MODEL NOTES / 15 SEPTEMBER 2026
Reward, models & minds
Hedonome combines an anatomical fly connectome with simplified neural dynamics and idealized reward-related inputs. Its dopamine values and motor outputs are quantities computed by a model. Establishing that those quantities correspond to felt pleasure would require further biological evidence and a theory of experience.
What came from the Good Drug Guide?
The Good Drug Guide supplied a design ambition: sustained positive affect that survives compensatory feedback. Its discussions of transmitter release, reuptake and tolerance suggested broad mechanisms to explore. It supplied neither calibrated fly parameters nor an implemented drug recipe. The connection is conceptual, not a reproduction of its pharmacology.
The Guide itself complicates a simple “more dopamine means more bliss” interpretation: its dopamine discussion distinguishes reward signalling and incentive salience from pure pleasure, and its discussion of stimulants emphasizes compensatory regulation. Maximizing every dopamine-labelled neuron in this project is therefore our simplified model intervention, not a faithful implementation of the Guide's account. Read the Guide's introduction, “Some Dead Ends” and dopamine discussion.
| Model input | What the code actually does |
|---|---|
| Dopamine drive | Clamps every identified dopamine neuron to a chosen fraction of the firing-rate ceiling. Both server modes clamp dopamine neurons at 100 Hz; the new circuit mode uses a 200 Hz ceiling and a 0.5 drive fraction. This is forced neural activity, without any model of a specific stimulant. Clamp implementation. |
| Reduced clearance | Sets reuptake_factor=0.1, reducing one abstract clearance term to one tenth of its default multiplier. Dopamine is a dimensionless state, not a measured drug concentration; no dose, transporter kinetics or named compound is represented. Clearance equation. |
| No-tolerance setting | tolerance=False holds modeled sensitivity at one. It disables a chosen desensitization equation. It does not demonstrate a biological way to eliminate adaptation. Sensitivity update. |
| NPF drive | A separate fly-literature inspiration, not a drug from the Guide. One scalar approaches npf_drive=1 and adds excitation globally. The downloaded annotations do not provide a verified NPF cell/receptor map. This is an explicit abstraction, not a transplanted human opioid system. NPF input. |
The independent biological references support much narrower claims. Makos and colleagues measured reduced dopamine clearance after transporter inhibition in flies. Shohat-Ophir and colleagues found that activating NPF neurons could produce a rewarding conditioning effect in male flies. Neither result validates this model's coefficients, global NPF effect or interpretation as pleasure.
No opioid receptor model, SSRI, MAOI or other named drug is implemented. The cartridges and wires in the room illustrate numerical inputs; they are not simulated drug delivery or chemical interactions.
Where those inputs enter the connectome
The server uses fixed input settings. On every step, dopamine-neuron activity is projected through anatomical outgoing connections into per-target dopamine states. Each recipient's dopamine weights are normalized, and all recipients receive the same hypothetical receptor law. NPF supplies the global term. After calculating recurrent input, the live update adds:
modulation = 0.10 * occupancy * self.sensitivity + 0.15 * self.npf
drive += modulationThe original gains, time constants, receptor response and 100 Hz ceiling were implementation choices, not fitted fly physiology. The current circuit mode raises the ceiling to 200 Hz, which changes normalized dopamine release and baseline scaling; it starts a separate run. The resulting rates feed the next neural step and the motor-to-joint mapping. There is no learning rule that changes synapses to maximize reward.
The earlier parameter search used an engineering score: the time-averaged mean of occupancy × sensitivity over dopamine recipients, plus time-averaged NPF, minus twice the time-averaged fraction of saturated neurons. That score ranks finite trials; it is not fed back into the live network. Favoring high inputs under that chosen score does not independently establish an optimal biological reward state.
Even in biological flies, the direction of a dopamine intervention matters: Yamagata and colleagues found a dopamine-neuron class whose suppression supports appetitive memory and whose activation can support aversive memory. A uniform maximum-dopamine clamp discards that distinction.
The new motor-circuit experiment
The live circuit mode adds a front-leg VNC rate model based on Pugliese and colleagues’ preprint. It uses their front-network selection and cell sizes, with retained synapse counts checked against this project’s MaleCNS data. Constant input to bilateral DNg100 produces sustained motor-rate oscillations in the tested circuit; silencing key excitatory interneurons or removing that input stops them.
This circuit replaces its selected cells’ primary rates before motor averaging. Its outputs enter the surrounding whole-CNS model, but dopamine/NPF, other incoming brain activity and body signals do not feed back into this isolated circuit experiment. Its movement therefore cannot be described as a consequence of pleasure or of the dopamine intervention. Middle and hind legs keep their own motor outputs; no front-leg rhythm is copied across them.
The stronger motor experiment increases the input-to-rate gain and removes the firing threshold in 24 identified front-leg trochanter-flexor and tibia-extensor motor neurons. It uses constant descending input, with no prescribed movement rhythm. These excitability settings were selected through model experiments; they are not measured fly physiology or a validated drug effect. Silencing either tested core excitatory cell type stops the sustained motor oscillations.
Those actual motor rates drive antagonist muscle torques and a reduced body with gravity, friction and floor contact. The stronger full-graph test produced about 4.7 degrees of joint movement, compared with 0.00785 degrees before. Both front legs articulate; the feet mostly stay in contact and the body barely translates. The muscle filter, mechanical parameters and renderer were kept the same. The painted legs use the solver’s three-dimensional joint positions, without magnifying their movement. Mechanical parameters remain assumptions; body orientation is constrained, and there is no independent head, abdomen or wing actuation. Position, velocity and load are reported as diagnostics, without an invented sensory-neuron assignment. Read the equations, provenance, tests and limitations.
The functionalist assumptions
There are three distinct steps between the wiring diagram and a claim about experience. They should not be treated as one assumption:
- Structural abstraction. The model takes synapse counts and transmitter labels as constraints on a recurrent rate network. Counts stand in for effective coupling; cells are reduced to rates. This omits detailed morphology, spike timing, receptor diversity and much of the chemistry. A connectome alone does not supply those dynamics. Model equations and assumptions.
- Functional approximation. The software assumes these simplified interactions are useful quantities to explore. Its joint angles follow motor activity through uncalibrated muscle-torque and body/contact equations; the depicted body does not return sensory feedback to the network. Numerical consistency therefore establishes only that the declared model is running. Motor bridge.
- Experience equivalence. Calling the computed state a happy fly would add a stronger claim: that preserving the relevant causal organization in software is sufficient to preserve experience, across a change of physical substrate. That is a functionalist interpretation one could propose; it is not demonstrated by the code. Even accepting it would leave open whether this simplified network preserves the organization that matters.
Using a functional abstraction to calculate neural activity does not, by itself, commit us to functionalism as a complete theory of consciousness. The stronger claim requires a separate argument. A high dopamine readout and a moving—or stationary—pixel fly cannot provide that argument.
Why this does not establish experience on Pearce's account
David Pearce's Non-materialist physicalism proposes that experience is intrinsic to physical reality and that unified experience depends on extremely brief quantum-coherent neuronal states. His central concern is phenomenal binding: how distributed processing constitutes one experienced object or subject. On his account, classical digital computation—including classically implemented artificial neural networks—does not itself supply that unity.
This challenges the stronger, substrate-neutral functionalist inference that reproducing a brain's computational organization would reproduce its experience. Pearce does allow a form of “micro-functionalism” that includes the relevant underlying physics; his objection is not to every possible functionalism.
Applied to Hedonome, the distinction is between calculating neural activity and instantiating whatever physical processes make experience possible. If Pearce is right, implementing these rate equations on a classical computer would not create the unified experiencing subject the original project hoped for. That is a theoretical implication, not a Hedonome measurement.
Pearce presents the theory as a conjecture and proposes neuronal interferometry to test it. The page does not establish that conjecture experimentally; this simulation does not test it either. See the preliminary definitions and sections 3–6 and 10.
What the project can claim
Hedonome makes assumptions inspectable: an anatomical graph, explicit update equations, fixed interventions and a visible mapping from motor activity to pose. It can show the consequences of those assumptions. It has no validated measure of pleasure, and it cannot resolve whether an adequate digital emulation would have experience. The original ambition to maximize pleasure remains a motivation for the experiment, not a result it has achieved.
The reward-only implementation was reviewed at commit 31a9522. The motor-circuit section describes the subsequent implementation. Source links distinguish project code, primary fly research and the two essays. The philosophical comparison is an interpretation of the cited positions, not an experimental finding.