mcptoolshop-rnd 1.1.10.0.0

English | 日本語 | 中文 | Español | Français | हिन्दी | Italiano | Português (BR)

The studio's research bench. Findings from any field come in fast as short Markdown entries, with every source tagged and every claim marked as checked or not. Experiments sit next to the entries they test. External catalogues are mirrored and rated, and a registry lists the studio tools research can call on. One command searches all of it.

Where it sits: the bench before readouts

Two stores hold the studio's knowledge, and they do different jobs.

Knowledge moves one way:

1. Bench. A finding lands here as an entry. Its claims start [unverified], and are marked [verified] only with a note of what checked them. Measurements made on our own machines go in as rig sources, with their harness under experiments/.
2. Internal shelf. Once a topic's load-bearing claims hold up, it is built into a knowledge base in readouts' private working repo, or added to one. Promotion is part of closing an experiment (the Director, 2026-10-09). When an experiment's result is committed here, its readouts-internal update follows in the same session or the next one, with its charts. Nothing sits on the bench for more than about a week. Regular small updates beat one big one. An experiment's README records its promotion as a line under its result: "Promoted: readouts-internal <kb> wave <n>, <date>", or why not yet.
3. Public shelf. A knowledge base is published to the public readouts repo when it is added to the export allow-list.

rnd readouts searches every readouts knowledge base from here, so one seat reaches both stores. Search the bench first, the shelf second, then research the gap.

Research here is not limited to the studio's own fields. Each entry records two things separately: what the knowledge is, and what it means for the studio (relevance: act | watch | reference). reference is a fine verdict.

Use it

Requires Python 3.10 or later and nothing else (standard library only). Runs on Windows, macOS and Linux. From the repo root:

python -m rnd search cuda graphs            # full-text over entries + catalogues (stemmed: pair finds pairs)
python -m rnd search --exact PAIR           # whole words, same case: the product, not the word
python -m rnd probe https://github.com/NVIDIA/NeMo-Relay   # licence, release, telemetry hits, already in rnd?
python -m rnd show 2026-10-07-cuda-graphs   # one entry, with backlinks
python -m rnd list --relevance act          # what needs doing
python -m rnd tools                         # instruments this seat can use
python -m rnd readouts splice glitch --any  # search the readouts knowledge bases too
python -m rnd catalog lanes                 # NVIDIA skills by lane, with studio fit
python -m rnd catalog list --fit adjacent   # skills worth using when the need arises
python -m rnd new "Paper title" --kind paper --field audio --tag pitch
python -m rnd check                         # validate every file (exit 1 on errors)
python -m rnd sql "SELECT tier, count(*) FROM sources GROUP BY tier"
python -m rnd bump --note "what changed"  # micro version bump + CHANGELOG section

Every listing command takes --json for agents. rnd.cmd (Windows) and rnd.sh (POSIX shells) are thin wrappers, so the command works from any directory.

To use the tool from other projects, install it from PyPI:

pip install mcptoolshop-rnd

That installs the rnd command, not the library: the entries live in this repository. The command finds a library in this order: --library DIR, then $RND_ROOT, then the nearest folder at or above the current one that holds entries/ and instruments/. Outside a library it stops with NO_LIBRARY; only rnd readouts works without one.

In Python, import mcptoolshop_rnd: that name always works. rnd is kept as a compatibility alias, but an unrelated PyPI package is also called rnd, and if both are installed it can replace the alias. The rnd command and python -m mcptoolshop_rnd keep working either way.

The library changes daily, so versions have five segments, MAJOR.MINOR.PATCH.MICRO.NANO. The first three version the rnd tool; MICRO marks a structural library change and NANO an ordinary update. rnd bump raises the last segment by default and writes a CHANGELOG section from the files changed since the last tag, so each update gets its own small tagged version.

Exit codes: 0 ok · 1 invalid library files · 2 usage error or not found · 3 runtime failure (an external tool, or an unexpected error). Errors print a code, a message and a hint; --debug adds the traceback.

Layout

Entry format

---
id: 2026-10-07-cuda-graphs        # defaults to the file name
title: CUDA Graphs
date: 2026-10-07
kind: concept                     # finding concept release paper tool catalog rig-fact event question decision instrument
relevance: reference              # act | watch | reference
fields: [gpu-computing]           # any research field, open vocabulary
tags: [cuda-graphs, pytorch]
---

## Summary
## Key points
## Studio relevance
## Claims
- [unverified] A checkable statement.
- [verified] A checked statement. (via: what checked it, date)
## Sources
- [primary] https://… — publisher

• Source tiers: primary (vendor docs, papers, repos), secondary (reputable write-ups), aggregator (summary sites, AI search output), user (supplied by a person: slides, notes), rig (measured on our machines).
• Claim confidence: unverified, verified, disputed, wrong. A verified or wrong claim must say what checked it with (via: …).
• [[entry-id]] links entries; rnd show lists backlinks.

Data packs

Processed data too big or too bulky for the clone is offered as a data pack: the bytes live on a Hugging Face dataset repo (private until published), and the library keeps only a small manifest. The integrity model comes from research-packs: a manifest, a SHA-256 for every file, a one-command verify, and a listing.

rnd datapack build natural-errors --source experiments/natural-errors  # reads datapacks/natural-errors/spec.json
rnd datapack list
hf download mcp-tool-shop/rnd-natural-errors --repo-type dataset --local-dir ne
rnd datapack verify natural-errors --copy ne/files                      # PASS when every byte matches

• Every file needs a licence in the spec. Model outputs carry their model's terms (for example "Built with Llama" for Llama 3.1 outputs). rnd datapack build refuses a pack that leaves a file uncovered.
• Every generating model is pinned (Ollama digest, HF revision or API model id), with its role.
• rnd check validates every manifest and fails on one edited by hand.
• The upload carries datapack.json beside the files, so anyone can verify without this repo: pip install mcptoolshop-rnd.

Collected experiment records

A tool repository commits its own pre-registered specs and run records (interlock's receipts), the way every repository commits its Atlas map. The library publishes only a trimmed copy of the numbers, made by rnd collect from the tool repository at pinned commits:

rnd collect                 # list the experiments that collect records
rnd collect e4b             # read each source with git show, trim it, write experiments/e4b/results/
rnd collect e4b --check     # recompute every output; exit 1 if any differs from the committed file

• experiments/<name>/collect.json names the repository, the sibling clone it's read from (--clone overrides), and each source: its path, the commit, the SHA-256 of its bytes, the output and the trim function in the experiment's extract.py.
• A pinned source never changes. Bytes that no longer hash to the pin are refused.
• Every record carries provenance: the repository, the source path and the SHA-256 of the bytes trimmed.
• Every output is public-safe before it is written. No machine path, home directory or environment field survives, and an experiment's module can add a stricter guard. Nothing is written on any refusal.

Probing a tool before a trial

rnd probe <url> … answers the first questions about a tool or model from each source's own API, not from a summary of a web page:

• GitHub (through gh api, so it uses your GitHub CLI login): the licence, whether it's archived, the last push, the latest release, and a code search for "telemetry". A hit is a lead to read, never proof; zero hits isn't proof either.
• Hugging Face: the licence tag, or for "other" the card's licence name and link; gating; the last change.
• PyPI and npm: the licence, the latest version and when it was published.
• rnd itself: an exact-word search for the name, so you know whether the library already holds it.

A source that can't be reached is reported as unknown, with the reason.

Watching sources

rnd watch checks the sources in docs/watchlist.json and reports what's new since the last check:

• NVIDIA's technical blog;
• the CUDA toolkit version;
• NVIDIA's new Hugging Face models;
• the NVIDIA skills catalogue;
• the releases of the engines and tools the studio runs, such as Ollama, OpenVINO, PyTorch, ComfyUI and Claude Code.

python -m rnd watch                  # what's new; records what it saw
python -m rnd watch --no-save --json # look without recording
python -m rnd watch --only ollama    # one source

What it saw is kept in a state file outside the repository ($RND_WATCH_STATE, else ~/.rnd/watch-state.json). A source's first check is a baseline: it records the current items and reports none as new. A pre-release is never reported as a source's latest version; a newer one is shown beside it. GitHub release notes and commit messages are kept in the JSON output, so a reader doesn't have to fetch each page. A source that can't be reached is reported as unknown, and its state is left alone so nothing is skipped silently.

The command is the deterministic half of a weekly ritual (Director, 2026-10-13). A scout reads the new items, probes any candidate tool with rnd probe, checks the library with rnd search --exact, and writes a short ranked report: what to file, what to try, and what to skip. Add a source by pull request, and say why it matters.

Relation to other studio tools

• readouts is the verified shelf this bench feeds (see above).
• research-os builds a gated, frozen evidence pack for one topic. A topic a decision depends on can graduate into a research-os pack, linked from the entry.
• repo-knowledge indexes the studio's own repos; this library covers knowledge from outside them.
• Findings gathered while designing something belong here too, so they outlive the session that produced them.

See python -m rnd tools for the full instrument registry, and docs/standards.md for how the workflow scores against the studio's workflow standards.

Security and trust

• Data touched: files inside this repo (entries/, instruments/, catalogs/, experiments/) and the index rnd.db, which it rebuilds. rnd readouts opens the readouts knowledge bases read-only. rnd sql runs against a read-only connection.
• Data not touched: nothing outside the repo and the readouts checkout. It stores no credentials and reads none.
• Network: none, except rnd catalog sync, which calls the GitHub API through your own gh CLI login when you run it.
• Permissions: ordinary file access. No elevated rights, no background service.
• No telemetry. Nothing is collected or sent.
• Public repo hygiene: before every push, the tree is scanned for home-directory paths and operator identity.

Report vulnerabilities as described in SECURITY.md.

Tests

bash verify.sh                               # tests, library check, index build, smoke
python -m unittest discover -s tests -t .    # tests only

Status and licence

Maintained by the studio and in daily use. Code: MIT. Entries: CC BY 4.0. The NVIDIA skills mirror in catalogs/nvidia-skills/catalog.json reproduces skill names and descriptions from NVIDIA/skills under that project's licences (Apache-2.0 for code, CC-BY-4.0 for skill text).

Built by MCP Tool Shop