The research watch¶
A dated sweep of new work relevant to offline voice input, plus the index that stops it repeating itself.
What is here¶
| File | What it is |
|---|---|
YYYY-MM-DD-digest.md | One run's findings, after curation — see below |
seen.json | Every paper ID already reported, so the next run shows only what is new |
How to run it¶
uv run python scripts/research-watch.py --days 30
uv run python scripts/research-watch.py --dry-run # look without writing
It queries the arXiv API over seven topic searches, each tied to a direction this project has written down. Roughly a minute, mostly spent being polite about the rate limit.
The digest is curated, not published raw¶
Every entry lands with **So what:** _(unreviewed)_ and must be either annotated or deleted before the digest is committed. That is the whole discipline. An automated feed nobody prunes is a feed nobody reads, and this repository has enough generated-and-ignored material already.
Curating the first run removed two entries the queries had matched incidentally — a voice-controlled game editor and a paper on LLM code security that used the phrase "speech input" in passing. Both queries were then narrowed, with the reason recorded in the script. That loop — run, prune, tighten — is the feature. The tool is worth keeping only for as long as someone does it.
What it will not do¶
- No PDFs. The repository blocks committed PDFs and the pre-commit hook enforces it. We publish a summary and a link to the publisher's copy; redistributing other people's papers is a copyright problem, not a convenience question.
- It is not in the daemon. ADR-019 enumerates every way data can leave the machine and fails the build on a new outbound call inside
src/yazses/. A literature watcher shipped in the product would be a sixth network path on a tool whose central promise is that it makes none you did not ask for. So it is maintainer tooling, run by a person who chose to. - It does not rank or score. Relevance is a judgement, and the "So what" note is where it goes.
Where the results end up¶
A digest that produces something durable should feed it forward rather than sit here:
- A direction worth pursuing → the direction page, through the same three gates as any other idea.
- A method for an open question → the research agenda.
- A term worth defining → the glossary.
- Nothing at all is a legitimate outcome, and the commonest one.