Asking the right questions

Attention is a Web of Trust for attention. The protocol only works if each kind answers one question. Mixing the questions produces likes that pretend to be trust, or trust that pretends to score a post.

32009  identity   →  Is this identity worth my trust?
32014  artifact   →  Is this artifact worth my time?

Answer trust first. Then consume ratings only from people who passed. Without kind 32009, a rating is a crowd score (platform likes, public review sites). With it, the rating is your network’s advice on attention.

flowchart LR
  trustQ["32009: worth my trust?"]
  wot["Walk positive p edges"]
  rateQ["32014: worth my time?"]
  chip["One scroll score"]
  trustQ --> wot --> rateQ --> chip

Kind 32014 is specified in NIP-32014.md. Kind 32009 is specified in NIP-32009.md. Attention ships both kinds: 32009 for trust hops, 32014 for ratings consumed from trusted identities.

Hard vs soft#

	Trust (`32009`)	Rating (`32014`)
Question	Is this identity worth my trust?	Is this artifact worth my time?
Value	`v` = `1` / `0` / `-1` / `""` (Delete)	`score` = `""` (Delete) or `0`–`100`
Graph	Positive `p` is a hop	Never a hop; never grants access
Feel	Hard, controllable	Soft, familiar (stars)

32009 can open a door (c=security:access:…) or expand whose judgments you read. A four-star review must never do either. Publishing both on the same subject is allowed; they mean different things.

Agents vs artifacts#

Trust UI for people and issuers (p, X user:id). These can issue statements.
Rating UI for dead things: posts, products, pages. They cannot issue trust.

A rating MAY target an npub or an X account as a statement. It still does not expand the graph. For people, the default action remains Trust / Distrust. Rating a person is a soft review, not “I will consume their trust.”

One scroll indicator#

While scrolling, show one number: the average of active 32014 scores at the nearest hitting degree — the same stop rule as kind 32009 trust. If you have rated the post, that score is the indicator. If not, average only the closest trusted cohort (people you trust, then the next hop, and so on). Farther ratings do not dilute nearer ones.

High → the network would invest time → read
Low → spam, slop, noise → skip
None → no signal yet → you decide

Who rated, which labels, and any comment belong behind a tap. The timeline is not a review thread.

Product policy (not a protocol MUST): labels such as spam or ai-slop MAY hide an item even when another trusted rater scored it high.

Do not duplicate the host#

Hosts already have social gestures. Attention adds the missing score.

Surface	What to offer	Why
X post (likes + replies exist)	Stars + labels + optional collapsed comment.	Discussion stays in X replies; `content` is the claim note, not a second reply timeline.
Product / page with no review	Stars + optional comment + optional labels	The host has no review. You are the review.

content stays optional on the wire. Empty content is a complete rating. On an X post, the composer is collapsed; inbound comments are shown on tap-through.

A public Like is not a rating: it has no dislike, it is not filtered by your web of trust, and it does not answer “worth my time.”

Labels are shortcuts, not a second protocol#

Optional l tags (insightful, genuine, funny, ai-slop, misleading, spam, …) are the why. They live on the same replaceable rating event. Clients MAY map them onto whole stars (insightful → 100 / 5★, genuine → 80, funny → 60, AI slop → 40, misleading → 20, spam → 0). One author, one subject, one site, one purpose context: one review.