The research bench. Readouts is the shelf.
Findings from any field land here fast, with every source tagged by tier and every claim marked as checked or not. Rig experiments sit beside the entries they test. When a topic holds up, it moves on to readouts, the verified knowledge bases.
Search
python -m rnd search cuda graphs
Both stores
python -m rnd readouts splice glitch --any
File
python -m rnd new "Paper title" --kind paper --field audio
What it holds
Messy on purpose, and honest about it.
Tiered sources
Every source is primary, secondary, aggregator, user or rig. AI search output is never passed off as primary.
Claims carry their checks
A claim is unverified until something checks it, and a verified claim names what did. Disputed and wrong claims stay visible.
Rig experiments
Measurements made on our own machines keep their harness, pinned inputs and result receipts under experiments/.
One search, two stores
rnd searches the bench, and rnd readouts searches every verified knowledge base, read-only, from the same seat.
Instrument registry
The studio tools research can call on (study swarms, evidence packs, local models, rented GPUs) with how to invoke each.
Rated catalogues
External catalogues such as NVIDIA’s 398 agent skills, pinned to a commit and rated for studio fit.
From bench to shelf
Knowledge moves one way.
| Stage | Where | What happens |
|---|---|---|
| Bench | rnd (this repo) | A finding lands as an entry; claims start unverified; rig measurements get a harness. |
| Internal shelf | readouts, private working repo | A topic whose load-bearing claims hold up is built into a knowledge base, or added to one. |
| Public shelf | readouts, public repo | A knowledge base is published when it joins the export allow-list. |
Use it
Search before you research
python -m rnd search cuda graphs
python -m rnd readouts listener --any
python -m rnd toolsFile and validate
python -m rnd new "Title" --kind finding --field audio
python -m rnd check # exit 1 on any errorAsk the index
python -m rnd list --relevance act
python -m rnd stats --json
python -m rnd sql "SELECT tier, count(*) FROM sources GROUP BY tier"Verify the repo
bash verify.sh # tests, check, build, smoke