Weekday scrape

How a human or a local Hermes loop refreshes the living YAML without inventing benches.

Updated Oct 12, 2026

The home card is data, not a layout. Weekday refresh means editing three files and, sometimes, adding a dated ROCm note. Do not redesign pages.

Files that move

File	Feeds	Touch when
`src/data/what-works.yaml`	Home “what works this week”	Any pin, headline, or honesty line
`src/data/qwen-recommendations.yaml`	Software / start / serve-Qwen picks	Daily driver vs fast agent vs optional
`src/data/stacks.yaml`	`/stacks` cards	Command-line TODOs, measured tok/s
`src/data/gpus.yaml`	GPU directory	Matrix or VRAM listing changes
`src/data/abliterated.yaml`	Abliterated table and method/variant cards	New notable local variant, a dead repo, or a cited card change

as_of is a calendar day. Bump it when you actually re-read sources. Leave it alone if nothing material changed.

Add src/content/updates/YYYY-MM-DD-….mdx when a line changes (new Core SDK, new vLLM extras index). Label it a snapshot. Historical notes stay archived. The YAML wins if a note disagrees.

Sources to open (every weekday)

ROCm 10.0 release notes and the live GPU matrix
AMD vLLM on ROCm — supported combo vs community extras
vLLM releases and the ROCm extra-index / -rocm image tags
Official Qwen cards the weekly YAML actually names (Qwen3.6-35B-A3B and Qwen 3.8 Flash this pass)
Optional: publisher cards for abliterated rows you already list — educational, not the default

If you do not find a newer stable ROCm vLLM or Core SDK, keep the last as_of. Do not bump the date for theater.

Edit rules

Reliability first. Official instruct stays the default until a measured, dated reason says otherwise.
Headline is one sentence: what stayed true, what is news.
No invented benches, prices, or download stats. Publisher file sizes are fine if you say they are publisher figures.
Unknowns stay unknown. “Unstated on the card” is a valid cell.
Companion profiles stay aligned with the YAML. qwen-reliable = daily driver; qwen-fast-agent = Flash. See Radeon Stack Lab.
Never invent measured VRAM, RAM, or flags. Leave  until the box owner fills them.
Copy that says “this week” on FAQ / start must match the YAML, or it must point at the home card.

Optional: Hermes as the overnight research loop

Local Hermes Agent can be the overnight pass that reads those sources and drafts a YAML diff. It is not a bot that publishes.

Serve official Qwen so http://127.0.0.1:8080/v1 or :8000/v1 already works.
Attach Hermes to that URL (do not let it download a random CUDA llama.cpp).
Give it the source list above plus “propose edits to src/data/*.yaml only; no benches.”
In the morning: read the diff, smoke-test on the card if a pin moved, then commit on main.

A golden prompt that finishes on GPU still beats a confident paragraph.

After you edit

npm run build

Commit the YAML (and the update note, if any) on main. Preview the home card before you walk away.