Curadoria viva do ecossistema MaleCNS, com a mesma régua para projetos próprios e independentes. O eixo científico avalia a força da evidência pública sobre o que cada projeto demonstra. Um null reproduzível pode ficar acima de uma demo chamativa sem controles. Interesse mede originalidade, ambição e fertilidade.
whole-connectome neurotransmitter-source/sign audit and converter
Estágio
released reproducible audit with converter, verification script and DOI
Origem
independent
Project ID
neuropunks-malecns-nt-audit
Repositório
neuropunks/malecns-nt-audit
Full-graph MaleCNS v1.0 audit of neurotransmitter source columns and polarity mapping, with a corrected three-class converter and a separate recomputation script for every published count. The evidence is unusually direct and reproducible for an infrastructure project: it operates on the released 25.6-million-edge graph and preserves exact accounting. The tier does not imply that its three-class polarity abstraction validates any downstream spiking model or behavioral claim.
Ver evidência e limites
Evidência mais forte
The audit covers the full 166,700-neuron, 25,582,938-edge MaleCNS v1.0 graph and reports concrete source-column/sign discrepancies rather than a sampled subset.
It quantifies four systematic error classes, including 884,631 Kenyon-cell output edges that change class under predicted versus consensus NT labels and 89,723 histaminergic edges affected by the legacy sign mapping.
The corrected conversion preserves an exact 25,582,938-edge total, while verify_audit.py independently recomputes the quoted counts from the raw feather files and emits explicit PASS/FAIL checks.
The analysis, converter and reproduction contract are released publicly with a citable Zenodo DOI.
Limitações
The audit validates annotation/source-column accounting, not biological dynamics, cognition or controller performance.
Treating dopamine, serotonin and octopamine as a separate modulatory class with zero instantaneous signed drive is a modeling choice; it does not reproduce receptor-specific neuromodulatory kinetics.
The verification script is independent code within the same project, not an external replication by another group.
Controles
Direct predicted_nt versus consensus_nt comparisons expose source-column sensitivity.
The converter can reproduce the legacy predicted/two-class path, enabling exact old-versus-corrected comparisons.
The verification stage recomputes published totals from raw official tables and the converted parquet rather than trusting cached summaries.
Missing for downstream claims: matched task-level experiments measuring how corrected versus legacy NT mappings change reservoir/control performance.
Histórico do tier
2026-09-26: initial placement -> scientific A / interest A, high confidence; full-graph reproducible audit credited, downstream dynamical/behavioral claims explicitly excluded.
This card scores the audit as a scientific infrastructure result, not as evidence that a particular MaleCNS simulator is behaviorally faithful.
whole-MaleCNS real-time Atari controller with causal circuit interventions
Estágio
working whole-brain real-time demo with paired interventions, circuit rewires and preserved negative controls
Origem
independent
Project ID
dohun1214-malecns-asteroids
Repositório
dohun1214/malecns-asteroids
Whole-MaleCNS v1.0 is run as a 166,700-neuron LIF system inside Atari Asteroids, with escape and pursuit readouts tied to named circuits. The strongest evidence is causal rather than score-based: paired replay, targeted circuit interventions, degree/sign/weight-preserving rewires, off-circuit and size-matched controls, dose responses and a cross-mode dissociation localize behavior to specific modeled wiring. The project also preserves counterevidence: a supervised MLP beats the fly on a held-out direction task, rule-based gameplay is statistically competitive, and a previously plausible movement metric was withdrawn after a random-motion falsification.
Ver evidência e limites
Evidência mais forte
Under identical taped stimuli, targeted DNp11 intervention changes action agreement to 56.9% and eliminates thrust, while two random cells leave actions exactly unchanged and 200 random cells produce a much smaller effect.
Degree-, sign- and weight-multiset-preserving rewires of LC4 projections push escape-bearing error from 60.7 degrees to 99.2 degrees, while same-size off-circuit and synapse-matched controls remain near the intact result.
LC10a-to-AOTU rewiring shows a monotonic dose response; at 100% rewiring the modeled directional information loss is 60.1% across 10 seeds while roughly half the signal amplitude remains, and matched other-output, escape-circuit and random-excitatory controls stay near zero loss.
A cross-mode dissociation preserves circuit specificity: LC10a intervention leaves flee-mode behavior unchanged at the reported resolution but collapses pursuit aiming, whereas DNp11 intervention strongly damages the flee readout.
The repository includes byte-identical branch replay, committed result artifacts, deterministic sanity checks and a 30-minute real-time soak rather than relying only on a live-demo narrative.
Limitações
The connectome is embedded in an engineered LIF model with engineered visual transduction and motor readouts; causal claims apply to this modeled controller, not directly to biological flies.
LC10a pursuit assumes arousal and repurposes a courtship circuit for Asteroids; the project explicitly marks that analogy as a strong modeling assumption.
A supervised MLP reaches lower held-out direction error than the fly readout, and whole-brain gameplay score is within one standard deviation of the rule-based controller, so the evidence does not establish general controller superiority.
Some closed-loop effect sizes are sensitive to Atari wraparound and observation details; the project documents multiple corrected measurement failures and large seed spread in the 100% pursuit rewire.
There is no independent external replication of the full intervention suite.
Controles
Paired stimulus replay and byte-identical state restoration separate intervention effects from ordinary trajectory divergence.
Random-cell interventions, off-circuit rewires, size/synapse-matched rewires and random-excitatory rewires test whether generic damage explains the observed circuit effects.
Random and supervised MLPs, a rule-based controller, random behavior and stillness provide conventional and behavioral comparators; importantly, the supervised MLP and rule controller prevent a claim of general MaleCNS superiority.
The repository withdraws a movement-direction metric after a random-motion control reproduced it, and retains uncertain or negative survival and absolute-rate findings.
Histórico do tier
2026-09-26: initial placement -> scientific A / interest S, high confidence; unusually strong within-model causal localization and topology-preserving controls credited, while biological validity and controller-superiority claims remain bounded.
The A tier is for intervention and falsification evidence inside the declared MaleCNS LIF model, not a claim that the simulation is a faithful biological fly or the best Asteroids controller.
reproducible whole-CNS closed loop with preserved negative mechanistic results
Origem
independent
Project ID
aravpanwar-amfly
Repositório
aravpanwar/amfly
Six synchronized copies of the full 166,700-neuron MaleCNS are coupled in a feedback installation in which one operator network selects stimulation targets and chamber state feeds current back to the operator. The strongest evidence is causal and within-model: pinned-data reproduction, isolation, perturbation propagation and several negative mechanism tests are explicit. The project does not establish that MaleCNS topology is uniquely responsible for the effects, and the visible convulsion is an authored direct motor drive.
Ver evidência e limites
Evidência mais forte
The published reproduction contract loads the full 166,700-neuron / 25,582,938-connection MaleCNS and reports 31 real-data tests passing, including exact counts, determinism, isolation, divergence, stimulus and analysis gates.
With phasic sensory drive, the stimulated instance diverges around step 18 and accumulates large spike-identity differences, while the four unstimulated chambers and operator remain bit-identical controls.
The project preserves failures that changed the interpretation: tonic drive synchronized away perturbations, raw DN argmax only reached two chambers, a moving reference produced a false operator divergence, and delayed stimulation of a settled network often changed only one neuron.
Driving the modeled escape pathway does not yield the desired motor response; the shipped visible convulsion instead injects current directly into 708 motor neurons and is explicitly labeled as authored rather than emergent.
Limitações
"Dopamine", punishment and electrode effects are modeled current injection, not demonstrated biological reinforcement, pain or tissue damage.
The operator mapping, dwell/hysteresis, chamber thresholds and body-motion amplification are engineered interfaces outside the frozen connectome.
The visible convulsion bypasses the escape circuit by directly driving motor neurons.
No degree-preserving rewire, random recurrent network or compact conventional controller is evaluated under the same closed-loop task.
Controles
Unstimulated whole-CNS instances and the operator provide exact within-run isolation controls; bit-identity is a stated invariant.
Phasic versus tonic stimulation, stable-reference repair and direct-motor-drive comparisons preserve negative and failure cases rather than tuning them away.
Missing: matched topology-destroying and non-connectome baselines capable of testing whether the specific MaleCNS wiring matters.
Histórico do tier
2026-09-26: initial placement -> scientific B / interest S, high confidence; whole-CNS causal/reproducible evidence credited, topology-specific and biological reinforcement claims withheld.
The tier credits controlled whole-network mechanism evidence, not the installation's reward/punishment metaphor or direct motor convulsion.
A frozen MaleCNS-derived recurrent substrate, hand-built visual front end and trained action head navigate the unmodified E1M1 level with strong held-out reactive behavior. The repository also preserves evidence that matched rewires carry more target information in an earlier task, so this supports useful frozen-connectome computation rather than superiority of biological wiring.
Ver evidência e limites
Evidência mais forte
On 100 untouched E1M1 test seeds, the connectome agent reaches the exit in 57% of runs and averages 92% route progress; widened starts plus sticky actions yield 64% exits.
Randomly reordered frames, black frames and flat-grey frames all reduce exits to 0%, while an open-loop replay reaches only 6%, bounding memorized-route and vision-free explanations.
The recurrent wiring remains frozen; training is confined to the downstream head.
An earlier topology gate reports degree- and weight-preserving rewires carrying more target information than the biological wiring (R2 0.62 vs 0.46), and the adverse result is retained.
Limitações
The eye is a hand-built motion/contrast model rather than a validated fly visual system.
The final head has about 5.4 million trained parameters and reads visual projection neurons as well as descending neurons.
No matched rewired/random recurrent control has yet been run on E1M1 itself.
Successful navigation depends on stochastic action sampling; the greedy policy reaches 0% exits.
The result covers one level on one difficulty setting.
Controles
Held-out random, always-forward, open-loop replay, shuffled-frame, black-frame, flat-grey and privileged-reference controls are reported on the same protocol.
Wide-start and sticky-action perturbations test sensitivity to starting state and repeated controls.
The earlier topology gate is adverse to a connectome-superiority claim, though it used a float16 feature cache and is not an E1M1 matched control.
Missing: E1M1 biological topology versus degree/weight-matched rewire or random recurrent substrate under the same eye, head capacity and training budget.
Histórico do tier
2026-09-26: initial placement -> scientific B / interest S, high confidence; strong held-out reactive controls credited, connectome-specific superiority explicitly withheld.
The tier rewards controlled closed-loop performance and preservation of adverse topology evidence, not a biological cognition claim.
MaleCNS subgraph on microcontroller with untrained escape readout
Estágio
hardware-verified subgraph execution and escape demo
Origem
independent
Project ID
slvdev-esp32-fly
Repositório
slvDev/esp32-ai
A 48,311-neuron MaleCNS subgraph runs directly on an ESP32-S3 with exact retained contact counts, reproducible host/device gold checks and an untrained looming-to-escape readout. The artifact strongly demonstrates compact execution of a biologically sourced pathway, but not a whole-CNS or topology-specific advantage.
Ver evidência e limites
Evidência mais forte
The released device graph contains 48,311 neurons and 9,462,135 directed edges with retained synaptic contact counts, and every inference step traverses the retained graph.
Device verification reports 9,275,712 graph-gold values with max error 3.9e-7, 960 escape-gold values with max error 1.8e-7, and 60 autonomous decisions replayed against the host with max error 3.0e-8.
Looming-detector stimulation makes known escape descending neurons dominate 20 same-size random visual-group controls; for DNp01, DNp02, DNp04 and DNp11, 0 of 20 controls reached the reported loom response.
Limitações
The board cannot fit the whole MaleCNS graph; the retained 48,311-neuron subset excludes the VNC and most optic-lobe intrinsic neurons.
The runtime uses engineered rate dynamics and positive contact-count weights; it does not reproduce transmitter signs, receptor effects, delays, biological plasticity or measured firing dynamics.
No degree-preserving rewire, random recurrent network or matched conventional circuit is reported for the escape computation.
Controles
Twenty random visual-projection groups of matched size provide a stimulus-to-readout specificity control.
Independent host/device graph gold, escape gold and autonomous-decision replay test numerical fidelity of the port.
A topology-destroying matched graph control remains missing for a MaleCNS-specific computation claim.
Histórico do tier
2026-09-26: initial placement -> scientific B / interest S, high confidence; strong hardware and pathway evidence credited, topology-specific and whole-CNS claims withheld.
The tier credits the verified connectome computation and deployment artifact, not biological realism.
A MaleCNS-derived haltere-to-wing experimental programme with a full-network forward simulator, explicit circuit sweeps, an inverse model validated by round-tripping through the simulator, and a Flappy Bird harness. The strongest evidence is mechanistic rather than behavioral: independent sweeps reproduce surprising indirect haltere effects, while the current game agent still holds altitude rather than avoiding pipes.
Ver evidência e limites
Evidência mais forte
Full-network cluster sweeps show SNpp12 and SNpp23 driving the wing-motor pool despite no direct measured synapses, while directly connected SNpp14 remains ineffective; an independent native-backend paired-cluster sweep reproduces the same ordering.
The inverse 67-motor-to-205-haltere MLP is checked against the real forward simulation, not only its labels; after targeted dataset repair it reaches IoU 0.75-1.0 across the named-cluster sweep and fixes a documented silent-target failure mode.
The Flappy harness records long multi-seed runs and corrected a prior scoring bug; the best reported T=560 runs clear 49 and 42 pipes over 6,000 ticks, while the project explicitly documents that this is fixed-altitude chance alignment rather than obstacle avoidance.
Limitações
Haltere stimulation is engineered host-side current injection, not measured haltere transduction physics.
The inverse model solves a simulator-defined inverse problem and does not establish biological uniqueness or real-fly causality.
The Flappy agent receives no pipe-gap information, so cleared pipes are not evidence of useful visual avoidance.
No matched rewired-connectome, random recurrent or compact conventional circuit is reported for the central haltere-to-motor findings.
Controles
Static one-hop wiring analysis is explicitly contrasted with full-network dynamics, exposing a direct-wiring explanation that fails.
Independent native and GPU sweeps cross-check the SNpp12/SNpp23/SNpp14 result.
The inverse-model workflow preserves marginal-frequency baselines, held-out validation, round-trip oracle tests and negative/failure cases.
A topology-destroying same-interface control remains missing.
Histórico do tier
2026-09-26: initial placement -> scientific B / interest S, high confidence; mechanistic replication and round-trip validation credited, behavioral avoidance and topology-specific advantage withheld.
The tier credits the controlled circuit evidence; the game's pipe count is not interpreted as learned avoidance.
MaleCNS visual-to-descending subcircuit controller for Chromium Dino
Estágio
results-bearing reproducible game controller with held-out benchmark and replicas
Origem
independent
Project ID
cobanov-flyjump
Repositório
cobanov/flyjump
An 80-cell MaleCNS visual-to-descending circuit drives the original Chromium Dino through a trained 243-parameter readout. The project has a clean train/validation/test split, three independent training seeds, reproducible artifacts and multiple behavioral controls. The published model completes 99/100 held-out 180-second courses and the replicas span 85-100/100. The tier stops at B because the compact circuit, engineered state encoder and trained readout are strong inductive biases, and no equally budgeted rewired or artificial recurrent control tests whether the measured MaleCNS topology is specifically responsible for the advantage.
Ver evidência e limites
Evidência mais forte
The published checkpoint completes 99/100 held-out 180-second Chromium Dino courses with mean survival 179.372 s; the same readout with the circuit silenced and the initial untrained readout both complete 0/100.
Three independently trained seeds complete 99/100, 85/100 and 100/100 on the same frozen held-out set, while validation and test are kept separate and the first declared training seed remains the published checkpoint regardless of later test scores.
The circuit is selected from measured MaleCNS anatomy before training: 80 cells, 1,296 directed measured edges and 26,029 synaptic contacts, with only the small action readout trained.
The repository publishes per-course benchmarks, checkpoints, training histories, deterministic environment pins and tests that reproduce benchmark, input sensitivity, silencing, graph reachability and rendered/headless parity.
Limitações
The eight game-state observations are engineered structured features rather than pixels or biological retinal transduction, and their assignment to visual cell types is arbitrary.
The modeled recurrence uses simplified signed tanh dynamics rather than measured membrane physiology.
Only an 80-cell selected subgraph participates in control; the remaining atlas is anatomical context rather than whole-CNS computation.
The handwritten rule comparator is intentionally simple and untuned, so the result does not establish superiority over strong conventional controllers.
No matched degree-preserving rewire, random recurrent network or artificial graph is trained with the same budget.
Controles
Same trained readout with the connectome circuit silenced, initial untrained readout, handwritten rule baseline, uniform-random actions and idle are evaluated on the frozen held-out courses.
Three independent training seeds reproduce high completion rates without test-based checkpoint selection.
Pinned original Chromium physics and sprites, deterministic seeds and rendered/headless parity reduce environment drift.
A topology-specific null remains missing, so circuit dependence is not interpreted as biological-topology superiority.
Histórico do tier
2026-09-26: initial placement -> scientific B / interest S, high confidence; reproducible held-out control and independent training replicas credited, topology-specific advantage withheld.
The B tier credits reproducible causal dependence on the modeled circuit and disciplined evaluation, not evidence that this MaleCNS topology is better than a matched artificial network.
controlled simulation results; raw-frame vision negative; no physical flight
Origem
independent
Project ID
hldrnwnv-fly-drone
Repositório
hldrnwnv/fly-drone
In a frozen virtual FPV benchmark, the original MaleCNS graph passes 20/20 tested courses while three degree-shuffled controls pass 1/20, 0/20 and 6/20. A conventional geometric baseline also passes 20/20 and is faster, and the project's raw-frame neural-vision variant fails all 20 trials. The B tier therefore credits a narrow topology result while preserving the stronger conventional baseline and the end-to-end vision failure.
Ver evidência e limites
Evidência mais forte
MaleCNS passes 20/20 frozen virtual courses versus 1/20, 0/20 and 6/20 for three independently degree-shuffled graph controls.
Exact same-frame replay records 150/164 correct direction choices for MaleCNS versus 80/164, 84/164 and 90/164 for the three shuffled controls.
The conventional geometric baseline also passes 20/20 and is faster, preventing a general superiority claim.
The separate raw-frame condition is negative: the conventional visual baseline passes 20/20 while MaleCNS and all shuffled conditions pass 0/20.
Limitações
The successful condition receives an externally estimated target direction rather than solving raw visual recognition.
The broader simulated vehicle stabilization remains outside the MaleCNS substrate.
The experiment is virtual; no physical flight is demonstrated.
The shuffled controls preserve degree but not every biological or spatial attribute.
Controles
Three degree-shuffled graph controls use matched training and evaluation budgets.
A conventional geometric baseline and exact same-frame replay constrain closed-loop confounds.
The raw-frame negative condition tests whether the result survives removal of the externally estimated target direction; it does not.
2026-09-26: initial placement -> scientific B / interest S, high confidence; topology-control result credited, conventional-baseline tie and raw-frame failure preserved.
This is a virtual simulation result; the card makes no claim about physical autonomous flight.
MaleCNS/MANC-derived locomotor circuits embodied in MuJoCo legged robots
Estágio
results-bearing open-loop embodiment with preregistered closed-loop negative pilot and engineered interleg coordination study
Origem
independent
Project ID
sakshaymahna-fly-cord-robots
Repositório
SakshayMahna/fly-cord-robots
A connectome-grounded fly walking project that reproduces published ventral nerve cord circuitry, maps connectome-derived motor-neuron activity into a MuJoCo body and then tests what is required to close the proprioceptive and interleg loops. Its strongest evidence is unusually informative negative evidence: four preregistered proprioceptive encoders find no regime where feedback materially changes the neural trajectory while preserving rhythm, and subsequent coordination work explicitly shows that usable interleg timing must be supplied downstream. The tier is B because the experiments isolate real failure mechanisms and preserve adverse results, but coordinated locomotion is not produced by the connectome alone and no matched topology null tests whether the measured wiring is specifically necessary.
Ver evidência e limites
Evidência mais forte
The preregistered closed-loop pilot tested four proprioceptive encoder formulations; across the stable regime trajectory effects remain about 0.0000-0.0010, while gains that produce effects around 1.01 destroy the rhythm, leaving 0 of 7 sweep levels meeting both preregistered criteria.
The project reproduces six published leg circuits and drives actual MuJoCo joint movement from connectome-derived motor-neuron activity in open loop before adding feedback.
Measured MaleCNS/MANC geometry rules out conduction delay as the needed 12 Hz antiphase mechanism: the longest inter-CPG distance is 432 micrometers, while the required 41.7 ms delay would imply an implausibly slow 0.0104 m/s conduction velocity.
Cruse-style downstream coordination with a pausable clock yields a hexapod tripod index of +0.282 and, without re-specifying the rules, left-right quadruped alternation of -0.743 after middle-leg removal; the repository explicitly attributes timing coordination to the engineered rules rather than the connectome.
Limitações
The closed-loop result is explicitly a pilot on parameter seed 641 with replicates 0-3; the preregistered main-experiment seed was not run.
Usable interleg coordination is supplied by an engineered pausable clock and Cruse-style rules downstream of the connectome because the modeled CPGs cannot be entrained without destroying their rhythm.
The primary walking circuitry is ventral-nerve-cord/MANC-derived rather than a demonstration that the entire MaleCNS jointly controls the robot.
No degree-preserving rewire, random recurrent network or conventional controller is matched to the same neural/interface budget for a topology-specific comparison.
Transfer to a quadruped demonstrates reuse of the engineered coordination rule, not autonomous discovery of a new gait by the connectome.
Controles
Four successively refined proprioceptive encoders are evaluated under the same preregistered decision rule, including explicit baseline-stability filtering.
The project separates rhythmicity from feedback effect size and preserves the null/destructive regimes instead of selecting a favorable gain after the fact.
Coordination experiments compare bounded output delay against a pausable-oscillator implementation and test the same rules after middle-leg removal.
A matched topology-destroying neural null remains missing, so B does not imply biological-topology superiority.
Histórico do tier
2026-09-27: initial placement -> scientific B / interest S, high confidence; strong preregistered negative and mechanism-discriminating evidence credited, while downstream engineered coordination and missing topology null bound the claim.
The strongest contribution is the clean boundary it establishes between what the measured fly circuitry supplies and what the robot still needs from engineered coordination.
whole-connectome spiking platform and frozen reservoir
Estágio
public reusable package with controlled pathway and communication experiments
Origem
independent
Project ID
alextitonis-fly-ai
Repositório
alextitonis/fly.ai
A reusable full-MaleCNS spiking platform with explicit pathway experiments, multiple applications and unusually useful negative controls. Side-specific looming and steering effects survive repeated noise seeds, while the project's own degree-preserving scramble and time-shuffle experiments show that stronger claims about special topology or temporal code do not always survive.
Ver evidência e limites
Evidência mais forte
The public FlyBrain package retains the 166,700-neuron, roughly 25.6-million-connection MaleCNS graph as a frozen recurrent substrate with task-specific encoders and fixed or trained readouts.
Direct feature-detector injection produces side-correct LC4/LPLC2-to-DNp01 and LC10a-to-DNa02 effects across six noise seeds.
The fly-to-fly communication experiment reports three runs with 50-shuffle permutation nulls and a degree-preserving scrambled-wiring control; at 2 ms the scrambled graph carries at least as much singer-state information as the measured wiring, explicitly falsifying the stronger topology claim in that regime.
Limitações
The core uses simplified point-neuron dynamics, rough transmitter-sign rules and no validated graded-neuron model; the project documents failures at the photoreceptor/lamina boundary.
Several applications bypass the eye by injecting known feature-detector populations or add task-specific readouts, so downstream success is not automatically evidence for autonomous whole-brain competence.
The simulations are not validated against recordings from the imaged fly.
Controles
Noise-seed replication, silence controls, permutation/time shuffles and a degree-preserving scrambled connectome are used in the reported communication experiments.
Negative results are preserved, including settings where vision does not modulate motor output and the 2 ms scrambled topology matches or exceeds the measured graph.
Histórico do tier
2026-09-26: initial placement -> scientific B / interest S, high confidence; controlled pathway evidence and adversarial nulls credited, biological-fidelity and topology-superiority claims withheld.
This card scores the reusable flybrain core; materially distinct applications may receive separate cards when they have their own evidence.
whole-MaleCNS combat game / escape-pathway controller
Estágio
combat prototype with full-graph escape pathway, real game smoke tests and pathway ablations
Origem
independent
Project ID
codeman1729-flybrain-arena
Repositório
CodeMan1729/flybrain-arena
A real 166,700-neuron MaleCNS computation is embedded in a playable Godot combat loop, with threat injected into annotated LC4/LPLC2 populations and DNp01/DNp03 activity used as an engineered dodge readout. Its strongest evidence is causal within the implemented model: edge ablation sharply reduces the escape signal and left/right input reversal flips the lateral response. The tier stops at B because the input encoding, dynamics and body controller remain engineered and no whole-topology matched null has been run.
Ver evidência e limites
Evidência mais forte
The runtime propagates through the full 166,700-neuron, 25,582,938-edge MaleCNS graph and reads annotated LC4/LPLC2 to DNp01/DNp03 pathways.
Direct-data checks report substantial real LC4-to-DNp01/DNp03 and LPLC2-to-DNp01 connectivity rather than assuming an escape pathway from labels alone.
A controlled local edge ablation drops reported escape strength from 1.0 to 0.088, providing causal evidence that the selected anatomical pathway materially drives the implemented readout.
Left/right input reversal, no-input cases, reset/lifecycle checks, real service/client tests and full game smoke tests are described as passed.
Limitações
Threat is injected directly at LC4/LPLC2 because the project does not model the full retinal-to-looming pathway; this is an explicit engineering intervention.
The signed activity-deviation dynamics are an engineering model, not a calibrated biological reproduction of fly neural dynamics.
Chase kinematics and the mapping from descending-neuron activity to dodge vector are engineered controllers.
The project does not establish that whole MaleCNS topology outperforms a matched rewire, random recurrent controller or a simple scripted dodge policy.
Controles
Pathway-edge ablation and left/right-input reversal directly test whether the selected model pathway contributes to the implemented dodge signal.
Occlusion, stale-response rejection and no-input tests constrain software/protocol alternatives but do not test topology specificity.
Missing: whole-network topology nulls and a matched simple controller under the same threat sensor and body dynamics.
Histórico do tier
2026-09-26: initial placement -> scientific B / interest S, high confidence; causal pathway ablation credited, whole-connectome superiority claim withheld.
results-bearing controlled pilot with negative calibration
Origem
independent
Project ID
8s8642-flyface
Repositório
8S8642/FlyFace
A frozen MaleCNS simulation is driven by controlled face-image stimuli and compared with a published human-rating dataset. The strongest result is a bounded negative: the configured network does not show a matching global ranking, and the proposed DNa02 orientation readout fails synthetic controls.
Ver evidência e limites
Evidência mais forte
At 900 ms, DNa02 chose the higher-rated image in 36/65 informative duels (55.4%, two-sided binomial p=0.457), with 35 silent duels.
Per-image DNa02 preference was essentially uncorrelated with the reference ratings (Spearman rho=-0.038; within-group permutation p=0.820).
Static/moving controls failed to validate DNa02 as an orientation-toward-image readout; no alternative among 473 bilateral descending types passed all four controls.
Limitações
The simulator inherits simplified MaleCNS dynamics from its pinned neural dependency and is not a recording of living neural responses.
The 900 ms exposure was selected after a 300/900 ms pilot, so it is not an untouched confirmatory endpoint.
No matched degree-preserving rewire, random recurrent network or no-connectome control isolates topology-specific effects.
Controles
Counterbalanced left/right presentation, mirrored/grayscale/shifted variants, within-group permutation tests and multiple-testing correction are reported.
Synthetic static and moving orientation controls directly challenge the assumed DNa02 readout and preserve the failed calibration.
A matched topology-destroying null remains missing for any MaleCNS-specific representation claim.
Histórico do tier
2026-09-26: initial placement -> scientific B / interest S, high confidence; controlled nulls and failed readout calibration credited, topology-specific claims withheld.
The scientific tier rewards the quality of the negative result and calibration discipline, not a ranking-prediction claim.
single-seed steering skill with connected CARLA demonstration
Origem
independent
Project ID
markunthank-flyhard
Repositório
MarkUnthank/flyhard
Full-scale MaleCNS-derived controller coupled to a simulated fly body and a passive steering wheel, with a held-out stationary steering skill and a causal grip-disconnection intervention in CARLA. The behavioral evidence is concrete, but one training seed and the absence of a matched shuffled or conventional-policy comparator prevent a topology-specific conclusion.
Ver evidência e limites
Evidência mais forte
The retained model uses 165,122 traced MaleCNS neurons and 25,563,197 measured neuron-pair edges.
A single trained steering run passed 100/100 held-out requested-angle trials after 0/100 before training.
In a 24-second CARLA demonstration, disabling the fly-to-wheel grip reduced peak steering by more than 99.98% under the same checkpoint and scripted route.
Limitações
The learned steering result is from one training seed and interpolation within the trained angle range.
Visual autonomous driving, coordinated learned pedal control and three-seed replication remain untested.
The current evidence does not compare the measured graph against a matched rewired graph or ordinary policy under the same training budget.
Controles
Matched before/after model with frozen interfaces and unchanged topology.
Grip-disabled intervention isolates the causal fly-body-to-wheel mechanical path.
A topology-destroying matched graph control and conventional policy comparator remain missing for the learned steering claim.
Histórico do tier
2026-09-26: initial placement -> scientific B / interest S, high confidence; strong embodied causal evidence, topology-specific claim withheld.
2026-09-27: evidence-source refresh -> connected-CARLA report and current repository head rechecked; scientific B / interest S unchanged because newer commits do not add a matched topology or conventional-policy comparator.
The tier credits the demonstrated embodied control chain, not a claim that MaleCNS topology is superior.
MaleCNS-derived browser game with controlled visual-circuit experiments
Estágio
playable game; neural visual propagation supported, useful navigation not established
Origem
independent
Project ID
dzhng-fly-escape
Repositório
dzhng/fly-escape
A browser game backed by a selected MaleCNS subgraph and careful visual-circuit experiments. Preregistered, multiplicity-corrected neural contrasts and silencing and occlusion checks support downstream propagation of modeled visual input, while motor pilots and paired gameplay results do not support useful attraction, avoidance or better navigation.
Ver evidência e limites
Evidência mais forte
All seven prespecified neural-vision contrasts pass simultaneous Bonferroni intervals across 847 comparisons, with 26 optical, graded-current, trajectory-equality and sham checks also passing.
Input silencing reproduces the dark-input neural trajectory; wall and furniture blocking reproduce dark, and reopening the optical path restores the response.
The retinal programme preserves failed primary endpoints: v2 spatial/brightness/occlusion fails while color passes, and a frozen v3 reserved-seed run still fails 2 of 9 spatial contrasts without post-result retuning.
Paired seeded gameplay runs show retinal input changes outcomes but reduces escapes in both reviewed rooms, which the project correctly reports as changed behavior rather than improved navigation.
Limitações
The runtime uses a selected seed-touching MaleCNS subgraph rather than whole MaleCNS, and omits retinal R1-R8 cells.
Retinal registration, receptive-field geometry, current scaling, neuron dynamics and body decoding are engineered model assumptions.
Controlled neural propagation has not been converted into a reliable player-controllable attraction or avoidance effect.
No matched degree-preserving rewire, random recurrent network or compact conventional circuit tests whether the biological topology is necessary.
Controles
Bonferroni-corrected preregistered contrasts, fixed held-out seed sets, input silencing, optical blocking, sham checks and trajectory-equality controls are preserved with raw evidence.
Failed motor and readout pilots, failed spatial endpoints and worse escape counts are retained rather than relabeled as success.
Missing: a topology-destroying same-interface control and a matched conventional controller for the navigation claim.
Histórico do tier
2026-09-26: initial placement -> scientific B / interest S, high confidence; controlled visual-circuit evidence credited, useful navigation and topology-specific claims withheld.
The scientific tier is driven by the controlled neural experiments, not the game's star score.
portable MaleCNS feeding-subcircuit simulator and mainframe reproducibility experiment
Estágio
results-bearing 501-neuron subcircuit with cross-platform deterministic execution; real IBM Z remains untested
Origem
independent
Project ID
mertefesensoy-onfly
Repositório
mertefesensoy/ONFLY
ONFLY extracts a 501-neuron MaleCNS sugar-to-feeding subcircuit, compares it with a much larger annotated MaleCNS reference and runs the same fixed-point experiment through portable C89/SoftFloat implementations including s390x under QEMU and MVS 3.8j under Hercules. Its strongest contribution is the unusually explicit reproducibility ledger: 19 golden requests agree across the recorded platforms and the scientific runs preserve both successful dose-response results and failures or post-result changes to acceptance criteria. The tier stops at B because the science is measured on x86, the decisive network files are not yet distributed for clone-only reproduction, the subcircuit carries a fitted compensating input, and no matched rewired/random topology control establishes MaleCNS-specific computation.
Ver evidência e limites
Evidência mais forte
On x86-64, sugar input at 40, 60, 120 and 200 Hz makes the MN9 feeding motor neurons fire in all 30 seeds at every rate, while zero sugar produces zero spikes in all 501 modeled neurons.
The compensated 501-neuron subcircuit stays within 10% of the 184,099-neuron annotated MaleCNS reference at 40, 60, 120 and 200 Hz under the project's reported protocol.
Nineteen golden requests have matching fingerprints across the recorded x86-64, Linux s390x/QEMU and MVS 3.8j/Hercules implementations; the project also records a GCCMVS code-generation fault found by byte-level stream comparison rather than hiding it behind the fingerprints.
The project preserves the failed magnitude comparison to the Shiu-style reference and explicitly records that two acceptance criteria were changed after seeing results.
Limitações
The scientific acceptance measurements are x86-64 results; MVS and s390x establish implementation agreement, not an independent biological or scientific replication.
The 501-neuron subcircuit includes a fitted compensating input for omitted neurons, and its 10 Hz comparison is excluded because the larger reference is too variable there.
The two network files needed for the full x86 suite are not distributed in the repository, so a fresh clone cannot yet reproduce the complete result path.
The project has not run on real IBM Z hardware or z/OS; the MVS result is under Hercules and the s390x result under QEMU.
There is no matched degree-preserving rewire, random recurrent network or conventional computational null testing whether the measured MaleCNS wiring is specifically necessary.
Controles
Zero-sugar trials, 30-seed response measurements, a much larger MaleCNS reference network and a re-run of the Shiu-style reference constrain the feeding claim.
Native, SoftFloat 3e and SoftFloat 2c backends plus big-endian and MVS recordings test numerical portability; response streams are compared beyond summary fingerprints.
Negative and post-hoc evidence is retained: the magnitude criterion originally failed, 10 Hz is excluded from the subcircuit acceptance comparison, and later criterion amendments are explicit.
Missing: a topology-specific null with matched dynamics and operating point.
Histórico do tier
2026-09-26: initial placement -> scientific B / interest S, high confidence; cross-platform reproducibility and preserved falsifiers credited, topology specificity and clone-only reproducibility withheld.
The B tier credits disciplined implementation and subcircuit evidence; bit-identical execution across emulators is not itself evidence that MaleCNS topology is biologically privileged.
whole-connectome embodied agent with local plasticity
Estágio
canonical v1 reproduced; plasticity demonstrated, behavioral gain not demonstrated
Origem
independent
Project ID
shute2004-virtual-fly
Repositório
shute2004/virtual-fly
Whole-MaleCNS embodied reconstruction coupled to FlyBody/MuJoCo with local plasticity. Canonical v1 provides a public reproduction path and demonstrates stored-weight changes, while frozen before/after evaluation does not show categorical behavioral improvement.
Ver evidência e limites
Evidência mais forte
Canonical v1 reports a whole-MaleCNS closed loop using a 166,700-neuron / 25,582,938-edge snapshot.
The canonical training run records 2,163,179 changed stored edges and publishes manifests plus a reproduction path.
Frozen initial and final evaluations produce the same categorical outcome, so weight change is not promoted into a behavioral-success claim.
Limitações
The short canonical run does not demonstrate improved behavior, generalization or long-term learning stability.
The simulator does not claim complete biophysical fidelity.
Asynchronous simulator trajectories are not claimed to be bit-identical across hardware.
Historical visual results are explicitly separated from canonical evidence.
Histórico do tier
2026-09-26: initial placement -> scientific B / interest S, high confidence; A withheld because the central behavioral learning payoff is not demonstrated.
working Snap Spectacles + Mac whole-connectome embodiment with public device demo and verification path
Origem
independent
Project ID
kvnloo-fly-brain-spectacles
Repositório
kvnloo/fly-brain-spectacles
An unusually complete physical embodiment of one full MaleCNS simulation per holographic fly, with room geometry, hands and objects translated into sensory drives and measured descending/motor activity returned to an AR body. The system is real and well disclosed, but substantial sensing, state, landing, gait and low-level control remain engineered and there is no matched topology control showing that MaleCNS wiring is necessary for the observed behavior.
Ver evidência e limites
Evidência mais forte
The repository states that all 166,700 neurons are simulated every step; the 16,000-neuron cloud is only a visualization sample.
A public Spectacles demo and an end-to-end verify_brain path exercise food turning, looming escape and bitter-stop behavior against the whole-brain server.
The Metal GPU kernel is documented as an ordered-delivery exact port, with the CPU path producing the same spikes more slowly.
The decision log explicitly separates measured neural quantities from engineered body state, controllers, landing rules, hunger gain, gait and perception helpers.
Limitações
The AR world is translated into hand-designed sensory channels; Gemini labels objects but does not itself choose actions.
Flight, wing rhythm, landing, walking gait, hunger and several body behaviors use engineered low-level rules around decoded neural commands.
The project intentionally injects at deeper feature detectors when pixel-level paths do not produce usable behavior in this model.
No matched rewired/random recurrent/simple controller is reported for the same AR tasks.
Controles
The project has explicit end-to-end scenario and brain-verification checks and documents cases where channels such as hearing produce no useful response.
CPU-versus-Metal equality is an implementation control, not a scientific topology control.
Missing: same sensors/body/controller with MaleCNS replaced by degree-matched rewire, random recurrent network or compact conventional controller.
Histórico do tier
2026-09-26: initial placement -> scientific C / interest S, high confidence; whole-brain physical embodiment credited, connectome-specific behavioral attribution withheld.
whole-MaleCNS plus NeuroMechFly/MuJoCo embodied fly simulator
Estágio
runnable whole-graph hybrid simulator with engineered ethology and separate circuit experiments
Origem
independent
Project ID
artem-x-meta-fly-arena
Repositório
artem-x-meta/fly-arena
A substantial embodied simulator that can prepare and run all 166,700 annotated MaleCNS entries and 25,582,938 connection rows alongside a NeuroMechFly/MuJoCo body, approximate retinal cameras and descending-neuron readout. The project is unusually explicit about its hybrid boundary: FlyGym supplies walking, while food search, arbitration, escape and several ethology demonstrations use engineered controllers; the public showcase GIFs are explicitly recorded with the connectome disabled. Scientific tier C credits the real whole-graph integration and inspectable experimental packages, but withholds stronger causal attribution until matched no-graph, rewire or conventional-controller comparisons show what the MaleCNS topology contributes to behavior.
Ver evidência e limites
Evidência mais forte
The preparation pipeline downloads and verifies MaleCNS v1.0, retaining 166,700 annotated non-glial entries and all 25,582,938 connection rows in the runnable sparse graph.
A connectome run exposes LIF activity through descending-neuron gait commands inside the same MuJoCo/NeuroMechFly environment, with checkpoints fingerprinting physics, neural and controller state.
The repository includes separate visual-pathway, graded-response, motion-selectivity and semantic-readout research packages and preserves negative boundaries such as the failure to replace the external visual controller with a validated neural alternative.
The README explicitly states that the showcase feeding, grooming, search and sleep clips use engineered controllers with the connectome disabled, preventing those demonstrations from being misattributed to MaleCNS.
Limitações
The walking layer uses FlyGym's engineered walking controller; motor neurons are not individually mapped to muscles.
Food search, behavioral arbitration and escape detection contain engineered logic, and the two-fly social demonstrations run without a neural graph.
The retinal interface uses approximate 96 x 96 cameras and project-defined LIF/background inputs rather than validated MaleCNS physiology.
The repository itself states that running the full graph does not demonstrate that natural behavior emerges from connectivity alone.
No matched degree-preserving rewire, random recurrent network or equally capable conventional controller establishes a topology-specific behavioral contribution.
Controles
Body-only and controller-driven modes are kept distinct from connectome modes, making the engineering/neural boundary inspectable rather than silently blended.
Checkpoint compatibility records code, dependency, graph and model fingerprints for reproducibility.
Separate diagnostic/ablation controls exist in the experimental packages, but the public main-behavior evidence is not a matched topology-causal benchmark.
Histórico do tier
2026-09-27: initial placement -> scientific C / interest S, high confidence; whole-graph embodied integration credited while controller-heavy public behaviors and missing matched topology nulls cap the scientific tier.
This is a strong platform and unusually broad systems integration; the C tier is about causal evidence for MaleCNS-specific behavior, not about software ambition.
native pickup-and-hold baseline with broader experimental controller
Origem
independent
Project ID
monomyth-fly-brain-codex
Repositório
monomyth/fly-brain-codex
MaleCNS-derived visual/proprioceptive controller for a simulated six-joint arm and gripper. The retained native baseline has repeated pickup-and-hold evidence, while the broader touch-gated controller is explicitly experimental. The public record does not yet isolate value attributable to MaleCNS topology from the learned head and engineered interfaces.
Ver evidência e limites
Evidência mais forte
The retained native baseline reports 19/20 pickup-and-hold trials in a defined 20 mm cube region.
Separate normal-UI runs verified pickup, five-second hold and scripted release/drop checks.
The deployed controller receives two rendered RGB views plus robot feedback rather than cube XYZ or task-phase oracle state.
Limitações
The stable result is restricted to a small placement region and one cube size; broader placements include missed or unstable grasps.
Release is an explicit scripted completion routine rather than a learned neural action.
Training selected connections, response offsets and a learned motor head prevents attributing success to fixed MaleCNS topology alone.
Controles
The project preserves separate stable, experimental and fit-only profiles instead of conflating their evidence.
A matched random/rewired recurrent or small-network comparator is not reported for the native success rate.
Histórico do tier
2026-09-26: initial placement -> scientific C / interest A, high confidence; repeated native task evidence credited, MaleCNS-specific mechanism unresolved.
MaleCNS visual-circuit simulation and branded-house avoidance art demo
Estágio
working replay-backed simulation with targeted circuit lesion
Origem
independent
Project ID
quillotaku-fly-away-from-security
Repositório
Quillotaku/fly-away-from-security
A deliberately playful MaleCNS simulation in which a 49,002-neuron visual-to-descending subgraph receives engineered staged visual drive and a decoder determines whether a fly enters or avoids a house. The published replay pipeline is reproducible and a targeted LC4/LPLC2 lesion flips the branded-house decision while leaving the stimulus injection in place, demonstrating a real causal path inside the declared model. Scientific credit is limited because the brand association and action decoder are explicitly invented, the effect is not a natural task, and no matched topology null or conventional controller establishes MaleCNS-specific utility.
Ver evidência e limites
Evidência mais forte
The pipeline rebuilds a 49,002-neuron, 1,026,595-edge MaleCNS subgraph lying between visual projection neurons and descending neurons, with source files verified by published hashes.
Under the published scenarios, the intact branded-house condition produces a +0.0487 decoder signal and avoidance, while lesioning the 146 driven LC4/LPLC2 neurons changes the signal to -0.0179 and entry even though the stimulus is still injected.
Alarmix and Casa Tranquila receive identical inputs and produce the same -0.0380 result, preserving the explicit boundary that the model distinguishes only the association deliberately encoded by the authors.
Replay generation validates timing, ID ranges, normalization and browser/pipeline agreement before artifacts are served.
Limitações
The mapping from a fictional brand sign to LC4/LPLC2 drive is intentionally invented; it is not biological evidence that flies recognize brands or alarm signage.
The LIF parameters and global synaptic gain are modeling choices rather than measured MaleCNS physiology.
The project uses a selected visual-to-descending subgraph rather than the full CNS, and the decoder from descending activity to approach/avoid is engineered.
There is no naturalistic behavioral task, independent replication, matched rewire/random-network control or strong conventional baseline.
Controles
A targeted lesion blocks propagation from the driven LC4/LPLC2 population while retaining stimulus injection, distinguishing pathway dependence from simply turning off the sensor.
Two non-target brand scenarios use identical inputs and yield identical outputs, checking that the implementation does not secretly distinguish them.
The browser recomputes decisions from replay data and checks them against the pipeline expectation.
Topology-preserving and topology-destroying matched nulls are absent.
Histórico do tier
2026-09-26: initial placement -> scientific C / interest A, high confidence; reproducible causal lesion credited while the intentionally fabricated semantic association and missing topology controls bound the scientific claim.
The scientific claim is intentionally narrow: the modeled MaleCNS-derived pathway causally carries an engineered signal in this simulation. The joke premise is not treated as biological evidence.
reproducible full-network browser model with engineered hybrid controller
Origem
independent
Project ID
mingdianliu-flybrain-playground
Repositório
mingdianliu/flybrain-playground
Browser-accessible full-network MaleCNS simulation with pinned source data, live sensory perturbations, anatomy telemetry and an embodied 3D room. Its reproducibility and model boundaries are strong, while the flight behavior is explicitly an engineered hybrid rather than evidence for an emergent MaleCNS motor controller.
Ver evidência e limites
Evidência mais forte
The model allocates state for essentially the full retained MaleCNS graph and exposes computed spikes, population rates and selected-neuron traces in the browser.
Pinned source hashes, graph-load fail-closed behavior and full-network tests make the numerical artifact independently rebuildable from official public data.
The room experiment reports deterministic synthetic task outcomes and full-network sensory-response checks under movable light, obstacles, odor and sound inputs.
Limitações
Visual retinotopy, sensory transfer functions, proximity encoding, stabilization and several steering assists are engineered rather than experimentally calibrated.
The documentation explicitly states that the downstream network is not a validated motor controller and that synthetic room outcomes are not predictions of real-fly behavior.
No matched rewired-connectome, random recurrent or compact conventional-controller result is reported for the embodied room task.
MaleCNS-derived policy/value model for Rubik's Cube
Estágio
public trained demo with legacy controlled results and corrected propagation code
Origem
independent
Project ID
lntegrals-flycube
Repositório
lntegrals/flycube-public
FlyCube places an 8,192-neuron measured MaleCNS subgraph inside a trained Rubik's-Cube policy/value loop and reports matched rewired, silenced and uninformed controls. Its public controlled advantage is still legacy evidence: those checkpoints propagated connectivity in the wrong direction; the code is corrected, but corrected-direction results are not yet the published benchmark.
Ver evidência e limites
Evidência mais forte
The preparation pipeline imports all 166,700 neurons and 25,582,938 directed edges, then extracts a reproducible 8,192-neuron input-to-readout subgraph with measured IDs, geometry, contact counts and transmitter-predicted signs.
Legacy matched 15-minute controls on the short suite report connectome 44%/57% solved in policy/search mode versus rewired 10%/14%, silenced 0%/14%, and random or uninformed-BFS 6%/14%.
The project preserves zero training overlap for the uniform test set, deterministic same-state policy checks, a negative DAgger result, and the original erroneous checkpoints rather than silently replacing them.
Limitações
The published positive benchmark uses legacy checkpoints trained with reversed connectivity propagation, so it cannot establish a correct-direction MaleCNS topology advantage.
Uniform random cube states are 0/100 solved, and the successful regime is shallow; the project explicitly makes no deep-solving or optimality claim.
Sticker encoding is engineered, neural units are rates rather than biological spikes, and the trained model uses only a reduced subgraph.
Controles
Matched rewired, silenced, untrained and uninformed/random baselines are implemented in the evaluation harness.
Teacher solvers are offline-only and blocked from live inference by tests.
A corrected-direction rerun of the matched controls is the decisive missing control for the strongest current result.
Histórico do tier
2026-09-26: initial placement -> scientific C / interest S, high confidence; strong control design credited but legacy direction-bug results prevent promotion.
The tier is intentionally conservative until the corrected propagation path reproduces the controlled advantage.
Public browser demo that runs a compact MaleCNS-derived recurrent operator in a Doom-like multimodal sensorimotor loop. A shuffled comparison is public, but the strongest topology-specific interpretation remains ahead of the evidence because representation matching and repeated held-out provenance are incomplete.
Ver evidência e limites
Evidência mais forte
A public interactive sensorimotor demo executes the compact MaleCNS-derived recurrent operator.
The public report includes concrete compression/throughput measurements and a degree-preserving shuffled comparison.
The sensor-to-descending-neuron mapping is described publicly.
Limitações
The compact MaleCNS arm and shuffled control are not fully representation-matched.
Repeated held-out runs, seed-level uncertainty and directly linked run artifacts are insufficient for a robust topology claim.
Sensors, motor mapping and game physics are engineered interfaces.
Controles
Degree-preserving shuffled wiring is reported, but not under a fully matched representation.
A repeated same-budget random recurrent or no-connectome control is not reported in the public article.
Histórico do tier
2026-09-26: initial placement -> scientific C / interest A, medium confidence; live artifact and shuffled control credited, topology claim narrowed.
This is a scientific-evidence tier, not the editorial tier of the associated blog post.
full-MaleCNS simulation path implemented; public browser demo uses MiniFly; hardware flight not yet demonstrated
Origem
independent
Project ID
spikecalls-flydrones
Repositório
SpikeCalls/FlyDrones
A real MaleCNS loading, simulation and calibration path is wired to a drone-control stack, but the public browser behavior is produced by the 850-neuron MiniFly stand-in and the hardware adapters have not yet been flight-tested. This is substantial system evidence, not yet evidence that full MaleCNS is a useful physical drone controller.
Ver evidência e limites
Evidência mais forte
The project implements a MaleCNS v1.0 loader, signed spiking simulation, named visual/descending populations and a calibrated motor readout.
The full-connectome path can be built, inspected, calibrated and run through the same simulator/runtime used by the control stack.
The project explicitly separates MiniFly browser demos, full MaleCNS execution, engineered interfaces and external safety control.
Simulator, retina, decoder, safety, protocol and MaleCNS-loader behavior are covered by tests.
Limitações
The browser demo and GIFs use MiniFly, an 850-neuron hand-wired stand-in, not the full MaleCNS connectome.
The real-drone adapters are documented as not yet flight-tested by the project.
The drone's own flight controller and an external safety governor remain outside the neural substrate.
Camera-to-neuron encoding and the descending-neuron readout are engineered.
The MaleCNS builder prefers consensus_nt and treats histamine as inhibitory, but maps dopamine, serotonin and octopamine to positive instantaneous synaptic drive; the public MaleCNS NT audit identifies those amines as a distinct modulatory class, so full-brain results should be checked under that alternative mapping.
Controles
A same-size random-graph benchmark is reported for throughput, but it is not a behavioral topology control.
Dry-run and simulator-first hardware paths provide engineering safety controls.
The independent MaleCNS NT audit (https://doi.org/10.5281/zenodo.22975837) provides a concrete transmitter-mapping sensitivity test for the full-connectome path.
Missing: full-MaleCNS versus rewired/random recurrent/simple-controller comparison under the same simulated or physical flight task.
Histórico do tier
2026-09-26: initial placement -> scientific C / interest S, high confidence; full-MaleCNS infrastructure credited, physical/full-brain behavioral claim withheld.
2026-09-26: transmitter-mapping clash review -> tier unchanged; consensus_nt and histamine handling align with the audit, while monoamine-as-fast-excitatory remains an unresolved dynamical choice.
full-scale Monaco driver plus implemented Kuala Lumpur road-law/perception transition; KL success not yet demonstrated
Origem
independent
Project ID
hotocoo-malecns-monaco
Repositório
hotocoo/malecns
Full-scale MaleCNS v1.0 spiking controller connected to a Monaco racing environment with anatomically named visual and descending/motor populations, deterministic evaluation tooling and a detailed training-loop audit. The engineering evidence is substantial, but no matched topology-destroying or simpler-policy control is yet reported for the central driving result.
Ver evidência e limites
Evidência mais forte
The runtime steps all 166,700 annotated neurons and a 6.24-million-edge thresholded graph with neurotransmitter-derived connection signs.
The public implementation includes deterministic multi-start evaluation, long-horizon and stress-track tools, and more than one hundred tests.
The reported calibrated/DAgger driver laps full-scale Monaco from all configured starts without a crash before ES refinement.
The current public revision adds Kuala Lumpur route geometry, Malaysia road-law data, camera/perception, scene, route and law-reward code with dedicated tests; this is credited as scope expansion, not as a successful MaleCNS KL driving result.
Limitações
The sensory transduction, readout projection, vehicle physics and optimization procedure are engineered around the task.
The graph builder uses consensus_nt and correctly treats histamine as inhibitory, but maps dopamine, serotonin and octopamine to fast excitatory drive; the public MaleCNS NT audit identifies those amines as a distinct modulatory class, so the driving result has not yet been stress-tested against that modeling choice.
The project does not yet report an ensemble of matched rewired graphs or a comparably trained small conventional controller for the same evaluation suite.
Repository audit quality and internal deterministic tests are not equivalent to independent replication.
Controles
Deterministic evaluation, stress tracks and a documented training-loop audit constrain implementation artifacts.
The independent MaleCNS NT audit (https://doi.org/10.5281/zenodo.22975837) supplies a concrete alternative transmitter-class mapping that should be rerun against the same lap suite.
A topology-destroying matched graph control remains necessary before assigning a MaleCNS-specific advantage.
Histórico do tier
2026-09-26: initial placement -> scientific C / interest S, high confidence; full-scale system and audit credited, topology attribution withheld.
2026-09-26: transmitter-mapping clash review -> tier unchanged; consensus_nt avoids the audit's source-column error, while monoamine-as-fast-excitatory remains a material dynamical sensitivity to test.
GPU whole-connectome driven forecasting demonstrated; autonomous rollout fails to reproduce the attractor
Origem
independent
Project ID
jangyeongsil-malecns-reservoir-computing
Repositório
JangYeongSil69420/malecns-reservoir-computing
A full MaleCNS-derived sparse recurrent reservoir with a trained linear readout reaches strong driven Mackey-Glass forecasts and preserves a clear negative autonomous-rollout result. The evidence shows that the implemented fixed connectome reservoir is computationally usable, but it does not isolate an advantage of biological wiring over matched random or rewired reservoirs.
Ver evidência e limites
Evidência mais forte
The project builds a sparse GPU recurrent reservoir from the full MaleCNS graph while leaving recurrent connectome weights fixed and training only a Ridge readout.
Reported held-out driven forecasting reaches R2 0.9933 at one step and remains above 0.8 at the 20-step direct horizon.
The autonomous 200-step feedback rollout is reported to fail quickly into a large-amplitude regular limit cycle, and that negative result is treated as a central limitation rather than omitted.
The repository provides an end-to-end notebook, pinned data acquisition/checksums, sparse GPU execution and explicit single-channel versus multi-channel measurements.
Limitações
The multi-channel gain confounds anatomical routing with richer engineered transforms of the same scalar signal: raw value, velocity and nonlinear acceleration.
No matched degree-preserving rewire, random recurrent network or simple conventional reservoir is reported under the same readout and feature budget.
Mackey-Glass is a synthetic forecasting benchmark and does not establish biological behavior or task-general usefulness.
Strong driven forecasting does not survive autonomous generation in the reported closed-loop rollout.
Controles
Single-channel versus three-channel injection is reported, but this is not a topology control because the information representation also changes.
The autonomous rollout acts as an adversarial generative check and produces a clear negative result.
Missing: biological topology versus degree-matched rewire/random ESN under identical input transforms, sampled-state dimension and Ridge protocol.
Histórico do tier
2026-09-26: initial placement -> scientific C / interest A, high confidence; strong reproducible driven result credited, topology-specific claim withheld.
MaleCNS subcircuit controller embedded in an artificial-life world
Estágio
results-bearing subcircuit experiments plus exploratory ecosystem
Origem
independent
Project ID
ikaankeskin-malecns-simulation
Repositório
ikaankeskin/malecns-simulation
An inspectable 2D artificial-life environment includes a verified 11-neuron, 32-edge MaleCNS DNg13 subgraph and explicit held-out controller comparisons. The project correctly treats motor-decoder reversal as an interface choice, while the larger ecology and social-learning layers remain engineered exploratory systems rather than MaleCNS mechanism claims.
Ver evidência e limites
Evidência mais forte
The committed DNg13 extract contains 11 verified MaleCNS v1.0 neurons and 32 connections and can be replayed deterministically.
A development-seed mapping sweep is followed by held-out comparisons against disconnected, input-silenced and shuffled-source controls.
On held-out seeds 10-19, the all-positive controller averages 4.0 collected items while the transmitter-filtered variant averages 4.1; the repository explicitly treats this as an interface result rather than evidence for food coding or inhibition.
Limitações
The neural substrate is a tiny selected subgraph, not whole-MaleCNS, and the decoder sign can dominate whether the agent approaches or avoids the pellet.
Most ecosystem, reproduction, communication, personality and lifetime-learning mechanics are external engineered rules with fixed MaleCNS topology.
Five-seed ecology/social comparisons are explicitly exploratory and do not establish general survival or cooperation benefits.
Controles
Disconnected, input-silenced and shuffled-source controller controls are evaluated on held-out seeds.
A transmitter-sign comparison preserves the null/near-null outcome instead of treating transmitter annotations as validated dynamics.
Whole-network rewires or capacity-matched conventional controllers are not reported for the larger artificial-life claims.
Histórico do tier
2026-09-26: initial placement -> scientific C / interest A, high confidence; real subcircuit and held-out controls credited, larger artificial-life behavior kept outside the MaleCNS evidence claim.
The card scores the DNg13/MaleCNS evidence; the broader world is an experimental scaffold, not a biological validation.
public full-connectome MuJoCo closed-loop demo with sensory A/B gates
Origem
independent
Project ID
noir-infini-pianist-fly
Repositório
Noir-infini/pianist-fly
Full-MaleCNS LIF simulation coupled to a MuJoCo fly and piano task, with olfactory input, lateral neural readout and physically verified key contact. The implementation makes its engineered IK and gating boundary unusually explicit, but it does not yet isolate a connectome-specific advantage against topology-destroying or simpler-controller controls.
Ver evidência e limites
Evidência mais forte
The public implementation runs a 166,700-neuron / roughly 25.58-million-synapse MaleCNS graph with static connectome weights and no backpropagation through the recurrent network.
Sugar-plume inputs stimulate annotated antennal sensory neurons; connectome activity gates left/right leg choice, while MuJoCo IK executes and verifies physical key contact.
The repository documents a five-gate scent A/B suite, including wrong-aim refusal and a symmetric centerline case, plus a reproducible setup path with pinned source downloads.
Limitações
The spatial sugar plume, hunger model, lateral readout gate, piano layout and inverse-kinematics trajectory are engineered interfaces rather than biological measurements.
The connectome selects a coarse lateral orientation while a conventional IK controller performs the detailed reaching and contact.
No matched degree-preserving rewire, random recurrent network or small conventional controller is reported for the same piano task.
Controles
Scent A/B and wrong-aim gates constrain accidental presses and test the sensory-to-side pathway.
A topology-destroying or same-budget conventional-controller comparison remains necessary for a MaleCNS-specific behavioral claim.
Histórico do tier
2026-09-26: initial placement -> scientific C / interest S, high confidence; whole-connectome embodiment credited, topology-specific attribution withheld.
Portable MaleCNS runtime with measured cross-backend execution and deliberately narrow scientific claims. Engineering evidence is strong; biological validation is explicitly outside the demonstrated result.
Ver evidência e limites
Evidência mais forte
Full-scale MaleCNS-derived execution is reported on CPU and WebGPU with parity for the tested configurations.
A deterministic fixed-point runtime contract and reproducible validation commands are public.
Measured, estimated and not-tested claims are labeled separately.
Limitações
Biological calibration and validation are explicitly not demonstrated.
Several accelerator and portability regimes remain untested at full scale.
Runtime execution alone does not establish a MaleCNS computational or behavioral advantage.
Controles
Cross-backend parity and deterministic propagation fixtures test the runtime claim.
Histórico do tier
2026-09-26: initial placement -> scientific C / interest A, high confidence; strong runtime evidence, no biological-mechanism claim.
end-to-end paper-trading system; useful learning not demonstrated
Origem
independent
Project ID
nftechie-stonkfly
Repositório
nftechie/stonkfly
A full retained MaleCNS graph is connected to real public Coinbase market observations and a fixed neural proposal interface, with paper trading, resumable state and explicit safety/audit boundaries. The project demonstrates an unusually complete end-to-end experimental apparatus, but neither profitable learning nor a MaleCNS-specific trading advantage has been demonstrated.
Ver evidência e limites
Evidência mais forte
The importer retains 166,700 MaleCNS nodes and 25,582,938 directed connections, with source hashes and derived graph arrays checked against committed locks.
Public Coinbase observations are rendered into an RGB stimulus, propagated through the retained graph, and converted by a fixed descending-neuron readout into buy, sell or hold proposals; the external guard may veto but not invent a replacement trade.
The paper-mode system preserves local neural state, audit logs, fees and resumable execution, and exposes a frozen-memory control mode for mechanism checks.
Limitações
The chart-to-photoreceptor adapter and DNp20/DNpe017 trading readout are engineered interfaces rather than discovered biological market semantics.
Profit/loss dopamine assignments and the KC-to-MBON memory rule are engineered; useful credit assignment and strategy improvement are explicitly unproven.
No held-out chronological result establishes profit, and no matched no-connectome, rewired or random-recurrent controller has demonstrated a MaleCNS-specific advantage.
Controles
A frozen-memory fixture mode is available to separate plasticity plumbing from weight-changing runs.
The repository explicitly requires independent starts, frozen-weight and shuffled-reinforcement controls, equal budgets, retention/reset tests, fees/slippage and cash/simple-exposure baselines before any learning claim.
A matched topology-null or no-connectome policy comparison remains missing.
Histórico do tier
2026-09-26: initial placement -> scientific C / interest S, high confidence; end-to-end auditability and whole-graph integration credited, trading-skill and learning claims withheld.
Scientific tier measures what the public artifact demonstrates, not whether live trading would be a good idea.
results-bearing; auxiliary controller line saturated; frozen Run 39 blocked on pinned binary cache
Origem
own
Project ID
franklinbaldo-whole-cns-speed-hold
Repositório
franklinbaldo/franklinbaldo.github.io
A results-bearing whole-connectome closed-loop programme whose strongest current conclusion remains negative and informative. The latest fresh whole-CNS result is Run 37: standardization did not rescue the responsive-16 representation, while the adapter-only arm stayed near the conventional PI floor. Runs 40-49 explored interface/controller timing but are now saturated and do not constitute new evidence for MaleCNS topology. Frozen Run 39 is the next whole-CNS discriminant and is blocked only on staging checksum-pinned MaleCNS binary assets into an eligible executor.
Ver evidência e limites
Evidência mais forte
Run 26 established causal participation of the 166,700-neuron, 25,582,938-edge whole MaleCNS in the C1 loop but did not beat the conventional PI baseline or establish topology advantage.
Run 37 is the latest fresh whole-CNS result: standardization improved conditioning but did not rescue responsive-16 spike/delta control, while adapter17 remained near the OBD-only PI floor.
The auxiliary Run 40-49 sequence preserved useful interface diagnostics but failed to unlock the whole-CNS path; the canonical state now marks that line saturated rather than manufacturing a Run 50.
Limitações
No robust topology-specific control advantage has survived the reported rewired, adapter-only and conventional comparisons.
Run 39 is planned and frozen but not executed; it has no selected whole-CNS lambda, fresh held-out MAE or topology result yet.
The remaining operational blocker is the exact checksum-pinned brain.npz and weights.npz cache in an eligible executor, not another controller-design question and not hosted CI.
Controles
Degree-preserving rewired graph.
Adapter-only no-graph control.
Conventional PI controller and causal history baselines.
Run 39 prospectively freezes OLS, fixed ridge and train-only-selected ridge without expanding sensors or readout capacity.
Histórico do tier
2026-09-26: initial placement -> scientific C / interest S, high confidence, matching the public whole-CNS negative-evidence assessment.
2026-09-26: status refreshed through Run 49 and canonical execution state; tier unchanged because interface/controller ablations did not create new topology evidence, auxiliary line marked saturated, Run 39 remains blocked on pinned binary assets.
The null result counts as evidence; scientific C reflects the current claim boundary, not failure to obtain a positive result.
MaleCNS-derived visual/olfactory subgraph rover and demo suite
Estágio
working public demo suite with engineering tests but no controlled behavioral benchmark
Origem
independent
Project ID
chihhsiangchien-malecns-rover
Repositório
ChihHsiangChien/maleCNS
Public browser/Python demo suite that traces MaleCNS visual and olfactory pathways into a 3D rover, neural telemetry workbench and arcade experiment. It ships real extracted connectome matrices and runnable code, but most of the sensing-to-action stack is an engineered Reichardt/matrix-control pipeline, and the published tests are primarily unit/synthetic checks rather than a controlled closed-loop comparison showing a MaleCNS-specific benefit.
Ver evidência e limites
Evidência mais forte
The repository ships extracted MaleCNS pathway matrices and a documented top-down tracing path for LC4, LC11, HS/VS and ORN-PN-LHON circuitry rather than using only a generic recurrent proxy.
The Python and browser implementations expose runnable visual-motion, connectome-matrix, neural-telemetry, 3D-rover and chemotaxis demos, with public GitHub Pages entry points.
The automated test suite checks matrix acquisition/export, retinal preprocessing, excitatory/inhibitory polarity handling and telemetry plumbing end to end at the module level.
Limitações
This is not whole-MaleCNS simulation: it uses selected visual/olfactory pathways aggregated into engineered low-dimensional matrices.
The unit tests exercise synthetic matrices and module behavior; they do not report a held-out navigation benchmark, repeated success rate or causal topology ablation.
Reichardt motion extraction, spatial binning, LC11 gain, odor field, wing/thrust equations and several escape behaviors are engineered outside the measured connectome.
The repository's narrative comparison with hand-written CV/PID/FSM is not a matched empirical baseline.
Controles
Synthetic looming, yaw, pitch and small-target stimuli provide pathway-level smoke tests.
Unit tests verify sign handling and module plumbing but do not destroy or replace the connectome topology.
Missing: matched rewired/degree-preserving graph, random recurrent network and conventional controller under the same 3D navigation task.
Histórico do tier
2026-09-26: initial placement -> scientific D / interest A, high confidence; real public artifact credited, behavioral/topology claims bounded by synthetic tests and engineered interfaces.
The tier treats the rover as a useful MaleCNS-derived demo, not evidence that the selected biological wiring outperforms a conventional controller.
reproducible single-seed closed-loop demo with external selector scaffolding
Origem
independent
Project ID
tolga-ileri-flycoder
Repositório
tolga-ileri/FlyCoder
A frozen MaleCNS-derived FlyBrain is placed in a real closed loop for a CSS centering task, but an external Q-biased action cursor and idle-commit fallback materially scaffold the successful behavior. The artifact is reproducible and unusually explicit about its interface boundary; it does not yet demonstrate a MaleCNS-specific control advantage.
Ver evidência e limites
Evidência mais forte
The demo keeps MaleCNS-derived recurrent weights frozen while routing experimental visual inputs through identified projection-neuron groups and reading descending-neuron activity.
A documented headless seed-64 replay reports centered=True after 10 attempts and 180 brain steps.
The implementation exposes whether actions came from neural commit, reject, or the external idle fallback, making the control boundary auditable.
Limitações
The published success is a single-seed deterministic demonstration rather than a replicated held-out evaluation.
On tied DNa02 laterality, an external selector moves toward untried or higher-Q actions; after eight quiet cycles it commits the current cursor even without a descending commit signal.
No matched rewired connectome, random recurrent network, no-connectome selector, or compact conventional controller is reported for the same task.
Controles
Connectome weights remain frozen and experimental encoder, decoder, reward and selector logic are explicitly separated from the MaleCNS substrate.
A deterministic headless replay is provided, but Q-bias-off, idle-commit-off and topology-destroying ablations are still needed to isolate the MaleCNS contribution.
Histórico do tier
2026-09-26: initial placement -> scientific D / interest A, high confidence; reproducible closed loop credited, externally scaffolded task success kept below topology-evidence tiers.
The tier evaluates the MaleCNS evidence, not the polish or creativity of the demo.
public architecture prototype; discriminating experiment pending
Origem
own
Project ID
franklinbaldo-flygenesis
Repositório
franklinbaldo/franklinbaldo.github.io
Public architecture prototype that reuses the FlyDoom MaleCNS recurrent substrate inside a morphology-evolution loop. The implementation is material, but the matched experiment needed to show a MaleCNS advantage over simpler alternatives has not yet been reported.
Ver evidência e limites
Evidência mais forte
A runnable browser prototype loads the public compact MaleCNS asset and routes synthetic sensory signals to descending readout.
The implementation includes morphology mutation, champion selection and a progressive terrain curriculum.
The article clearly separates MaleCNS topology from synthetic body, sensors, adaptor and fitness.
Limitações
No reported evolved-body outcome, fitness trajectory or held-out generalization establishes the proposed scientific payoff.
No matched MaleCNS-versus-random-reservoir or small-network result is reported for the same search budget.
Functional protein labels are abstractions rather than molecular simulation.
Controles
Matched random-reservoir and small-network comparisons are proposed but not yet results.
Histórico do tier
2026-09-26: initial placement -> scientific D / interest A, medium confidence; implemented demo, prospective scientific question.
working practice-mode integration with reward-modulated plasticity and steering scaffolds
Origem
independent
Project ID
tsungyumr-malecns-lol
Repositório
tsungyumr/malecns_LoL
A materially implemented MaleCNS-to-game loop with virtual vision, reward-modulated internal plasticity and detailed action observability. The current executed policy is still substantially shaped by deterministic steering and tactical scaffolds, so the project has not yet demonstrated that MaleCNS dynamics or plasticity are responsible for competent game behavior.
Ver evidência e limites
Evidência mais forte
The runtime loads a cached MaleCNS graph, keeps the biological graph as a fixed prior and adds checkpointed reward-modulated plastic components on selected internal edges.
The project separates dry-run observation from optional practice-mode actuation and records both brain-requested and actually executed movement.
Plasticity can be disabled and checkpoints can be ignored for fresh runs, providing useful infrastructure for future controlled comparisons.
Limitações
The current graph weights are unsigned, the 128-sector visual adapter is not official MaleCNS retinotopy and the motor projection is fixed and synthetic.
A deterministic enemy-steering layer and a separate tactical value learner can determine or override executed behavior while MaleCNS motor mappings remain unstable.
No reported held-out comparison yet establishes a behavioral advantage for MaleCNS plasticity over frozen, rewired, random-recurrent or conventional policies.
Controles
The runtime exposes plasticity-off, fresh-checkpoint and brain-move-versus-executed-move observability.
A matched frozen/null/controller evaluation is still required before promoting the central learning claim.
Histórico do tier
2026-09-26: initial placement -> scientific D / interest S, high confidence; real closed-loop integration credited, MaleCNS-specific gameplay claim remains prospective.
The project explicitly scopes actuation to private/practice/custom research use; the tier reflects the public research artifact.
fruit-fly ecology simulator with optional whole-MaleCNS neural control and visualization
Estágio
released interactive simulator with optional neural bridge
Origem
independent
Project ID
welkinhh-flyworld-malecns
Repositório
welkinhh/FLYWORLD-MaleCNS
A released Godot ecosystem can attach a whole-MaleCNS service to a selected adult, inject food/threat channels into declared sensory groups, read locomotor/escape populations and optionally use those readouts for action selection. The wider ecology remains playable without the brain and is governed largely by simplified external rules, so the current evidence is integration and observability rather than MaleCNS-specific capability.
Ver evidência e limites
Evidência mais forte
The Python brain service instantiates the full FlyBrain/MaleCNS model and addresses taste plus left/right looming inputs and forward, steer, escape, backward, punch and kick output groups.
The Godot adapter carries source-ID-addressed neural activity across a versioned WebSocket protocol and rejects stale responses before visualization or optional control use.
The game exposes an optional brain-control path for the selected adult while explicitly distinguishing genuine MaleCNS activity from the local proxy.
Limitações
Only the selected fly is stepped through the remote neural model in the interactive path; the wider ecology continues under authored rules and can run with no neural service.
Foraging, reproduction, predators and environmental dynamics are simplified game mechanics and are not biological predictions.
No matched no-brain, rewired, random-recurrent or conventional-controller benchmark is reported for behavior or survival.
Controles
The adapter prevents proxy responses from being presented as genuine neural activity and validates request context before accepting neural results.
The repository explicitly states that neural signals are connectome-based computations, not living neural recordings.
Behavioral ablations against proxy, no-brain, rewired and conventional controllers remain missing.
Histórico do tier
2026-09-26: initial placement -> scientific D / interest A, high confidence; whole-MaleCNS integration credited, behavioral attribution withheld for lack of matched controls.
This is currently strongest as an embodiment/visualization artifact, not as evidence that MaleCNS improves ecological behavior.
whole-connectome market experiment / paper-trading system
Estágio
end-to-end paper mode with frozen walk-forward experiment gate; no demonstrated profitable edge
Origem
independent
Project ID
marketcalls-openfly
Repositório
marketcalls/openfly
A unusually complete MaleCNS-to-market experimental system with whole-connectome simulation, walk-forward evaluation, strong non-neural trading controls and a dopamine-plasticity arm. The core scientific question is still unresolved: the repository explicitly reports no demonstrated profitable edge and does not yet publish the held-out result that would pass its own experimental gate.
Ver evidência e limites
Evidência mais forte
The system simulates the 166,700-neuron MaleCNS connectome and maps market observations through 4,146 photoreceptors into the recurrent substrate.
The experiment harness separates train, validation and untouched test windows and requires comparison against fixed-entry, matched random-entry, shuffled-label and flat controls.
The plasticity arm stimulates named dopamine populations and updates 7,835 Kenyon-cell-to-memory-output connections while retaining a frozen-memory twin comparator.
Paper trading, replay, execution safety and experiment infrastructure are implemented end to end, with 328 backend and 13 frontend tests reported.
Limitações
No held-out result demonstrating predictive or trading value has been published; the README explicitly states that no profitable edge is demonstrated.
Trade timing, risk sizing, stops, targets, execution and guardrails are conventional external rules rather than decisions made by MaleCNS.
The statistical readout is trained externally on future market movement targets.
Live trading remains locked behind an explicit opt-in and a passed experiment.
Controles
The planned walk-forward gate includes fixed 09:20 straddle, same-count random-entry, shuffled-label and stay-flat controls.
The plasticity arm is compared against a twin with frozen memory.
Missing: a matched rewired MaleCNS, random recurrent or small conventional representation under the same encoder/readout/trading protocol.
Histórico do tier
2026-09-26: initial placement -> scientific D / interest S, high confidence; experimental design and implementation credited, scientific tier held down pending the preregistered held-out result.
Fonte canônica: cartões OKF em knowledge/malecns-projects/. Mudança de tier exige evidência material. Resultado negativo não é punição: um controle que derruba a hipótese melhora o registro.