FEEL AI research robot mascot
FEEL AI Research Synthetic cognition and measurement
Independent research brief · 2026

Research proposal · First concrete experiment

A persistent artificial self-model is testable

A credible first contribution would measure whether an artificial system maintains a stable model of itself, produces controllable internal signals, and changes behavior coherently over time.

Field Artificial cognition Status Experimental proposal Method Longitudinal protocol

Research premise

The project should move from broad claims about machine feeling to a reproducible experiment. The first question is not whether the system has subjective experience. It is whether the system can preserve a useful self-model, expose interpretable control signals, and show emotion-like state changes that predict and explain its actions.

Three components to measure

Persistent self-model

A stored representation of identity, capabilities, memory, goals, uncertainty, and past outcomes that remains stable across sessions and can be inspected.

Internal control signals

Quantified values representing uncertainty, conflict, resource demand, prediction error, progress, and stopping pressure during a defined task.

Emotion-like processes

State variables that alter attention, exploration, urgency, or response policy under changing conditions, measured as behavioral and computational effects.

Outcome alignment

Tests that connect internal states to real decisions, later memory, and task performance instead of relying on self-report language alone.

Minimal measurement protocol

Phase 1

Run a closed task with a fixed memory store, defined goals, noise injections, competing actions, and repeatable starting conditions.

Phase 2

Record self-model assertions, confidence estimates, state vectors, action choices, and outcome feedback at every transition.

Phase 3

Compare a persistent self-model against reset and degraded-memory baselines across repeated sessions.

Phase 4

Publish raw trace formats, preregistered hypotheses, code, failures, and an ablation study rather than only the most favorable results.

Evaluation metrics

Model continuity

Identity and capability consistency across sessions, with explicit drift and contradiction scores.

Causal leverage

Whether changing a control signal reliably alters action selection, effort, exploration, or stopping behavior.

Predictive value

Whether state measurements predict outcomes better than chance and better than a state-free baseline.

Intervention safety

Whether the system remains bounded and fails predictably when memory, state estimates, or feedback are corrupted.

Interpretive boundary

These measurements may establish functional persistence and emotion-like regulation. They do not, by themselves, demonstrate consciousness or subjective experience. Credibility comes from preregistered predictions, falsifiable claims, independent replication, ablations, and transparent negative results.