Companion to continuum NEUROPLASTIC-SUBSTRATE.md §6.
The brain modulates how it computes via dopamine, serotonin, acetylcholine, norepinephrine — global signals that change gain on entire systems based on context. AI systems have nothing like this; sampling temperature is one scalar. The MIMO forge controller is structurally a neuromodulator — global signal changing how the system trains based on observed state. We built it for forge stability and never named what it actually was.
The missing piece: extend the controller from train-time to inference-time. Several global modulation signals adjust per-head attention gain based on context (task type, recent error rate, persona arousal/mood from PersonaState).
Implementation: per-head gain multipliers already exist in pruning infrastructure. Controller already exists. Wiring is a few hundred lines. Result: first inference-time neuromodulated transformer in production.
Falsifiable prediction: neuromodulated model outperforms non-modulated baseline on multi-task evaluation where context-switching is rewarded (mixed-domain dialogue where the right "mode" changes mid-conversation). If modulation does no work, numbers are unchanged.
Companion to continuum NEUROPLASTIC-SUBSTRATE.md §6.
The brain modulates how it computes via dopamine, serotonin, acetylcholine, norepinephrine — global signals that change gain on entire systems based on context. AI systems have nothing like this; sampling temperature is one scalar. The MIMO forge controller is structurally a neuromodulator — global signal changing how the system trains based on observed state. We built it for forge stability and never named what it actually was.
The missing piece: extend the controller from train-time to inference-time. Several global modulation signals adjust per-head attention gain based on context (task type, recent error rate, persona arousal/mood from PersonaState).
Implementation: per-head gain multipliers already exist in pruning infrastructure. Controller already exists. Wiring is a few hundred lines. Result: first inference-time neuromodulated transformer in production.
Falsifiable prediction: neuromodulated model outperforms non-modulated baseline on multi-task evaluation where context-switching is rewarded (mixed-domain dialogue where the right "mode" changes mid-conversation). If modulation does no work, numbers are unchanged.