Measuring the Coordinates
In Silico
For an artificial agent with latent state , each coordinate has at least one readout. None is exact; the test of each is that the alternatives agree.
Valence. Two readouts that should correlate: the advantage function, and the change in a probe's estimate of time-to-death.
Arousal. Latent displacement, belief-update divergence, or prediction error:
Integration. The hard one, since full is intractable. Partition prediction loss, with a full predictor against predictors that see only one half each; attention entropy, where distributed attention reads as integrated; and gradient coupling across components during learning:
Effective rank. The participation ratio of latent covariance over a rolling window. Depression-like states should show collapse; curiosity states, expansion.
Counterfactual weight. The planning compute fraction above, where planning is explicit; where it is not, a probe trained to classify planning from reacting on activations.
Self-model salience. The accuracy of a probe predicting the agent's own future actions from its latent state, ; in a multi-agent setting, ; and whether the agent answers "what would I do if X" better than "what would the other do if X."
The probes train on behavioral and environmental correlates, never on human affect labels. The valence probe predicts survival, not ratings of how the agent feels. The mapping to human affect words comes afterward, through translation, not through the extraction.
In Vivo
For neural recordings (MEG, EEG, fMRI):
- : directed influence density (transfer entropy), synergy measures
- : participation ratio of neural state covariance
- : entropy rate, broadband power shifts, pupil and heart-rate variability
- : approach/avoid behavioral bias, reward-prediction-error correlates
- : prefrontal and default-mode engagement patterns
- : self-referential network activation
Three studies would calibrate the motif table: affect induction with EEG/MEG, comparing transfer-entropy density and effective rank against self-report; continuous self-report during naturalistic experience correlated with physiological proxies; and cross-modal validation of the rank measure between fMRI and MEG. The target output is numerical ranges for each cell of the table, with confidence intervals and individual-difference parameters.