cat architecture.md
MINT runs a local-first research infrastructure on a Mac Studio. FeedMe reads hundreds of sources three times a day and ranks what matters to the lab. Yesterday in AI is written from its shortlist, with Seth choosing and approving every issue. A searchable paper corpus, with its own paper-fetching and review services, sits behind the lab's Slack agents and a read-only server for collaborators' assistants. Shared services handle credentials, model access, backups and the public route in. Public pages are static files deployed through GitHub Pages.
trace content-pipeline
FeedMe is the shared collection and ranking layer. The full 04:00 run starts all seven fetchers at once, drops repeats and stale items, scores every new item against the lab's interest profile, summarizes long pieces, and groups the survivors into stories on the day's slate. The 12:00 and 20:00 runs are top-ups with six fetchers that merge new stories into the same slate. Items count as seen only once the slate accepts a run, so a failed run is fetched again. A watchdog checks the morning run thirteen times before 08:00 and restarts it if it stalls.
- X
- 5 Bluesky feeds
- ~380 feeds, journals, podcasts
- Newsletter inbox
- arXiv + PhilPapers
- 04:00 full run
- 12:00 top-up
- 20:00 top-up
- Day's slate
- up to 300 sources
- Morning picks
- 20 to 25 to #firehose
- SethMarks the slatein the curator app
- Selector19:30 to 23:30, about 40 sources, continuity check
- SethApproves the pitchproposal at 23:00
- Writer + deskdraft, Read mores, edits, preview
- Releasewebsite, one email per subscriber, #mint-community
- Picks + sharesmorning picks, and the share bot every 30 s
- #firehosethe lab's shared stream
- Paper botevery 5 min, fetches papers posted there
- #paperspapers cross-posted; requests filed for any it can't get
- Intake#papers every 10 min, folder overnight, Drive every 30 min
- Fetchuniversity library and a local model
- Analyze41 questions and six summaries per paper
- Single writerone locked writer; archive to Drive and Zotero
- Serve@Minty, the corpus server, and the Corpus Map (Sundays)
Source adapters
Products and agents
A sign-in keeper holds the browsers that read paid sites, the university library and X, tests each login at 09:10 and 17:10, and emails Seth when one needs renewing. Each source is judged against its own recent normal, and Seth is told once when an important source fails and once when it recovers.
trace corpus-ingestion
Paper ingestion is a queue-backed set of focused services. Papers
arrive from the #papers channel, a local intake folder
and a shared Google Drive inbox; papers that an ordinary download
cannot get are found by a separate fetching service working through
the university library. Each paper is fingerprinted so repeats are
set aside before any costly work, then identified, converted,
embedded and analysed. One writer owns every database change, and
the archive keeps the original PDF on Drive and an entry in the
lab's Zotero library.
-
corpus-intake-localWork the local intake folder overnight, every 5 minutes. -
corpus-intake-slackTake papers from Slack from 08:00 to midnight, every 10 minutes. -
corpus-ingest-handoffHand unfinished work over at midnight and top up the overnight queue. -
corpus-analyzeAnswer 41 research questions and write six summaries per paper. -
corpus-db-writerCheck each handoff against its archived files, write it, and read it back. -
corpus-tidyClear intake copies once the archived copy is confirmed.
du -sh corpus/
The corpus contains both PDF-backed papers and sources captured
directly from the web as Markdown. Each document is represented at
several levels: bibliographic metadata, full text, searchable chunks,
embeddings, structured research-question analysis, and hierarchical
summaries. Every Sunday the corpus is regrouped into topic clusters,
and each paper keeps a list of its ten closest neighbours. The
current document embedding policy is
Voyage voyage-context-3, 2048d.
Explore the public projection in the Corpus Map. The public map exposes research structure and links, while the working corpus remains on lab infrastructure.
query @Minty
Tag @Minty in a lab channel with a question and it
offers buttons for a mode and a model. A model rewrites the question
several ways, the agent pulls matching corpus papers, and the answer
cites only that evidence, with real links added by the program. Long
answers and deep reviews arrive as a PDF. Tag it on a paper or PDF
and it adds the paper to the corpus instead.
| Mode | Best use | Searches |
|---|---|---|
Fast Search |
Quick factual or bibliographic lookup | Paper corpus |
Fast Review |
A short review of a focused question | Paper corpus |
Deep Review |
Substantive literature synthesis, delivered as a PDF | Paper corpus |
Conversation |
Follow-up discussion in the thread | Paper corpus and thread context |
AI governance |
Questions about AI laws, bills and policies | Governance dataset, refreshed daily |
Teaching |
Questions about training models from human feedback | Open textbook and course videos |
gpt-6-astra, with
high
reasoning for intelligence-heavy work.
cat AGENTS.md
Minty is the lab's research-management agent. The root
AGENTS.md is the always-on contract for identity,
operating boundaries, verification, and collaboration. Each
persistent workstream keeps one THREAD.md with its rules,
sources, decisions and blockers, so a new agent can pick it up.
The old root fact digest and session database are retired archives. Durable lessons belong in the active runtime's memory, and exact history stays in that runtime's native transcript store.
launchctl list | sort
launchd owns persistent and scheduled services. The table below is generated from Minty's canonical daemon registry and reconciled with the jobs currently loaded on the host. 54 user jobs and 2 system jobs are loaded.
| Service | Purpose | Schedule | Observed state |
|---|---|---|---|
loading snapshot |
Reading the public infrastructure inventory. | - | • Loading |
calendar daily-jobs
| Time (ET) | Service | Work |
|---|---|---|
| 04:00 | feedme-daily |
Full source collection, scoring and the day's slate, then the Slack picks |
| 04:42 to 08:00 | feedme-daily-watchdog |
Thirteen checks on the morning run; restart it if it stalls |
| 05:15 | minty-chatbot-sync |
Rebuild the site assistant's records when they change |
| 09:10 and 17:10 | yinai-health-check |
Test the paid-site, library and X sign-ins |
| 12:00 and 20:00 | feedme-incremental |
Top-up runs that add new stories to the day's slate |
| 19:30 to 23:30 | yinai-selector |
Choose about forty sources for the next issue |
| 23:00 | yesterday-in-ai |
Plan the issue for Seth's approval |
| After approval | yesterday-in-ai-write-consumeryinai-proposal-release |
Write, edit and preview the issue; publish and send once approved |
| 00:00, 06:00, 12:00, 18:00 | rclone-backup |
Back up lab data; the 06:00 run archives the corpus and mirrors off-site |
| 06:00 and 18:00 | guide-updater |
Refresh and publish this page's system snapshot |
| Sunday 04:00 | paper-map-updater |
Regroup the corpus if due and redraw the Corpus Map |
| Sunday 05:00 | slack-index |
Refresh the Slack search index |
| Continuous | corpus-* |
Queue-backed paper intake, analysis, and writes |
| Persistent | corpus-agent |
Answer grounded corpus questions and add tagged papers in Slack |
| Every 30 seconds | share-to-minty |
Score and post links shared in Slack |
| Every 5 minutes | news-pdf-bot |
Fetch papers linked in #firehose |
Exact launch intervals and loaded state are listed above. The schedule table highlights the main lab-facing rhythm rather than duplicating every watchdog and maintenance interval.
help lab-members
Bots
Slack channels
| Channel | Role |
|---|---|
#mint-community |
Lab discussion, shared links, and each Yesterday in AI issue link. |
#firehose |
Morning news picks and shared links; papers posted here are fetched. |
#papers |
Paper intake for the corpus, with ingestion replies. |
inspect agent-runtime
Scheduled agents send model calls through a shared router that rotates through the lab's model sign-ins and rests any that hit a usage limit; a probe checks every sign-in every ten minutes without spending a call. Jobs get credentials by name from a signed helper that keeps them out of code and logs. Deterministic services use ordinary code where a model adds no value; research synthesis and judgment-heavy work use higher reasoning effort.
| Task class | Reasoning effort | Typical work |
|---|---|---|
| Intelligence |
high
|
Research, synthesis, review, and complex judgment |
| Functional |
low
|
Implementation and bounded operational tasks |
| Deterministic |
low
|
Mechanical checks where a model is still required |
ls integrations/
Yesterday in AI
A daily briefing on AI safety, governance, ethics, and capabilities, chosen and approved by Seth, delivered to 188 enabled readers.