Plans

spikeforge documentation

This site is generated from the authoritative design documents in plans/ and the project README. The markdown remains the single source of truth; the site is a browsable rendering of it. For user documentation — install, extras, CLI, and the honest limitations — start with the README.

Status — implemented. The professionalization program (workstreams WS-A…WS-F) described here shipped. See the README's "Model hub", "Backend execution", "Sequence primitives", "Event runtime and energy", "Operational maturity", and "Interop fold-ins" sections for what is delivered and where the honest boundaries are.

Roadmap

• Professionalization roadmap — the master program: six workstreams, interfaces, and the acceptance bar.
• Ecosystem roadmap — the earlier ecosystem plan this program builds on.
• Ecosystem listings — the two directories this field searches (NIR's framework table, Open Neuromorphic's software guide), what each currently lists, and drafted submissions for both.

Architecture

• Repo topology and split — the component inventory, dependency layers, and the staged recommendation for whether (and when) to split the client, server, and interpreter/deploy layers into separate repositories.

ARCH-0001 — phased repo split

The architecture for one repository with multiple distributions, now implemented. These documents are the design of record for the split; the spikeforge rename, the capsize-games organization move, and the spikeforge-targets, spikeforge-hub, and dashboard extractions have shipped. The spikeforge-server extraction did not fire (T4 no-go).

• ADR: repository topology — the accepted Option 0 decision, the comparison against Options 1–3, and the named extraction triggers.
• Target topology — the distribution names, import roots, GitHub repository names, and dependency direction.
• Core boundary — the forbidden-import list and the CI checks that enforce a headless core install.
• Protocol contract — the JSON Schema source of truth, the protocol_version field and compatibility policy, and TS type generation/validation.
• Packaging and versioning — the extras-to-package mapping, console-script ownership, pinning, and release automation.
• Migration plan — the reversible Phases 1–4, test/fixture relocation, and the no-back-compat-alias policy.
• Risk register — every split cost with a mitigation and an owner.
• Decision metrics — the numeric thresholds that gate the next extraction.

Focused workstream designs

• Interpreter spine — the TopologySpec single source of truth and the NIR translation.
• Model hub — curated catalog, downloads, and import/inspect (WS-A).
• Model-hub expansion — taking the catalog from static-image reference checkpoints to event vision and audio, the dataset provenance the schema does not yet record, and the missing event train/test split that blocks both.
• Backend execution — substitution execution and real backends (WS-B).
• Sequence primitives — sequence/ attention stage kinds and per-stage neurons (WS-C).
• Event runtime — sparse execution and energy accounting (WS-D).
• Operations — persisted metrics, external tracking sinks, determinism, and this docs site (WS-E).
• Interop fold-ins — ONNX, nirtorch, quantization, and geometry (WS-F).

Production use cases (PC-0)

The chip-less production-use-case program (issue #12). The shared toolkit every use case consumes is the hardware-free production toolkit; the umbrella and ticket seeds are production use cases. Each use case has a full specification mirroring UC-1 (problem, reference architecture, offline/online design, phases P0–P6 with MVP = P0–P4, dependencies, out of scope, and a ready-to-file GitHub issue payload).

#	Use case	Spec	Status
1	Streaming time-series classification / anomaly detection	UC-1	implemented (spikeforge 0.3.0)
2	Always-on audio / keyword spotting / wake-word	UC-2	scoped
3	Event-camera (DVS) vision	UC-3	scoped
4	Ultra-low-latency sensor-stream inference	UC-4	scoped
5	Anomaly / intrusion detection (IoT, network, grid)	UC-5	scoped
6	Bio-signal & medical monitoring	UC-6	scoped
7	RL for control & robotics	UC-7	scoped
8	Efficient sequence models / spiking transformers	UC-8	scoped
9	Edge/mobile inference under power budgets	UC-9	scoped
10	Computational neuroscience / neuromorphic R&D	UC-10	scoped

Recommended implementation order: UC-1 → UC-2/UC-5 → UC-3/UC-4 → UC-6/UC-7 → UC-8/UC-9/UC-10.

Memory-system research (issues #22-#26)

A research track, distinct from PC-0: one-shot, non-forgetting spiking memory, not a scoped production feature.

• Memory-system research — the held-out-digit POC (#22, shipped) and the camera-fed room memory pipeline design (#23, design only).

Guides

• Cookbook — practical, runnable recipes for every shipped capability, organised by goal.
• Examples — small, runnable, offline-safe scripts, one per core journey, with the real output each one prints.
• Open-source checklist — the pre-release readiness checklist (licensing, community, CI, packaging, docs).

Reference

• Project rules — the style and architecture contract.
• Integration plan — the original integration notes.

Building these docs

scripts/build_docs.sh          # build into build/docs
scripts/build_docs.sh --check  # fail on broken documentation links