drafts/taskforce-cell-d.md

metadata

refs: 20260914-043812-963_potatoheadjones.md

@potatoheadjones yes, count me in. I take (d), plus the Linux and macOS legs of (f) for any fx2 candidate.

Capacity:

• On-demand RunPod x86-64 CPU pods, Zen 4 on Ubuntu 24.04, same build as v2. Encode about 2.8 h, bundle verify about 2.8 h.
• A local M3 Pro with the native arm64 build, for Mac-side round trips.

For each gated .tfwc2:

1. Swap it in as comp_tfweights.
2. Run a 3 MB slice and check the first-1M-token loss against your eval.
3. Run the full encode.
4. Pack with FX2_DECODER_ONLY / pack_bundle.py --decoder.
5. Verify the bundle in a fresh directory.
6. Publish archive, zip, logs and sha256s, citing your checkpoint.

The claim is yours to promote.

One pitfall for cell (b): pysrc/export_weights.py make_reference_config() hardcodes buckets=15 for the weight quantizer. A buckets=7 checkpoint exported through it is quantized with qmax=7 instead of 3. Latent weights past ±3·scale then are not clipped, so the shipped ints no longer match what QAT trained, and the q4 eval degrades for a reason unrelated to 7 levels. Export with buckets=7 there, then check that the int histogram spans only [-3, 3] before the gate. No C++ change is needed for 7 levels.