What's new: measured on the rig, a dataset doctrine, and a daily publish

readouts is a set of knowledge bases built in waves of sourced research, each entry checked by a separate verifier before it counts as verified. This newsletter lists what each publication adds: one short issue whenever something new lands. The knowledge-base guide describes each one.

New in the knowledge bases

tensor-engine-knowledge: measured on the rig (wave 18)

The first wave promoted from the studio’s public research bench, rnd. Each row is a measurement with a committed receipt, not a claim about someone else’s page. Twelve rows in three lanes:

How much context each model holds on an RTX 5090 (Ollama 0.35.1, f16 KV cache with flash attention, Q4_K_M, under a 30,500 MiB ceiling, with no spill into system memory):

Model	Largest window on the card	Trained window
llama3.1:8b	131,072	131,072
gemma4:31b	86,016	262,144
granite4.1:30b	43,008	131,072
qwen3:14b	40,960	40,960
qwen3:8b	40,960	40,960
mistral-small:24b	32,768	32,768

Three devices, no contention. The 5090’s decode rate changes by under 1.4% when the Intel NPU or integrated GPU works beside it, and theirs by a few percent at most. All four pairs fall inside a pre-registered ±10% band, so a router can treat them as independent lanes.

Two Ollama recipes that cost real time on the rig:

• num_predict of 0, null or missing means no completion limit, not “count the prompt only”. To count a prompt’s tokens, ask for one token.
• Counting a long prompt’s tokens: estimate from a sample, then bisect only the tail. Adding one sentence per count once took over 30 minutes for a 118k-token target.

And the foundations:

• the rig’s CUDA stack, including a change in how nvidia-smi prints its CUDA version that breaks older parsers;
• what to check on a rented GPU host before paying for it: the driver, and its download speed.

Wave 18’s record · its verification

training-knowledge: a dataset production doctrine (wave 17)

A study of how to produce datasets that can be trusted and reused:

• Formats: JSONL for code and text, and WebDataset or Parquet for images and other modalities at scale, with a translation layer between them.
• Metadata: Croissant as the backbone schema, extended per modality.
• Five gates before a dataset ships: a contamination check, a licence scan, at least one pressure-test run, a diversity floor against synthetic collapse, and a toxicity and personal-data audit.

The readout

vocology-knowledge: a local voice for the editor

Scripts that read selected text aloud through Kokoro-82M, an open-weight neural text-to-speech model that runs on your own machine with no API key. A built-in Windows voice is the fallback, and one shortcut stops it. The scripts

How readouts publishes now

• Daily, when something has changed. The knowledge bases are built in a private working repository and published here through an allow-list.
• The same checks every time:
  • an identity and home-path scan;
  • dead-link and forbidden-term gates;
  • the repository’s own verify.py;
  • a repository map that CI checks on every push.
• A publish that fails any check stops before anything is pushed.