What happened
Graze published an operational description of midterms.graze.social, its planned US midterm-election experience for November 3, 2026. The project runs on Bluesky and the underlying atproto network. Graze presents it as an alternative to engagement-ranked social feeds for a predictable but fast-moving news event: a collection of feeds whose inclusion logic, sources, and editorial rules can be examined by readers.
The system has 54 feeds: one for every state and Washington, DC; a national feed; a national desk drawing on 218 vetted sources; and a DC desk drawing on 25 local sources. Its open feeds are named for people on ballots rather than hashtags. Graze says the project covers 1,392 candidates across 508 Senate, House, and governor races. A reader can sort a feed by new or trending, and can open a feed definition to see how it selects posts. Graze says readers who disagree with a definition can remix it.
The editorial model combines prepared structure with live moderation. Graze is asking Bluesky users to take a state desk, suggest useful posts, flag poor material, and, for people with prior curation experience, help edit feed rules and pin results. Volunteers can also host or join watch parties. People with co/core accounts can lend machines to help produce hourly summaries. Graze says it will send volunteers directions before election night and is prioritizing people who have built Graze feeds or worked in news, while aiming to involve everyone at some level.
Automation is present but bounded. Each race receives a daily summary marked “machine-written · can be wrong.” Every theme in a summary links to the underlying posts. Graze describes the summary as an entry point into a feed, not a replacement for the feed itself. The source frames this design as an answer to the weaknesses of hashtags and opaque ranking systems: names match ordinary writing more reliably than tags, while public feed definitions and active editors make the selection process visible.
Graze also cites a prior election feed built with WNYC and Gothamist for the New York mayoral race. According to Graze, that feed delivered 16.7 million posts to 114,000 readers and reached 1,200 posts per second at peak. The new project is a larger-scale test of the same premise: that a newsroom can own an editorial distribution surface on an open social network instead of relying on a platform’s ranking choices.
What the archive adds
This is a new set of operational facts about a known project, rather than a restatement. Our earlier Field Note, Graze's election night dashboard demonstrates the power of open algorithms, covered the dashboard’s newsroom-facing use of open social data. The new source supplies the scale, desk structure, volunteer workflow, and the design of its machine-written summaries.
Trend signal
This supports Model answers become task interfaces. Graze does not present a generated summary as a final answer; it constrains it to a race-level orientation layer, labels its fallibility, and attaches each theme to the posts behind it. That is a bounded task interface with an evidence path, rather than a standalone text-generation surface.
Why now
Election night has a fixed date, while moderation, source selection, and feed definitions need preparation before the volume and volatility of live results arrive. Graze published the operational plan on October 14, with November 3 identified as election night, and is recruiting volunteers in advance.
Why it matters
For builders, the useful lesson is not that a model can summarize political conversation. It is that a high-stakes information product can assign the model a narrow role and retain inspectability around it. Graze exposes the feed rules, distinguishes human curation from automated summaries, flags that summaries can be wrong, and routes users to the underlying material.
That arrangement does not eliminate judgment: source vetting, feed definitions, moderation, and pinned results remain editorial work. But it creates separable components that can be reviewed and revised. Teams building AI-assisted dashboards, support systems, or internal intelligence tools can apply the same division: make the model’s output a navigational layer, show its basis, and keep ownership of the rules that decide what enters the system.
What happens next
- Graze will fill more state-desk volunteer roles before election night. (3 months, 70%). Watch for: changes to the public state map or additional volunteer instructions.
- The project will test whether its machine-written summaries remain useful under live election volume. (3 months, 75%). Watch for: published summaries with linked source posts and post-election reliability reporting.
Limits and uncertainty
The source is Graze’s own description of a planned project, not an independent evaluation of its accuracy, moderation quality, or resilience under election-night traffic. The cited WNYC and Gothamist figures are attributed to Graze. The material does not establish how the machine-written summaries are generated, what model is used, how hourly summaries are reviewed, how volunteers are vetted, or how disputes over feed definitions are resolved. Public definitions make selection inspectable, but the source does not demonstrate that readers will inspect them or that openness prevents misinformation from entering a feed.
Skill to build now
Choose one generated summary in an existing product and turn it into a bounded evidence interface this week. Define its permitted scope, add a visible uncertainty label where appropriate, and require each claim or theme to point to the source item that supports it. Then assign a named reviewer or operating rule for changing the inputs. The practical goal is not merely better prose; it is a summary that helps a user inspect, verify, and correct the system.
Reader takeaway
Graze’s election desk is an experiment in making automated synthesis subordinate to a visible editorial system. Its important design decision is the evidence trail: machine-written race summaries are labelled as fallible and linked back to the posts that produced their themes. For AI product teams, that is a more durable pattern than treating generated text as the endpoint.
Open question
Can a volunteer-run, publicly remixable feed system maintain enough editorial consistency during a national election night while still responding quickly to unexpected events and misinformation?