Text model
Implementation
What happens
Design choices
Failure if wrong
One controlled experiment changes one feed ingredient, compares it with the control, and preserves what was learned.
Decentralized execution, centralized attention.
You control it. A feed generator can run on your server and return its own ordered list of post URIs.
Portable by protocol. The feed record points through a DID to your service endpoint.
AppView dependency. The client’s backend fetches your skeleton and turns URIs into complete posts.
De facto centralized. Bluesky’s app controls prominent discovery, ranking, defaults, and client presentation.
Can a feed charge users? Not natively. The current feed contract has discovery, likes, and delivery—but no paid subscription or entitlement primitive. Treat direct paid-feed access as unsupported unless a client adds it.
Where can Hyperblues earn? Sell the surrounding service: private team feeds in a compatible client, feed design and operation, analytics, moderation, sponsorship, or feeds that lead to a paid community/product. Keep the public Bluesky feed useful as distribution.
Showing One Per Person by default. Sign in to use your Bluesky pinned-feed order.
Select a row to drive the model above.
| Feed / goal | Source → selection | Ranking | Risk → next experiment |
|---|---|---|---|
| Orbit Personal relevance |
People you recently liked → Recent posts from high-affinity authors |
Affinity × recency, with exploration | Reinforces an existing social bubble → Reserve more slots for weak ties |
| Exploration Discovery |
Follows of your favorite people → Accounts you do not already follow |
Shared connections, then recency | Popularity inside one neighborhood → Reward distinct social neighborhoods |
| Quiet Images Visual serendipity |
Quiet, image-heavy accounts nearby → Recent posts containing images |
Quiet-author signal × recency | Image presence is a weak quality signal → Learn which visual themes earn saves |
| Liked By Reciprocity |
People who recently liked your posts → One recent post per person |
Interaction recency | A like may be incidental, not affinity → Weight repeated reciprocal interaction |
| Peers Peer discovery |
Accounts in and near your graph → Authors with a similar audience size |
Follower-count proximity × recency | Audience size does not imply shared interest → Combine peer size with topic affinity |
| One Per Person Author diversity |
Your following timeline → Newest original post per author |
Chronological | Treats every author and post equally → Tune freshness without losing coverage |
Ten candidate feeds scored against promise, relevance, discovery, coverage, supply, control, feasibility, and resilience.