ecosystem listings - Capsize-Games/spikeforge GitHub Wiki

Ecosystem listings — getting spikeforge into the two directories this field searches

Status: both submitted on 2026-09-15, after the 0.3.5 release put the corrected README and metadata on PyPI.

The NIR pull request deviates from the draft below in one place, deliberately. The draft assumed a plain ✓ for Read from NIR. spikeforge loads a NIR graph and executes it through the independent interpreter; it does not reconstruct a native module the way the other listed frameworks do (nir_bridge/stage_builders.py and stage_mapping.py are export-direction only, and nothing builds a TopologySpec from a graph). The pull request states that difference and invites the maintainers to mark the cell ⬚ if they read the column as "convert to native" -- overclaiming a cell in the table this project's positioning rests on would cost more than the row is worth.

Why this is the promotion work worth doing

Searching for the name already works: "spikeforge spiking neural network" returns the GitHub repository and all seven PyPI projects at the top. What does not work is categorical discovery — someone browsing the NIR ecosystem, or searching "SNN framework comparison", never encounters the project. Two directories account for almost all of that traffic, both are free, and both explicitly invite submissions.

This is deliberately sequenced after the PyPI and documentation fixes. Reviewers of either submission will click straight through to https://pypi.org/project/spikeforge/ and https://docs.spikeforge.net/, so those pages had to be right first.

1. NIR's framework support table

Where: https://github.com/neuromorphs/NIR — the README's framework support table.

Verified state (2026-09-14): ten frameworks listed, with columns Framework, Write to NIR, Read from NIR, Examples:

Listed framework Write Read
hxtorch (BrainScaleS-2) ✓ ✓
jaxsnn (BrainScaleS-2) — ✓
Lava-DL — ✓
Nengo ✓ ✓
Norse ✓ ✓
Rockpool (SynSense Xylo) ✓ ✓
Sinabs (SynSense Speck) ✓ ✓
snnTorch ✓ ✓
SpiNNaker2 — ✓
Spyx ✓ ✓

spikeforge is absent, despite supporting both directions. The README has no documented contribution process for the table, so the route is a pull request adding the row and a linked example.

What makes the row worth accepting rather than just longer: none of the ten re-executes an exported graph through an independent interpreter and reports numerical drift against the original model. That is a NIR-specific capability, which is exactly the argument to make — not "please list us."

Draft row:

| [spikeforge](https://github.com/Capsize-Games/spikeforge) | ✓ | ✓ | [spikeforge examples](https://github.com/Capsize-Games/spikeforge/blob/main/examples/04_nir_export_validate.py) |

Draft PR description:

Adds spikeforge to the framework support table.

spikeforge (BSD-3-Clause, pip install spikeforge) reads and writes NIR through spikeforge.nir_bridge. Beyond export/import it ships an independent NIR interpreter that re-executes the exported graph and reports per-tensor numerical drift against the source model, so an export can be validated rather than assumed:

pip install "spikeforge[nir]"
spikeforge-verify validate --topology conv_net   # non-zero exit outside tolerance

Runnable example: examples/04_nir_export_validate.py. Docs: https://docs.spikeforge.net/interpreter-spine.

Before filing, check: that examples/04_nir_export_validate.py still runs end to end on a clean install, and that the drift claim in Interpreter spine still matches the shipped tolerance.

2. Open Neuromorphic's software guide

Where: https://open-neuromorphic.org/neuromorphic-computing/software/ — the de-facto directory for this field.

Verified state (2026-09-14): 27 SNN frameworks, 8 data tools, 1 simulator. Neither "spikeforge" nor "Capsize" appears. The page states:

"Help us keep the software guide comprehensive and up-to-date. Suggest new frameworks, data tools, or corrections by opening an issue on our GitHub repository."

Route: an issue at https://github.com/open-neuromorphic/open-neuromorphic.github.io/issues/new/choose.

Draft issue body:

Framework: spikeforge Repository: https://github.com/Capsize-Games/spikeforge Docs: https://docs.spikeforge.net/ Install: pip install spikeforge License: BSD-3-Clause Language / backend: Python, built on snnTorch and PyTorch

What it is: an SNN toolkit that sits on top of snnTorch rather than replacing it. Rate/latency/delta/random spike encoding, LIF training across eight image datasets and four neuromorphic event datasets, NIR export with an independent interpreter that re-executes the exported graph and reports numerical drift, a per-target deployment capability matrix, an event-driven energy estimator (labelled as an estimate, not hardware-measured), and a live browser training/introspection dashboard.

Maturity: pre-1.0, published on PyPI, CI green, reference accuracy numbers published at https://docs.spikeforge.net/benchmarks.

Before filing, check: the maturity sentence is still true, and that the benchmarks page is live at that URL.

What not to do

Neither submission is a place to claim more than the project delivers. The capability matrix's spec-only targets (spinnaker2, speck, xylo) are registered but have no installable SDK integration; the energy figure is an estimate; the model hub carries a handful of this project's own reference checkpoints, not a zoo. Every one of those is stated plainly on the project's own surfaces, and the submissions must not soften any of it — a listing earned by overclaiming is worse than no listing.