01 · The sentence
Every week a machine reads about three hundred candidate stories and sorts them. Then two or three people read the top of that list, vote, argue and choose about twenty. Nothing enters an issue without a person having opened the source; nothing is dropped from the top of the list without a logged reason.
The sorting is documented so that it can be argued with. If you think the method is wrong, the rubric is the thing to argue about — write to news@phynetic.ai.
02 · Where candidates come from
Five kinds of source feed one inbox. Each candidate keeps the source it came from, the number of outlets that ran it and the time it was first seen.
feedsarXiv cs.RO, company blogs and trade publications, YouTube channels, GitHub releaseshourly
public APIsarXiv, Hacker News, Reddit and Hugging Face, filtered by a fixed query sethourly
GDELTNews in 65 languages matching a published query sethourly
search passa logged set of web searches, run and archivedSaturdays 08:00 WIB
reader submissions/submit — a URL and an optional note; no account, no replyas they arrive
Duplicates are clustered into one story; the cluster keeps the earliest source as primary. About 300 stories a week reach the inbox.
03 · The rubric
A model scores each story 0 – 5 on eight criteria, reading the fetched text — never the headline alone — and writes one line of rationale that curators see beside the score. The weighted sum is scaled to 0 – 100.
criterionweightwhat a 5 looks like
Relevance to Physical AI
gate — below 3 the story is skipped, whatever else it scores
7
Robots learning, moving or sensing in the physical world — or the data, simulators and tools that make that possible.
6
Changes what a team would do this quarter: a shipped product, a released dataset, a result that moves a benchmark by double digits.
6
A primary source with numbers, method and conditions — a paper with code, a dataset with a card, a video that is continuous and unedited.
5
The first time this has been shown, not an increment on something already covered.
2
A lab or company whose moves the field watches.
2
A single source so far, and we would be the first outlet to carry it.
1
Closes a loop on something we ran — the weights promised in July, the trial announced in May.
1
A continuous, unedited recording of the system doing the thing claimed.
score = Σ (weight × criterion) ÷ 1.5 · weights sum to 30 · maximum 100 · integers on purpose — they are meant to be read, not tuned
04 · Three buckets
TopRead by every curator. Most of an issue comes from here.70 – 100
MaybeRead by at least one curator; enters the issue on a vote.45 – 69
SkipSpot-checked, not read. A story found here later is logged as a ranker miss.below 45, or relevance below 3
05 · The diversity rule
An issue has five sections in a fixed order, each with a target count. A company appears at most twice in an issue and at most once in a section; Phynetic counts as a company. Targets are targets, not limits — an editor can run a section long, and the count turns amber when they do.
Robots & Hardware5
Models & Research5
Data, Sim & Tools4
Industry & Funding4
Quick Links8
Upcomingevents, no target
06 · What the team overrides
The ranker proposes; the team decides. Curators vote keep, maybe or reject with a one-line reason; an editor of record resolves the votes and owns the order. The editor may:
—reorder within a section, and move a story between sections
—bring a Maybe or Skip story into the issue — an override, logged with the reason
—leave a Top story out — also logged, also with a reason
—rewrite the model's draft summary; the edited text is what ships, and a claims check flags any sentence the source doesn't support
A second person, never the author, approves each issue before it is scheduled. How much of the ranker's top 20 the team keeps, and how many items come from outside Top, are recorded for every issue.
07 · What we don't do
—No paid placement. A sponsor's item is scored, edited and placed like every other item, and carries the label SPONSOR.
—Phynetic's own items carry the label FROM PHYNETIC, appear at most once an issue, and are held to the same rubric.
—No open tracking. We count clicks on links that pass through news.phynetic.ai and nothing else — no pixels, no third-party scripts, no third-party fonts, on any page or in any email.
—No invented claims. Every summary must be supported by the fetched source text; when the source is an abstract only, the summary says so.
08 · Changelog
v1September 2026First rubric: eight criteria, scored 0 – 5 by a model against the fetched text; relevance is a gate at 3.