miscellaneous research

---

id: miscellaneous-research title: Miscellaneous Research tags: [python, apple-silicon, cuda, video, cli, neuroscience, fmri] created: 2026-08-29 updated: 2026-08-30 freshness: fresh

Miscellaneous Research

Summary

The drawer for live research instruments that do not belong in security, OSCAL, agents, or local inference. Two repositories: videocortex (Apple Silicon) and videocortex-spark (NVIDIA DGX Spark / GB10) — the same TRIBE v2 encoding instrument on two machines. Drop a clip, see which cortical regions the model's average subject would fire. This is not a holding pen for unfinished tools; it is the place for work that is real and public and just not in the rest of the taxonomy.

Key claims

• videocortex is the encoding direction: stimulus in, predicted fMRI out, for TRIBE v2's average subject. It is not a decoder and it is not a mind-reader. ^src:videocortex
• The Mac repo is the local instrument around an upstream model: Metal instead of a silent CPU fallback, a doctor preflight that fails in five seconds instead of twenty gigabytes in, and plates / PIP overlays you can put in a slide. ^src:videocortex
• Upstream device="auto" never picks Metal. videocortex resolves CUDA → MPS → CPU, rewrites all four hard-coded device: cuda fields in the published checkpoint, and patches WhisperX float16-on-CPU plus Llama 3.2 GQA so Metal does not LLVM-abort. ^src:videocortex
• The renderer is a separate install from the model. pip install -e . gives working draw / synthetic samples with nilearn only; .[predict] is the torch + tribev2 half. ^src:videocortex
• Colour limits are computed once over the whole run. Per-frame restretch is treated as a lie — a quiet TR would render as vividly as a startling one — and a test fails if it creeps back in. ^src:videocortex
• The wrapper is MIT. TRIBE v2 weights are CC-BY-NC-4.0; meta-llama/Llama-3.2-3B is gated. ^src:videocortex
• The command deck is stdlib HTTP on loopback :8730 — Host pin, non-loopback refuse — matching Cleanroom / Skillspector / STRIDE-Lite. Overlay from the deck does not touch the model. ^src:videocortex
• videocortex-spark is the GB10 sibling, not a silent CPU fallback and not a Metal tree with the serial numbers filed off. CUDA 13 / aarch64; doctor refuses a CUDA-12 wheel that looks empty; UMA headroom comes from /proc/meminfo, never nvidia-smi. ^src:videocortex-spark
• The Spark deck had a circular loadRun → show("runs") → loadRun re-entry: visiting the runs view stacked nested invocations, rebuilt the detail pane, and left the player flickering at 0:00. show() is now the only caller. ^src:videocortex-spark
• Spark adds two verbs the Mac stub does not yet document: export (one-file WebGL brain.html, run-wide scale) and sonify (|predicted BOLD| as loudness, shared scale, Destrieux stand-in voices). Same honesty contract. ^src:videocortex-spark
• Encoding is not attention: visual cortex still fires for an unexpected walker in a counting-task clip. The Spark tree ships the recipe (examples/encode-is-not-attention/); do not vendor someone else's gorilla footage. ^src:videocortex-spark

Open questions

• Do export and sonify land back on the Mac tree, or do they stay Spark-only?
• ScreenLens is video-in / captions-out. Is there a useful shared "video as stimulus" seam, or are they just both videos?
• The loopback command-deck pattern now appears in five public repos. Does it become a shared primitive, or stay copy-pasted?

Related

• local-inference-mlx (ScreenLens is the org's other video-in tool; videocortex is not MLX)
• local-inference-dgx (screen-lens-dgx is the Spark video-in counterpart; videocortex-spark is cortex, not captions)
• developer-tools (Cleanroom's loopback deck is the same local-only posture)
• ai-security-pentesting (STRIDE-Lite / Skillspector share the loopback GUI habit, not the science)

Articles

None yet that are about this work.

Sources