Filing research

One entry per topic. Scaffold it, fill the sections in your own words, then validate.

python -m rnd new "Singing pitch trackers" --kind finding --field audio --tag pitch
# edit entries/2026/2026-10-07-singing-pitch-trackers.md
python -m rnd check

The file

---
id: 2026-10-07-cuda-graphs        # defaults to the file name
title: CUDA Graphs
date: 2026-10-07
kind: concept
relevance: reference              # act | watch | reference
status: active                    # active | draft | superseded (optional)
fields: [gpu-computing]           # at least one; 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

Frontmatter is a small YAML subset: scalars, inline [a, b] lists, block lists, lists of maps and block scalars.

Kinds

kind	for
`finding`	a fact or result learned from a source
`concept`	an idea, technique or principle worth knowing
`release`	a product, toolkit or library release
`paper`	a paper or formal study
`tool`	an external tool, library or catalogue item
`catalog`	a pointer to an ingested catalogue
`rig-fact`	a measured fact about our own machines
`event`	a talk, webinar or conference
`question`	an open question being tracked
`decision`	a decision taken on the strength of research
`instrument`	a studio tool the seat can use (lives in `instruments/`)

Two questions, kept apart

What is it? Summary and Key points describe the knowledge as it stands, with no studio framing.

What does it mean for us? Studio relevance says what to do with it, and the relevance field sums that up:

act: something should change because of this.
watch: it may matter; revisit when it moves.
reference: worth knowing; nothing to do.

Source tiers

Every source line starts with its tier.

tier	means	needs a URL
`primary`	vendor docs, papers, repos, model cards	yes
`secondary`	reputable write-ups of primary material	yes
`aggregator`	summary sites, AI search output	yes
`user`	supplied by a person: slides, notes, pasted text	no
`rig`	measured on our own machines	no

Pasted AI-search output is aggregator or user, never primary. If a claim matters, find its primary source.

Claim confidence

tag	means
`[unverified]`	not checked yet (the default)
`[verified]`	checked; must end with `(via: …)` naming what checked it and when
`[disputed]`	sources conflict
`[wrong]`	checked and false; must say what showed it, with `(via: …)`

Verification is proportional. Check a claim when it looks wrong, when sources conflict, or when a decision depends on it. Do not run a verifier over every entry by default.

Interim results are fine. Say they are interim in the title and the claim, and update the entry when the rest lands.

Links

[[entry-id]] links one entry to another. rnd show lists each entry’s backlinks, and rnd check warns about a link that points at no entry.

What `rnd check` enforces

Errors (exit 1): unreadable frontmatter, a bad id, a missing title, a date not in YYYY-MM-DD, an unknown kind, relevance or status, no fields, an unknown source tier, a URL-less primary/secondary/aggregator source, an unknown claim tag, a verified/wrong claim without (via: …), a duplicate id, and an instrument missing instrument_status or invoke.

Warnings: an entry with no sources (except rig-facts and instruments), and a [[link]] to no entry.

Writing rules

Your own words. Never paste copyrighted text into an entry.
No home-directory paths, secrets or anything that is not ours to publish. The repo is public.