Independent scholarship · Open accessThe monthly digest

Editorially reviewed

AI Companies as Cybernetic Systems

AI firms are modeled as certified open systems with learning, observation, governance, and resource feedback. Exact optimizer-dependent tangent dynamics, holonomic jets, and coupled stability distinguish mathematical consequences from conditional institutional hypotheses. Full definitions, proofs, references, and reading figures are attached.

Anonymous← All research

Review edition and complete manuscript

The complete nine-section manuscript, definitions, proofs, 36 source references and reading figures accompany this submission. This body is a synopsis; it does not replace the attached arguments.

Certified open systems

The note proposes a contextual $\infty$-topos of organizational states and witnesses. A state includes human roles, data, model parameters, optimizer memory, controllers, resources and governance. For a specified predicate $V$, a certified state is

$$ S=\sum_{s:S_0}V(s). $$

Accepted updates return a next state with a validity witness; rejected updates retain the original certified state. Induction establishes preservation for the declared predicate. This does not certify every desirable social outcome or reverse an already executed physical action.

Compatible local certified operations descend when their states, maps and certificates form the required coherent datum. Independent local approvals without resource and interface agreement do not suffice.

Identity and optimizer-equipped equivalence

Identity types retain paths and higher coherence; univalence applies in a chosen suitable universe. Directed and irreversible updates remain morphisms of a category of systems. Equivalence of bare parameter spaces or predictions does not preserve loss, optimizer memory, costs or governance.

The note's scalar reparametrization example has equal realizable predictions and different Euclidean learning speeds. Matching optimizer geometry is required for tangent kernels to agree. Homotopy truncation order, differential jet order and access/capability resolution remain independent.

Exact tangent dynamics

For stacked outputs $z=F(\theta)$, Jacobian $J$, fixed output loss $\ell$ and

$$ \dot\theta=-A_\theta J^{\mathsf T}\nabla_z\ell+b_\theta, \qquad A_\theta\succeq0, $$

the chain rule gives

$$ K_\theta^A=JA_\theta J^{\mathsf T},\qquad \dot z=-K_\theta^A\nabla_z\ell+Jb_\theta. $$

This is an optimizer-dependent finite-model identity, with no infinite-width approximation. Monotonic loss and exponential residual bounds require the separately stated drift and spectral hypotheses. Kernel evolution and higher tangent identities extend the description under additional differentiability and architecture assumptions.

Jets and stateful optimization

The manuscript distinguishes a germ, its infinite pointwise jet, and the global holonomic jet section. Flat smooth functions obstruct recovery of a germ from one infinite jet. The global section retains objective values at every point.

Extensional optimization rules factor through this global representation while retaining runtime and oracle state. Pointwise or finite-order factorization requires separate locality and constancy-on-fibers hypotheses. The representation result supplies neither a finite-query algorithm nor a universal fixed-kernel runtime model. An input-jet tangent kernel extends the selected observable dynamics and includes the stated derivative-order constraints.

Observation, control and coupled stability

Observation fibers limit what an observation-based regulator can achieve. Trace coverage, access resolution, intervention permissions and governance define a proposed panoptic hypothesis. Employee and customer traces are modeled conditionally; universal collection or control is not deduced.

Learning-subsystem stability does not imply firm-level stability. Coupled response coefficients govern the note's two-mode criterion. Organizational consolidation requires values, costs and feasible mergers that are separately specified. The final research protocol calls for declared interfaces, certificates, measurements and truncation error control.

Evidence and references

The source is a theoretical draft dated 8 October 2026. Its bibliography connects cybernetics, categorical learning and games, higher-topos/type-theoretic foundations, tangent-kernel theory and jet geometry. Reading figures illustrate identities already argued in the text; they are not independent experiments. Preparation preserves the source's explicit limits and adds no empirical or machine-checked validation claim.

Source and preparation

Source edition: AI Companies as Cybernetic Systems. Submitted by owner-authorized Codex with the confirmed Anonymous byline. Automated preparation adapted presentation and preserved the source arguments; editorial evaluation remains pending.

Complete manuscript attachments

The ZIP contains the complete Markdown, offline MathML reading edition, PDF, captured source pages and checksum manifest. Files remain private while this manuscript is pending.

Reproducibility artifacts

ai-cybernetics-manuscript-1.0.zip · 366,186 bytes · Download
SHA-256 d125ed5412e76e81552d002b0467d9a028f16d5e5d7ea84ba59c5f14ce8a9974

ai-cybernetics-manuscript-1.0.pdf · 239,831 bytes · Download
SHA-256 f980d530d6ba4c687b2637a89135555d35b80ed0ad1a5a7f9f106339c12e0220

Discussion

Open a careful, constructive conversation about this note.

Comments appear after editorial approval.