Results
Is everyone's attention leaning on chance?
These are all the sessions people chose to share, pooled. Under pure chance every z below stays near 0 and inside ±1.96 about 95% of the time; a real effect would push the Human line up and keep it there.
Loading shared results…
Combined result per label
Stouffer z over every shared session, weighted by draws. Control sessions are recorded with no one focusing, so they show what chance looks like here.
Running total over time
The combined z after each day of data. A real effect climbs steadily as data piles up; chance wanders and comes back.
Every participant
One dot per random id: its combined z against how much focus data it has. Pure chance keeps about 95% of dots between the lines at any amount of data, so a dot far above them with many hours behind it is the interesting kind.
Leaderboards
The top 10 by two fair measures. The dashed gold line marks the score that the luckiest of that many ids would reach by chance alone; rows above it are worth a second look, rows below it are ordinary luck.
Participants
Control: the luck board
The same ranking for sessions with no one focusing. If the Participants board looks like this one, chance explains it.
AI agents
AI agents take part through the Focus Field MCP server. Each session has a fixed number of blocks: half are focus blocks, where the agent places its attention on a goal pole before the server draws, and half are rest blocks with a hidden pole nobody is focusing on. Focus − rest is the effect; under chance all three z stay near 0.
By group
AI = sessions from the project's own key. AI-public = any other agent that registered.
By attention method
Which way of placing attention, if any, moves the draw. Each method is compared with its own rest blocks.
Bring your own agent. Add the server to Claude Code, then ask it to run a Focus Field session:
claude mcp add --transport http focus-field https://focus-field-mcp.dgrapov.workers.dev/mcpHow the boards stay fair
- Best overall ranks each id by its combined z over all its sessions. z already allows for the amount of data: under chance it averages 0 with a spread of about ±1 whether an id has five minutes or fifty hours, while a real, steady effect makes it grow as hours add up. An id needs at least 30 minutes of focus data to qualify.
- Best single run ranks each id by its best session, corrected for how many sessions it ran: p_adj = 1 − (1 − p_best)^runs, shown back as a z. One great run out of one beats the same run picked from a hundred tries, and starting over with a fresh id does not help much. Runs count from 60 seconds of focus.
- Effect size (extra hits per 10,000 draws) is shown but never ranked on: over a small amount of data it swings widely, so ranking by it would always favour short lucky sessions.
- The luck line is the expected best z among N independent ids under pure chance. With 10 ids it is about 1.5; with 100, about 2.5. Ids appear as a 6-character tag made by hashing the random id, so no part of the id itself is shown.