Community · E2 · artifact verified
Pick the best line of a shared story
Jev Yarn is a party game where everyone writes the next sentence and Jev picks the winner: one taste request scores every line on four dimensions, and Nouls handle room filters and whether the story feels finished.
01 · Role in the system
What Jev does here
Every round, all requests go to api.typesafe.ai in parallel. One taste request carries four Choice questions - Funny, Flows, On theme, Surprising - over every candidate line, returning a probability per line per dimension that code mixes with the theme setter's slider weights into one share. Detail requests in chunks of twenty lines ask one Noul (would the story feel finished if this were added?) plus one Noul per room filter; a filtered line cannot win, its author sees why, and nobody else ever sees it. Past minimum length the closure probability blends into the share and decides whether the story ends. The finished story is rated by four Score questions on five described levels, combined by the room's weights into a score out of 100. Candidates are shuffled and lettered A, B, C, so Jev never sees player names. The author reports 100 players make a round six requests.
Taste as four choices, moderation as a threshold
The party game's machinery is a small lesson in judgment composition. Every round, one taste request carries four Choice questions — Funny, Flows, On theme, Surprising — over every candidate line, returning a probability per line per dimension that code mixes with the theme-setter's slider weights into one share. Nouls handle the room filters and, past a minimum length, whether the story feels finished — the closure probability blending into the share until it ends the story.
Fairness details are structural: candidates are shuffled and lettered so Jev never sees player names, a filtered line cannot win and its author alone sees why, and the finished story is scored by four five-level Score questions combined by the room's weights. The author reports a hundred players make a round six requests.
A moderation threshold, named as one
The limitation that matters most is stated flatly: moderation is a probability threshold — filters fire past 0.5 — not a human panel, and one room is deliberately unfiltered. Player picks nudge slider defaults but never change a round's outcome, and scores wipe when the room empties.
Within the games cluster, this record shows judgment composition at its most layered — taste, filtering, closure, and final scoring as separate typed questions over one shared state — while being honest that the safety layer is a threshold, not a person.
02 · Control boundary
Where Jev sits
One taste request (four Choices over all lines) plus chunked Nouls for filters and closure; deterministic code owns weighting, filtering, and the reveal.
Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.
03 · Known limits
What this evidence does not prove
- Moderation is a probability threshold (a filter fires past 0.5), not a human panel; one room is deliberately unfiltered.
- Player picks nudge the room's slider defaults but never change a round's outcome, and scores wipe when the last player leaves.
- The six-requests-per-100-players figure is author-reported from the author's own runs.
04 · Attribution
Public sources
This is a Community record: the project was published by a third-party community author.