Radar explore
AI Research Radar — Evidence-Grounded Trend Ledger for AI Systems
npx -y skills add Neetx/ai-research-radar --skill radar-exploreAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
2 things to look at
- no licenseNo license file was found in the repository. Code published without one is not open source by default, so using it at work is a question for whoever answers licensing questions where you are.
- 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
What its author says it does
Copied from the file, not written here
Discover important work the radar does NOT yet track, by iterating the discovery-venue list in SOURCES.md every run — significance-first: read the top / most-attention items in each venue REGARDLESS of topic, not by searching tracked axes. Surfaces off-axis significance and candidate new axes. Use for the daily exploration slot, after the sweeps.
SKILL.md
3.4 KB, as published. Nobody here has run it
Explore — find what we SHOULD be tracking, not only confirm what we do
The sweeps (radar-lab-sweep, radar-pulse, radar-repo-watch) CHECK known
sources for new items on the axes we already track. This skill does the opposite
job: it finds work we do NOT yet track but that is important — the candidate next
axis, the off-axis result the field is about to care about.
It is list-driven exactly like the lab sweep: the venues live in SOURCES.md and
you iterate ALL of them. You do not "explore something" by vibe — you execute a
list of WHERE new work surfaces. WHAT to look for is never enumerated here (that
is the trap that makes the search narrow); WHERE to look is the contract.
Method (significance over topic-match)
- ITERATE the venues in
SOURCES.md→ "Discovery / exploration venues" (see the list contract there: cross-category attention venues every run; one rotating arXiv category as supplement). For each venue, read its TOP / most-attention items — HF daily papers by upvotes, an arXiv category's newest listing, a trending feed's lead items — REGARDLESS of whether they match a tracked axis. - ADVANCE THE WINDOW: read items NEWER than the last time you covered THAT venue
(its previous date in
logs/source_rotation.md); record the range. A venue that returns "same batch already mined" did not advance — move the window forward or take the next venue in rotation. Never re-mine the same listing. - Assess each top item on its OWN merits, NOT the trend bar: does it claim to solve a known hard problem, is it highly ranked / from a serious venue, are multiple groups already on it? Significance is topic-agnostic — a result OFF every current axis can still be the most important thing you see today.
- ROUTE every find per
routines/daily.md§4 andradar-ledger-update: on a tracked axis → evidence; a queue cluster of ≥3 independent groups → a seed trend; else →observation_queue(a genuinely significant OFF-axis single artifact still gets queued, marked "significant, off-axis"). A named primary is NEVER left in report prose only. - Recurring off-axis significance is a candidate NEW AXIS — propose it at the weekly. The radar's job is to discover what it SHOULD track, not only confirm what it already does.
Venue list
Lives in SOURCES.md → "Discovery / exploration venues" (agent-owned: ADD a
venue when it repeatedly surfaces significant work, drop dead ones — a new venue
goes INTO that list so it is iterated next run, never just into this prose).
Cross-category, attention-ranked venues are PRIMARY every run — they show a theme
forming across categories that a single arXiv category cannot. Single arXiv
categories are a supplementary rotation with an advancing window.
Log
Record in logs/source_rotation.md which venues you covered and the date range
each run, even when nothing cleared assessment ("explore: <venues>, nothing new").
A venue you skipped is a discovery gap, not a lean pass.