From install to action.
Python 3.9+ and NumPy. No CUDA required. The package is currently installed from source; it has not yet been published to PyPI.
git clone https://github.com/freeman-1984-coder/flybrain-sdk.git
cd flybrain-sdk
python -m venv .venv
diff --git a/site/contribute.html b/site/contribute.html
index 5945818..15ce27d 100644
--- a/site/contribute.html
+++ b/site/contribute.html
@@ -4,7 +4,7 @@
-
+
Skip to contentflybrain_sdk Community / Open sourceBuild with us.
No GPU needed.
Fork the repository, make a focused change and open a pull request. Contributions are reviewed before merging, and pull requests run automated tests.
Useful first contributions
- Add a versioned model URL with size, license, attribution and checksum.
- Build a small MaleCNS/FlyWire subgraph importer with tests and a model card.
- Extend the Godot adapter or add Unity and other environments.
- Port the LIF reference core to WASM and compare traces with Python.
- Improve examples, translations and error messages.
The contribution loop
- Open an issue for a substantial design change, or pick an existing issue.
- Fork, branch and install the Python development dependencies.
- Make the change and add meaningful tests or examples.
- Run tests and the quickstart, then open a pull request.
Validationpython -m pip install -e ".[dev]"
pytest
python examples/quickstart.py
diff --git a/site/demos.html b/site/demos.html
index 0c0c076..0219d9c 100644
--- a/site/demos.html
+++ b/site/demos.html
@@ -4,14 +4,14 @@
-
+
Skip to contentflybrain_sdk Demo kits / Same core / No CUDA required
Observe. Connect.
Make it your own.
One Python session loop connects neural activity to a small game environment or synthesized sound. Run a preset, inspect the inputs and controls, then replace an adapter.
Choose the experience: the game sandbox and circuit lab run live simulations. The demos below are verified recordings from Python; playback controls inspect a run without recalculating its brain.
Full-brain voxel world · recorded GPU experiment
All 139,255 FlyWire neurons run on an A16 GPU and control a ground body through an explicit DNa02 readout. Compare the actual trajectory with a silenced-input control. It moves, but curves away from food; no successful navigation or training is claimed.
Full FlyWire brain · CUDA benchmark
All 139,255 proofread neurons and 16.85 million source rows on a real NVIDIA GPU. CPU/CUDA numerical checks, throughput, checkpoint replay and continuous simulation; this is a full-graph software benchmark, not a trained game policy.
Reflex School · reward-trained readout
A fixed real 313-neuron circuit, three learned external action weights, and an actual A16 GPU run. On 80 balanced held-out two-choice trials, accuracy changed from 50% to 100%. This is readout learning, not biological plasticity.
Godot connectome game adapter
No CUDA required. Run an actual Godot 4 scene with the Python CPU SDK. Godot owns obstacles and movement; a replaceable encoder and neural readout supply steering. Includes pause, paired world/brain saves, restore and stop-on-disconnect behavior.
-Python 3.9+ and Godot 4.7.2 · two terminalsgit clone --branch v0.4.0a4 https://github.com/freeman-1984-coder/flybrain-sdk.git
+Python 3.9+ and Godot 4.7.2 · two terminalsgit clone --branch v0.5.0a1 https://github.com/freeman-1984-coder/flybrain-sdk.git
cd flybrain-sdk
python -m pip install -e .
python examples/godot/bridge.py
@@ -19,7 +19,7 @@ Full-brain voxel world · recorded GPU experiment
All 139,255 FlyWire
godot --path examples/godot
Alternatively import examples/godot/project.godot in Godot and press F5. Wait for the bridge to print its ready address, then click Run / pause. The artificial toy works offline after installation. For the real 313-cell model, restart the bridge with --model male-cns-escape-v1 --download.
The scene advances 20 ms per acknowledged action, independently of rendering. Save completes any outstanding action before writing a checkpoint. Restore both the brain and the world together. The local HTTP bridge is a development example; it is not a hosted multiplayer service.
-Godot setup and adapter guide Get the editable scene
+Godot setup and adapter guide Get the editable scene
Verified with toy and real circuits against Python feedback traces and whole-brain checkpoints. Real anatomical wiring uses assumed LIF dynamics and engineered steering; no trained avoidance or biological fidelity is claimed.
Neural dodge
A player moves left or right as obstacles descend. Neural outputs determine steering; the environment reports the movement actually applied. Both recordings use the same seed, real 313-cell model and six-second duration.
Watch the neural run Compare silenced outputs
diff --git a/site/foraging.html b/site/foraging.html
index a918c99..010ea3e 100644
--- a/site/foraging.html
+++ b/site/foraging.html
@@ -22,7 +22,7 @@ 把完整脑，放进一个小世界。
0 / 2000 ms
绿色：实际轨迹黄色：食物气味源拖动场景：旋转观察角度
这一帧，脑子输出了什么？
等待实际记录。
如何读这个实验
比较正常神经输入和关闭嗅觉入口的同初始状态对照。轨迹来自实际记录；网页不会让角色自动朝向香蕉。
结果以记录为准，包括没有接近或接触食物的情况。
哪些是连接组，哪些是我们的假设？
全部脑神经元和连接来自 FlyWire v783。输入只覆盖一个 DM1 食物气味通道，不代表完整香蕉气味或视觉识别。脑模型采用简化的突触 LIF 动力学。
动作层是未经训练的工程映射：DNa02 两侧放电频率之差控制转向、频率之和控制速度。尤其是速度映射没有生物学验证。身体是地面运动模型，没有飞行、肌肉或六足步态仿真。记录中的移动不等于可靠觅食。
哪些是连接组，哪些是我们的假设？
全部脑神经元和连接来自 FlyWire v783。输入只覆盖一个 DM1 食物气味通道，不代表完整香蕉气味或视觉识别。脑模型采用简化的突触 LIF 动力学。
动作层是未经训练的工程映射：DNa02 两侧放电频率之差控制转向、频率之和控制速度。尤其是速度映射没有生物学验证。身体是地面运动模型，没有飞行、肌肉或六足步态仿真。记录中的移动不等于可靠觅食。
怎么知道该刺激哪些嗅觉神经元？
香蕉图标在这里标记气味源的位置。我们没有让模型识别画面中的香蕉，也还没有重建香蕉的多种挥发物混合气味。
当前输入依据是：乙酸乙酯可以激活 Or42b 嗅觉通路，相关神经元投射到 DM1。我们根据固定版本的 FlyWire 细胞注释，选择左侧 35 个、右侧 33 个 ORN_DM1 神经元；它们只是入口，全脑的 139,255 个神经元仍参与计算。
diff --git a/site/fullbrain.html b/site/fullbrain.html
index d78885c..8c9b498 100644
--- a/site/fullbrain.html
+++ b/site/fullbrain.html
@@ -29,7 +29,7 @@ 实际资源占用
测试卡显存 7.82 GiB；结束时 CuPy 内存
这里比较的是累计脉冲。刺激刚开始时存在短暂的左右差异，因此不能断言完全没有方向信息。另外，双侧刺激停止后的 200 毫秒内，下降神经元仍新增了 859 次脉冲；这种持续活动需要进一步校准，不能直接称作生物记忆。
新引擎通过 5 组真实 GPU 检查；完整图在 200 个有嗅觉输入的时间步中，5 个检查点的全体神经元电压、突触状态和频率最大差异为 0。全脑状态连同输入随机数状态，恢复后 20 步完全重现。同机 204 项软件测试通过，零跳过。
每组包含 100 毫秒静息、500 毫秒刺激、200 毫秒恢复，以 0.1 毫秒步长运行。每组实际耗时约 102–104 秒，包含记录开销。这不是实时模拟，也不与上方不同参数的旧基准直接比较。输入是 Or42b/DM1 单通道食物气味近似，尚未校准为香蕉的真实化学剂量。
-
+
Full FlyWire synaptic CUDA experiment: sensory-only input reaches ALPN, MBON and descending populations; matched no-odor and silenced-input controls remain silent. DNa02 laterality does not reverse with stimulus side. Navigation is not established.
旧基准配置：为什么信号停在入口
2026 年 9 月 13 日，在 NVIDIA A16-8Q 上运行相同的完整 139,255 神经元网络，只刺激经标注确认的 68 个 ORN_DM1 感觉神经元。每组从同一静息状态开始，包含 100 毫秒静息、500 毫秒刺激和 200 毫秒恢复。
@@ -38,9 +38,9 @@ 实际资源占用
测试卡显存 7.82 GiB；结束时 CuPy 内存
结果：入口刺激和关闭操作有效，但这组参数没有产生下游放电。嗅觉投射神经元在采样时刻观察到的最高电压约 0.252，阈值为 1；有微弱膜电位响应，不等于完全没有信号。
原因已经定位到基准测试的权重配置：每个神经元的输入连接绝对权重之和被归一化到 2.5。在当前不应期下，即使上游以允许的最高频率同步放电，无外部电流的下游电压上限也只有 0.918，低于阈值 1。因此不能靠提高气味强度解决，也不能把之前的数值测试当作已验证的感觉—动作模型。
这次 GPU 环境中 181 项软件测试通过。数值验证通过与行为验证通过是不同的结论。上方的新模型实验已经补充并验证突触衰减、传递延迟与感觉刺激；本节保留旧配置的负面结果，便于复现和比较。
-
+
Full-brain olfactory CUDA probe: sensory inputs fire, downstream spiking absent with the benchmark normalization. This is a recorded negative functional result, not live inference, banana recognition, or successful navigation.
开发者如何复现
使用实验分支 feat/cuda-reference，按复现指南下载官方数据并安装可选 CUDA 依赖。
+开发者如何复现
使用 v0.5.0a1 版本，按复现指南下载官方数据并安装可选 CUDA 依赖。
python scripts/validate_fullbrain.py \
--data-dir data \
--output results/fullbrain.json
diff --git a/site/index.html b/site/index.html
index 3a04e9e..5365176 100644
--- a/site/index.html
+++ b/site/index.html
@@ -4,7 +4,7 @@
-
+
Skip to contentflybrain_sdk Python SDK / Open source / AlphaNo CUDA
required.
Bring a neural circuit into your game.
Load a model. Give it input. Read an action.
Play the live sandbox Python quickstart View source ↗CPU readyMIT licensedOffline demoquickstart.pyfrom flybrain import FlyBrain
brain = FlyBrain.load("toy", backend="cpu")
@@ -17,4 +17,4 @@
brain.save("brain.checkpoint.json")
restored = FlyBrain.restore(
"brain.checkpoint.json"
-)
What runs today: a synthetic offline circuit and an opt-in, 313-neuron real MaleCNS subgraph on your CPU. Select cells, inject currents, inspect activity and bind your own output channels. Try it in the browser, export a recording for Python replay, or download an editable HTML demo. The real model uses assumed dynamics. The experimental CUDA branch now has a complete FlyWire brain benchmark: 139,255 neurons. Watch the full-brain GPU voxel recording: neural output drives movement, but the current readout does not reach food. The released v0.4 alpha remains CPU-only; WASM is planned.A small API, from input to actionKeep the game loop yours.
01 / STIMULATESend a sensory signal
Food, looming on the left or right, and touch are mapped to explicit inputs in the demo circuit.
02 / STEPAdvance neural time
Run fixed simulation ticks on NumPy. Inspect voltages, spikes and firing rates. No GPU setup.
03 / ACTRead motor intensity
Bind your own named output channels, or use the toy walk, turn and jump values. Save an experiment and continue its trajectory.
A model catalog, not a giant installDownload only what you want to use.
Source URLs, versions, sizes and licenses live in a lightweight catalog. Assets stream to a local cache when you explicitly request them.
Model source Today Download Toy LIF circuit Runnable, synthetic Included MaleCNS escape circuit Runnable, experimental real subgraph 3.8 MB, SHA256 verified MaleCNS v1.0 Raw connectome data Choose connections, annotations or transmitters FlyWire v783 Raw connectome data Choose connections or neuron IDs
Build in the openThe next circuit could be yours.
Contribute a model source, a real-data importer, an engine adapter or a WASM implementation. You can start with a laptop and a focused pull request.
Start contributing