Research program / Teleodynamic AI

Learning as coupled structure, parameters, and resources

MiRust treats Teleodynamic Learning as a specific, falsifiable research paradigm—not as a synonym for an adaptive agent, a resource monitor, or a language model. This page translates the published definitions into engineering records, observability requirements, and experiment gates.

Coupled teleodynamic learning loops A fast inner loop adapts parameters inside a current structure. A slower outer loop evaluates structural actions. An endogenous resource ledger couples both loops, while observations update history and phase diagnostics. Inner dynamicsparameter adaptationOuter dynamicsstructural actionor noopEndogenous resource ledgercost · reward · viability · history
Conceptual reading aid. It is not an implementation diagram and does not imply that GGUF currently instantiates these dynamics.
Site statusResearch
Primary sourcearXiv preprint v1
Source reviewed2026-06-25 UTC
Implementation evidenceNot verified

Working definition

A narrow term with five definitional commitments

The March 2026 preprint Teleodynamic Learning: A New Paradigm for Interpretable AI defines the paradigm through five commitments. A system missing one of them may still be adaptive, resource-aware, or self-regulating, but should not be labeled teleodynamic under that paper’s definition.

Source status: the primary source is an arXiv v1 preprint. MiRust has reviewed the published text but has not independently reproduced its proofs, software, or benchmark results. Reported DE11 values remain author-reported research evidence.

Open the primary preprint record

Evidence status for the five source-defined Teleodynamic Learning commitments. Defaults record definition evidence only; implementation and reproduction remain unverified.
CommitmentCriterionStatusEvidence or next test
Two-timescale dynamics Parameter adaptation and discrete structural action are separately defined, coupled, and observable. Defined by source Definition reviewed against arXiv:2603.11355v1; no GGUF implementation evidence or independent reproduction has been recorded.
Endogenous resource coupling An internal resource is changed by learning actions and affects later action viability. Defined by source Definition reviewed against arXiv:2603.11355v1; no GGUF implementation evidence or independent reproduction has been recorded.
Local teleodynamic objective Candidate actions expose predictive, structural-complexity, and resource terms before selection. Defined by source Definition reviewed against arXiv:2603.11355v1; no GGUF implementation evidence or independent reproduction has been recorded.
Emergent structural halt No-op can dominate structural alternatives endogenously while inner adaptation may continue. Defined by source Definition reviewed against arXiv:2603.11355v1; no GGUF implementation evidence or independent reproduction has been recorded.
Phase-structured behavior Under-structuring, teleodynamic growth, and the equilibrium/over-structuring boundary can be reconstructed from retained trajectories; structural freeze remains separately observable. Defined by source Definition reviewed against arXiv:2603.11355v1; no GGUF implementation evidence or independent reproduction has been recorded.

System state

Represent the dynamics as explicit state, actions, and transitions

The useful engineering move is to replace metaphor with fields that can be serialized, traced, tested, and challenged.

H

Hypothesis structure

The currently available rules, modules, routes, experts, graph nodes, or other representational units. Structural actions alter this set.

θ

Parameters

Continuous or quasi-continuous values adapted inside the current structure. Parameter updates do not themselves prove structural change.

E

Endogenous resource

An internal state variable changed by learning outcomes and action costs, and consulted when deciding which actions remain viable.

τ

History

The ordered trajectory of observations, predictions, actions, costs, rewards, structural states, and diagnostics needed to explain path dependence.

S = (H, θ, E, τ)γ : S × X × Y → O × S

Engineering reading: the current state, observation, and feedback produce an observable result and a successor state. The result should expose prediction, confidence, selected action, objective terms, and resource change rather than hiding them behind a single model output.

Constraint closure

Name the beneficiary before calling behavior end-directed

The reviewed preprint identifies the hypothesis ensemble as its beneficiary: successful hypotheses receive resource and persist, while weak hypotheses lose support. That is a paper-specific claim, not a reusable answer for every adaptive system.

Identity

What is maintained?

Define the entity whose continuity makes the action history intelligible: an ensemble, rule registry, adaptive overlay, task policy, or another bounded structure.

Viability

What counts as healthy?

Publish measurable viability variables, acceptable ranges, degradation signals, replacement rules, and failure conditions.

Reciprocity

Which constraints sustain each other?

Show how one process maintains conditions needed by another and how the second process returns support. A one-way feedback signal is not sufficient evidence of closure.

Authority

Who can stop or replace it?

Human operators retain deletion, rollback, reset, and safety authority. A beneficiary model does not grant software an unrestricted objective to preserve itself.

Homeostasis is not stasis: a maintained system can remain dynamically active while restoring bounded organization after perturbation. Demonstrating this requires perturbation-and-recovery traces, not anthropomorphic language.

Phase diagnostics

Growth is a time-series claim, not a screenshot

A claimed phase transition must be recoverable from recorded trajectories. A final compact model is insufficient evidence.

  1. 01

    Under-structuring

    Available structure is insufficient: loss remains high, structural actions are frequent, and resource is spent faster than useful predictions replenish it.

  2. 02

    Teleodynamic growth

    Loss declines while structural actions become selective and predictive benefit, complexity, and resource cost remain in active tension.

  3. 03

    Equilibrium / over-structuring boundary

    Noop may dominate and complexity may stabilize; continued growth despite diminishing return is the over-structuring condition the dynamics must expose.

Operational signature

Structural freeze with active inner dynamics

A collapse in structural transition rate while parameter adaptation continues is evidence relevant to emergent halt. It must be separated from a timeout, maximum-move limit, or disabled structural phase.

Concrete instantiation

DE11 is a rule learner, not a language model

The Distinction Engine v11 described by the preprint uses logical Forms, natural-gradient parameter updates, tropical selection, and discrete genesis or wedge actions. It should not be conflated with a transformer, a TinyLM runtime, or a general browser agent.

Structure

Interpretable hypotheses represented as Forms. Structural actions create or replace hypotheses rather than adding transformer layers.

Parameter adaptation

Natural-gradient updates use a diagonal Fisher approximation. The paper explicitly identifies this approximation as a limitation.

Action selection

A local objective balances immediate predictive loss, structural-complexity change, and energy consumption. It is locally rational, not globally optimal.

Reported results

The authors report 93.3% on IRIS, 92.6% on WINE, and 94.7% on Breast Cancer. MiRust has not independently reproduced these results.

Material limitation: the paper reports 55.9% on DIGITS and identifies scalability, diagonal-Fisher approximation, hyperparameter sensitivity, initial-energy dependence, tropical calibration, and limited structural actions as open limitations.

Auditing the source, not just quoting it

The preprint contains testable tensions—not settled implementation facts

A serious project guide should preserve the source’s claims and also record where its formal appendix, empirical narrative, and terminology require independent checking.

01

Emergent halt versus hard caps

The empirical section says external caps were rarely binding. Appendix B.3, however, guarantees eventual freeze by the configured maximum structural moves or structural-phase step limit. A GGUF experiment must therefore show noop dominance before non-binding caps and publish cap-free or expanded-cap counterfactuals.

02

“Bounded” resource versus unbounded accumulation

Appendix B.4 states that when resource decay is one and expected net reward is positive, the energy variable grows without bound. An implementation must define decay, clipping, saturation, normalization, and restart semantics before claiming a bounded resource process.

03

Phase vocabulary varies inside the paper

The definitional section names under-structuring, teleodynamic growth, and over-structuring. The empirical phase section describes an equilibrium/over-structuring regime and visualizes fixed-complexity equilibrium. MiRust records both instead of silently choosing one terminology.

04

Objective direction is inconsistent

Action selection is defined as minimizing the local objective and the formal freeze condition uses a lower-or-equal noop score. Definition 2.4 says noop wins when its score exceeds alternatives. Implementations must publish the comparator and test it explicitly.

Evidence status: these are editorial observations from the reviewed v1 text. They are not a peer review, proof audit, or independent reproduction.

Browser mapping

Map theory to measurable proxies without pretending a proxy is physics

A browser implementation cannot generally read trustworthy thermal state, total device power, or complete system memory. The internal resource variable must therefore be defined as an algorithmic ledger built from available measurements and explicit policy—not casually equated with battery charge or thermodynamic energy.

Candidate resource-ledger fields for a browser experiment
Field Observable proxy Required caution
Compute cost Worker CPU duration, wall-clock duration, dispatch count, or calibrated operation estimate Wall time includes scheduling noise and does not directly measure joules.
Memory pressure Owned WASM bytes, GPU buffer bytes, active hypothesis count, cache size, or allocation failures Browser APIs do not expose complete device memory or reliable cross-browser pressure signals.
Predictive return Task-specific score, calibrated feedback, error reduction, or accepted utility event User approval is not automatically ground truth; reward design can dominate behavior.
Structural cost New nodes, parameters, rules, adapters, routes, or persistent bytes Cost must include maintenance and inference overhead, not only creation.
Resource state Deterministic ledger updated from documented rewards, penalties, decay, and action costs This is an internal model variable unless independently calibrated to physical energy.
Phase evidence Transition rate, noop run length, complexity, score, and resource trajectory over time Hard caps and timeouts must be recorded separately so they are not mislabeled as emergent halt.

Claim boundary

What does not qualify by itself

Static regularization

A fixed penalty in a global loss does not create an endogenous resource state or coupled structural trajectory.

A timeout or memory guard

An externally imposed stop can be good engineering, but it is not an emergent structural halt.

Neural architecture search

An outer search with a fixed external budget is not automatically teleodynamic; the coupling and endogenous resource requirements still apply.

Minimum Description Length alone

An MDL-style fit-versus-complexity objective has no direct analog of the paper’s endogenous resource state or path-dependent action process.

Generic agentic AI

Planning, tools, memory, self-description, or autonomous execution do not demonstrate the five commitments, a beneficiary, or constraint closure.

Local execution

Running in a browser, on a small model, or without a server does not establish teleodynamic learning.

Evidence gate

A minimum falsifiable experiment record

  1. Define the beneficiary

    Name the entity whose persistence is maintained, and identify observable health variables rather than using “the agent” as an unexplained placeholder.

  2. Define both timescales

    List parameter actions and structural actions, their cadence, eligibility, and how inner updates continue after structural freeze.

  3. Specify resource dynamics

    Publish initial state, rewards, penalties, decay, action costs, units, bounds, and restart or death behavior.

  4. Log every candidate action

    Record local objective components for noop and each structural alternative before selection.

  5. Diagnose phases

    Retain transition-rate, complexity, utility, and resource trajectories with seeds and sample ordering.

  6. Run ablations and counterfactuals

    Disable resource coupling, vary initial energy, change ordering, and compare hard-stop controls to test whether the claimed behavior is specific.

GGUF reference envelope

Start with the smallest architecture that can falsify the claim

MiRust recommends separating the predictive model from the structural controller at first. This makes the source commitments observable without pretending that a transformer must mutate online.

Level 1

Fixed predictor + teleodynamic controller

A TinyLM, SLM, embedding model, or rule scorer remains fixed. The controller adapts an interpretable structure such as hypotheses, routes, memory, or task adapters under an endogenous resource ledger.

Preferred first experiment

Level 2

Bounded adaptive overlay

The base model remains fixed while a low-rank adapter, sparse routing layer, retrieval memory, or rule registry receives continuous updates plus discrete structural actions.

Higher attribution risk

Level 3

Full model co-evolution

Both model parameters and architecture change under the declared resource dynamics. This is the strongest claim and requires substantially stronger rollback, reproducibility, and safety evidence.

Research only

Evidence maturity

Move from definition to reproduction without skipping a rung

The site records evidence independently for each commitment. One observed behavior does not promote the entire project, and implementation plans do not count as observations.

  1. 01

    Source-defined

    The criterion is explicitly grounded in the reviewed preprint revision.

  2. 02

    Design-mapped

    A proposed component, state field, action, and observation map to the criterion.

  3. 03

    Experiment-ready

    Implementation revision, data, baselines, metrics, and failure criteria are fixed.

  4. 04

    Observed

    Raw traces show the behavior under the declared method and environment.

  5. 05

    Independently reproduced

    A separate run or implementation reproduces the result with compatible evidence.

Minimum GGUF implementation-evidence handoff
Record Required value Why it matters
Implementation identity Repository, release, commit, artifact checksum, and build UTC Prevents guide prose from being mistaken for executable evidence.
Commitment mapping Concrete inner actions, structural actions, resource transition, local objective, noop, and phase diagnostics Makes each definitional commitment independently reviewable.
Environment Browser, OS, device, CPU/GPU, memory limits, enabled APIs, and hard safety guards Separates algorithm behavior from host constraints and fallback behavior.
Trace bundle Append-only events, candidate scores, selected actions, resource trajectory, structure lineage, seeds, and ordering Supports replay, phase analysis, and falsification.
Result status Observed, failed, inconclusive, or independently reproduced, with limitations Prevents a prototype, plan, or screenshot from silently becoming a support claim.

Review the MiRust–GGUF project boundary

Source and claim register

Keep definitions, interpretations, proposals, and observations separate

Each layer has a different evidentiary weight. The site should never promote a conceptual analogy or implementation plan into an observed capability.

Definition source

Teleodynamic Learning

Teleodynamic Learning: A New Paradigm for Interpretable AI, arXiv:2603.11355v1, submitted 2026-03-11. This is the source for the five commitments, DE11, and the author-reported experiments.

Open the reviewed preprint

Conceptual background

Teleological causality

Deacon and García-Valdecasas, A thermodynamic basis for teleological causality, 2023, DOI 10.1098/rsta.2022.0282. MiRust uses it as conceptual context, not as proof of a software implementation.

Open the DOI record

MiRust layer

Editorial interpretation

The state fields, evidence ladder, browser proxy table, falsification protocol, and GGUF handoff are MiRust engineering interpretations. They are labeled as guidance rather than claims made by the source paper.

Implementation layer

GGUF evidence

No implementation revision, trace bundle, benchmark, or independent reproduction is bundled with this WordPress release. The implementation project remains separately governed and unverified here.

Safety and interpretation

Do not turn systems language into anthropomorphic claims

Teleodynamic Learning is presented as a machine-learning paradigm. It does not establish consciousness, sentience, moral patienthood, biological autopoiesis, or an unrestricted right for software to preserve itself. Human operators retain authority over deployment, deletion, rollback, data retention, and safety limits.