are the on-ramp to it — not a substitute for it.

01 · The forcing function

No cell left blank

Every cell must carry a declared position — no ingredient ungoverned. The discipline that turns a checklist into a completeness guarantee. Architectural precedent: OMB Compliance Supplement, Part 3.

02 · Conformance

S / B / T / M

Four native categories test the 19 primitives: Structural (is it declared?), Behavioral (does it act as declared?), Temporal (does it hold over time?), Mechanical (can a machine check it?).

03 · The governance floor

MVGP

Before any system-specific governance, a Minimum Viable Governance Posture always applies — three universal domains on one baseline control framework (COSO) and one provenance bridge (W3C PROV + BFO).

peoplemoneyknowledge

04 The COSO on-ramp

Five governance disciplines

A navigational overlay — not a partition, and not the test surface. The five disciplines group the primitives and map onto COSO Internal Control's five components, giving a practitioner a familiar door into the matrix. Sixteen primitives group under a discipline; three — intent, work, capacity — have no COSO home and are tested directly through Mechanical conformance.

01Authority MappingWhere the right to act comes from, and who is accountable for it.COSO · Control Environment

02Decision BoundaryWhere a decision is made, and what it is allowed to touch.COSO · Risk Assessment

03Constraint PropagationHow a rule binds, travels, and wins a conflict.COSO · Control Activities

04Evidence ObligationsWhat artifact must exist to prove the system operated as claimed.COSO · Information & Communication

05Accountability ResolutionHow failure is surfaced, owned, and resolved.COSO · Monitoring

The disciplines are the door. The matrix is the room.

05 The rows of the matrix

The ten ingredients

Every world model requires all ten. These are the rows of the matrix. Miss one and the system has a blind spot governance can't reach.

01Purpose

What the system is for, declared where it can be pointed at.

02Distinctions

The categories it draws and treats as real.

03Entities

The things it tracks, and how they're registered.

04Constraints

The rules that bind it, and which one wins.

05Time

How it stays continuous and current across change.

06Observation

How it takes the world in, and on whose authority.

07Relations

How its things connect and commit to each other.

08Transition Logic

How one state becomes the next, legibly.

09Memory

What it retains, and what it is architecturally allowed to hold true.

10Uncertainty

What it knows vs. declares vs. derives vs. cannot see.

Uncertainty is graded by four truth types — authoritative · declared · derived · opaque — so the framework can tell knowledge from assertion.

06 The evidence spine

Written from conviction. Grounded in evidence.

A framework written from conviction is still just assertion until the evidence has its turn. Before this draft went out, the founding research corpus — 20 Research Reports distilling research sprints across roughly six decades of literature — was walked across all 190 control objectives: 412 audit-style findings against 190 objectives.

82^%

155 of 190 control objectives are evidence-backed by an auditable corpus

For those, the framework does not merely claim; it cites.

One objective pushed back

The research challenged exactly one control objective — the Memory-ingredient cell that expected structural invariances to be learned from accumulated evidence. Five independent research traditions converge that they cannot be: invariances must be architecturally imposed. A framework that engages its own counter-evidence is more trustworthy, not less.

35 objectivesCoverage gaps, named — not hidden

They cluster in the operational and AI-native surfaces — activation, namespace, interpretation, environment-interface, work, entity-registration — exactly where landscape research thins out. The named agenda for the lab's next research cycle.

It governs itself first

The framework's last entity-wide control requires the governance system to examine itself by its own rules. GrytLabs runs on the substrate this framework describes; every discipline asked of a governed system, the lab submits to first.

A system that cannot expose its structure cannot be governed — only trusted; and trust without structure is the arrangement this lab was founded to replace.

The convictionPower is legitimate only when it is governed, visible, and reversible. That holds for the AI systems this framework governs — and for us.
WMGF-001 · the load-bearing commitment

07 Boundaries

What it is not

It does not replace COSO, ISO, or NIST.

It tests the structural connective tissue those frameworks assume but never specify. It sits underneath them, not in front of them.

It is not a standard.

No certification, no conformance gate. This is implementation guidance.

It is not substrate-neutral — by choice.

The 19 primitives are DLP's, and DLP-Core is the reference substrate. That specificity is what makes the objectives testable rather than aspirational. Neutrality would be the evasion, not the virtue.

It is not final.

Not certified, not yet peer-reviewed. Published open, on purpose, so it can be tested.

08 Two readers, one page

Who it's for

Governance practitioners

Internal auditors, compliance officers, risk managers, AI-governance leads who already carry COSO / ISO / NIST and want to see exactly what this adds. Spoken to in control-objective, finding, and attestation language — the bridge to your vocabulary is built in.

Founders & operators

Building or deploying world-model-adjacent systems without an audit background. Every concept grounded in plain language, framed by what goes wrong without it. No audit function required to start.

It doesn't ask you to choose between the auditor's rigor and the builder's speed. It's one structure that serves both.

09 To v1.0

Where it's going

1

Amend the one challenged objective — structural invariances architecturally imposed, not learned.

2

Refinement pass over objectives the corpus showed necessary but understated — sharper requirements and language.

3

Research program for the 35 gaps — the operational and AI-native surfaces; the next research cycle.

4

Formal peer review — the full evidence base laid open beside the framework, so reviewers test the grounding.

Buckminster Fuller's test — “if the solution is not beautiful, I know it is wrong” — is the lab's evidentiary standard: when the evidence says the solution is not beautiful, we change the solution, not the evidence.

Each revision is published open on Zenodo under a versioned DOI.

10 The invitation

Read the working draft. Test it. Tell us where it's wrong. The structure is exposed on purpose.

Read the working draft →Download the PDF ↓research.grytlabs.ai research@grytlabs.ai

Basquiat · pinned to the board

The apex purpose

The people with the most to lose from ungoverned systems are rarely the people in the room when those systems are designed. A world where holding a governable model of your own organization is ordinary — not a privilege of those who can afford an audit function — is why equity is written into this lab's apex purpose.

World Model Governance Framework (WMGF-001) · Working Draft v0.5.0 · Cameisha Smith, CIA · DOI 10.5281/zenodo.21220866 · CC-BY 4.0research.grytlabs.ai · research@grytlabs.ai · © 2026 GrytLabs Dynamics Inc.