Roadmap

This is a plan, not a promise. Items will move up or down based on community engagement.

v0.1 — spec draft + L0 emulator ✅ (this release)

• Layered architecture (spec/v0.1-architecture.md)
• Process scheduler spec + 3 policy implementations
• Neural migration spec + L0 emulator (literal weight copy)
• Health monitoring spec + L0 implementation
• Reference Izhikevich and Hodgkin-Huxley neuron models
• Pair-based STDP + homeostatic normalization
• Mini kernel exposing the stable behavioral API
• Mock Cortical Labs CL1-style API surface
• CLI + end-to-end demo + pytest suite

v0.2 — first L1 driver (seeking maintainer)

We want at least one L1 driver for real MEA hardware before v0.2. That means: someone with a CL1, FinalSpark, Open-Ephys, or compatible platform implements neural_os.core.neuron.NeuronModel and neural_os.core.organoid.Organoid against their hardware.

Target: ship by end of 2026 Q4, contingent on hardware access.

• At least one L1 driver PR merged
• L1 conformance test harness
• Spec clarifications discovered during L1 work
• Real trend history for health monitoring (the L0 emulator currently fakes this)

v0.3 — multi-culture scheduling and delta migration

Once two or more L1 drivers exist, we can meaningfully test multi-culture scheduling and migration between cultures.

• Multi-pool scheduler that owns more than one organoid
• Delta migration (spec/neural-migration.md §9): instead of re-training from scratch, transfer only what changed since the last snapshot
• Concurrent migration: two cultures migrate toward each other in parallel

v0.4 — reproducible benchmark suite

The current one-bit classifier is too trivial to drive spec evolution. We need real benchmarks.

• Pong (Cortical Labs DishBrain task, rebuilt on this stack)
• Pattern classification on a public dataset (MNIST → spike patterns)
• Delayed match-to-sample (working memory)
• Each benchmark ships with a reference performance curve on L0
• A spec for publishing a benchmark result (with HealthReport included, per the ethics doc)

v0.5 — driver ecosystem

• ≥3 independent L1 drivers from ≥3 different hardware vendors
• Driver conformance test suite that vendors can run in their own CI
• A public dashboard of driver conformance status

v1.0 — spec freeze

After at least 2 independent L1 implementations have shipped stable results against the v0.4 benchmark suite, we freeze the spec at v1.0 and adopt strict semver.

• ≥2 L1 implementations
• v0.4 benchmarks passing on each
• Ethics document has had ≥3 external reviews
• All v0.1 open questions in spec docs are either resolved or explicitly marked “deferred to v2.0”

Beyond v1.0 (intentionally vague)

• Multi-organoid distributed execution (a “cluster” of dishes)
• Hardware acceleration for the emulator (CUDA, Loihi 2)
• A standard “brain-state exchange” format (think: container images for organoids)
• Investigation of whether the spec generalizes beyond cortical organoids — e.g., to cardiac organoids, retinal organoids, or whole-brain slices

Non-goals (explicitly out of scope for the foreseeable future)

• Genetic engineering of the cultures. That is upstream of the OS.
• Closed-loop experimental ethics. That belongs in docs/ethics.md and the originating lab’s review, not in the spec.
• A standard machine-learning model. The OS provides the substrate; applications implement their own learning algorithms on top.
• A standard MEA hardware specification. That is a vendor concern.