Stonkfly
Experimental MaleCNS controller with market-chart inputs, paper trading, and optional Coinbase integration. Includes modeled reinforcement and memory; profitable learning has not been demonstrated.
Synapses change under a modeled rule grounded in fly biology — for example dopamine-gated change on Kenyon-cell → MBON synapses in the mushroom body. Learning begins here.
Dispute this rating →The evidence
No evidence, no level.Each simulation bin writes the updated efficacies back into the connectome's own weight array whenever weights are not frozen.
https://github.com/nftechie/stonkfly/blob/78ef3e05ab0fa086032098558d893667068944a0/stonkfly/neural/brain.py#L331-L352The rule is the Huang/Luo 2024 baseline-centered anti-Hebbian eligibility model, driven by measured KC and DAN firing rates with declared bounds and trace constants.
https://github.com/nftechie/stonkfly/blob/78ef3e05ab0fa086032098558d893667068944a0/stonkfly/neural/rule.py#L1-L65Reinforcement enters as a current pulse into the identified reward or aversive dopamine population, and the same observation loop runs the plasticity rule; the decision readout only counts spikes.
https://github.com/nftechie/stonkfly/blob/78ef3e05ab0fa086032098558d893667068944a0/stonkfly/neural/controller.py#L49-L97The authors state that no profitable learning, strategy improvement or biological replication has been demonstrated, and that applying one rule to both memory compartments is their own unvalidated extension.
https://github.com/nftechie/stonkfly/blob/78ef3e05ab0fa086032098558d893667068944a0/docs/model.md#L53-L59
The upstream of the fly-wirehead neural core, and the same verdict for the same reasons: an approximate LIF kernel over the full retained MaleCNS v1.0 graph, with a baseline-centered anti-Hebbian rule adapted from Huang/Luo 2024 writing new efficacies into the live weight array on 7,835 existing KC->MBON07/MBON11 edges during the run. What is specific to this repo is the closed loop: incremental portfolio P&L above a deadband schedules a 200 ms current pulse into the 15 PAM11 cells, and P&L below it pulses the 2 PPL101 cells, so the reinforcement reaches the plasticity rule as actual dopamine-cell spikes rather than as a scalar written onto a weight. The trade decoder is a fixed spike-count threshold on DNp20/DNpe017, not a trained readout. Rate the simulation, not the returns: the project claims no edge. No validation mark: docs/model.md states flatly that no profitable learning, strategy improvement or biological replication has been demonstrated by the repo's tests, and that applying one rule to both the alpha1 and gamma1pedc compartments in this male reconstruction is the authors' own unvalidated extension.
Record
Nearby
All projects →Ratings on thelearningfly.com are proposals, not verdicts. Every one of them is arguable in public.