Project summary

This project funds a 16-week proof of concept for AATM: a distributed, continuous, label-free sensorimotor architecture designed to test whether a finite artificial substrate can build reusable sensorimotor coalitions from an open world, and later reactivate one of those coalitions without external stimulus.

The core question is not whether this system can outperform current AI models. It will not. The goal is narrower and falsifiable:

Can a continuously active architecture, with no object labels, no cognitive timestamp, no central object registry, and no reset between experiences, learn distributed sensorimotor pathways through movement, prediction, feedback, homeostatic state, and local plasticity?

The technical appendix is already specified in v0.3: architecture, invariants, data structures, experiments, ablations, metrics, implementation plan, logging, and falsification conditions.

Public project context:
https://www.grdprocess.ch/distributed-intelligence/

Technical requirement V0 (still in French): https://www.grdprocess.ch/wp-content/uploads/2026/06/AATM_Specification_reproductible_noyau_sensorimoteur_reconstruction_v0_3.pdf

What are this project's goals? How will you achieve them?

The goal is to build a reproducible test bench for the first AATM proof of concept: the Distributed Sensorimotor Core.

The system will test whether several mechanisms can be made to work together in a minimal software environment:

• continuous operation without cognitive resets;

• no label, object ID, timestamp, absolute pose, or episode number entering the cognitive core;

• local sensorimotor sampling in a 2D world;

• distributed micro-units and columns where no unit contains the complete object;

• prediction through movement and local transitions;

• recognition as stabilization of a coalition under external constraint;

• reconstruction as endogenous reactivation when external input is absent;

• vector homeostasis as internal pressure and selection bias;

• causal ablations to distinguish real architectural effects from shortcuts.

The implementation will start with a controlled 2D environment: entities made of shared primitives, local sensors, simple movement actions, and recurrent distributed units. The system will then be tested under multiple conditions: correct movement, permuted movement, novel signals, eyes-open recognition, eyes-closed reconstruction, homeostatic recall, topological accessibility, virtual myelination, and internal/external attention regimes.

The success criterion is not “AGI”. Success means producing interpretable evidence that specific mechanisms are causally active. For example, if removing efference copy does not degrade prediction, then the sensorimotor hypothesis fails. If removing feedback does not degrade reconstruction, then reconstruction likely comes from a shortcut. If an isolated column performs as well as the full network, the distributed representation hypothesis is not validated.

Negative results are useful. The project is designed so that failure should be informative rather than ambiguous.

How will this funding be used?

The requested funding is CHF 50,000 for a 16-week implementation and validation program.

The funding will be used for:

• researcher runway during the 16-week build and validation period;

• implementation of the standalone Python/NumPy test bench;

• experiment automation, logging, reproducibility, and reporting;

• multi-seed validation runs and ablation matrix;

• storage, backup, compute maintenance, and infrastructure;

• documentation, public write-up, and communication of results;

• optional targeted technical review if needed.

Planned work packages:

• Weeks 1–2: freeze specification, API contracts, invariants, YAML configuration, seeds, leakage tests, 2D world.

• Weeks 3–4: continuous engine, body, sensors, event queue, residual activity, logging.

• Weeks 5–6: local prediction, sparse routing, local error, movement ablations.

• Weeks 7–8: vector homeostasis, per-dimension effects, associations.

• Weeks 9–10: distributed coalitions, recurrence, inhibition, revisitation, open-world test.

• Weeks 11–12: reconstruction and attention, feedback, eyes-closed condition, internal/external gating.

• Week 13: virtual myelination and delay reversal experiments.

• Week 14: baselines and multiple seeds.

• Week 15: validation freeze, holdout seed, statistical analysis, falsification tests.

• Week 16: repository, report, video demonstration, and go/no-go recommendation for the next phase.

Who is on your team? What's your track record on similar projects?

The project is currently led and implemented by Gaetan Duchateau through GRDprocess Sàrl, Switzerland.

I am an independent researcher working at the intersection of cognitive architecture, informal logic, structural reasoning, and AI systems. My work focuses on whether some cognitive properties are better treated as emergent consequences of architecture rather than as features to be directly programmed.

Relevant prior work includes:

• AATM conceptual framework and technical specification;

• a simple homeostatic agent prototype testing bounded internal variables and non-apathetic behavior;

• multi-agent homeostatic loop experiments;

• dialectical/reinforcement experiments on model behavior under different training regimes;

• THE FRAME, an operational system for structural decomposition and consistency analysis of normative prescriptions;

• public documentation through GRDprocess.

The current proposal is narrower than the full AATM research program. It asks for funding only to implement and test the first reproducible sensorimotor core.

What are the most likely causes and outcomes if this project fails?

The most likely causes of failure are technical and scientific.

Main failure modes:

• movement may not causally improve prediction;

• distributed coalitions may fail to stabilize;

• one global attractor may dominate the network;

• activity may fragment instead of forming reusable pathways;

• reconstruction may depend on leakage or shortcuts rather than feedback;

• homeostatic selection may be too weak or too dominant;

• novel signals may be forced into existing categories;

• the architecture may require a level of biological detail not captured by the current abstraction.

The project includes direct falsification criteria. Examples:

• If removing movement changes nothing, the mechanism is not genuinely sensorimotor.

• If an isolated column identifies as well as the full network, the representation is not distributed in the intended sense.

• If removing feedback does not impair eyes-closed reconstruction, reconstruction is probably caused by a shortcut.

• If every novel signal is forced into an existing entity, the system behaves as a closed-set classifier.

• If the same attractor always wins regardless of internal state, the system collapses into unregulated topological dominance.

The main successful outcome would be a reproducible proof of concept showing separable causal effects: dependence on movement, persistence without reset, distributed representation, feedback-driven reconstruction, homeostatic selection, and novelty accommodation.

The main negative outcome would still be valuable: a documented failure showing which architectural assumption does not hold under this implementation.

How much money have you raised in the last 12 months, and from where?

CHF 0 raised for AATM in the last 12 months.

I have applied to several grant or funding channels for related AI research, including LTFF, Emergent Ventures, SFF and Coefficient Giving, without getting funding.

This Manifund request is intended to fund the first reproducible implementation and falsification test of the AATM Distributed Sensorimotor Core.