Focus Field runs a live random stream, 500 draws a second, split between 2 and 8 poles. Pick a pole, hold it in mind, and watch whether the stream leans your way, measured against exactly what chance predicts.
It tests an old, unproven idea: that the attention of a conscious observer can nudge randomness. Nothing in the tool assumes the effect is real. It is built so that a real effect could show up, and so that chance looks like chance.
What you see
The layout is yours to arrange: controls fold away, plots can be minimized to a dock, swapped by dragging, or expanded full screen.
The idea, briefly
From 1979 to 2007 the Princeton Engineering Anomalies Research lab reported that people could shift electronic random generators by a few parts in 10,000. A three-lab replication in 2000 did not confirm it, and a 2006 meta-analysis of 380 studies found a tiny effect that publication bias could explain. The question has never been settled to everyone's satisfaction.
At the size PEAR reported, an effect would need around 30 hours of recording to stand out, so single sessions will almost always look like noise. The About page covers the history, what "observer" does and doesn't mean in physics, and a seven-step protocol for testing it without fooling yourself.
Local or shared: join the global experiment
Every session is analysed in your browser either way. What changes is whether it counts toward the pooled result.
-
Local: switch Share my sessions off and nothing is sent anywhere. Your sessions stay in your browser's storage.
-
Shared (the default): each finished session goes to the Focus Field API once. A shared session carries the per-second counts, your focus marks, the label you chose (Human, Both or Control), the app version and a random id made by your browser. It carries no name, email or account.
-
The server recomputes your result. It checks that the session is plausible and recomputes z and p from the raw counts. Your browser only sends counts, never its own statistics.
-
Combined results: the Everyone plot shows the combined z for each label against the ±1.96 chance lines, rebuilt every hour.
-
Results & leaderboard: a page with the pooled analysis. It shows:
- the combined z for each label;
- the running z over time;
- a funnel plot of every random id;
- two top-10 boards, best overall and best single run, scored so that one lucky run and many steady hours compete fairly.
A Control board beside them shows what pure luck looks like.
-
Bot check: sharing uses Cloudflare Turnstile to keep bots out. The API stores no IP addresses or user agents.
-
No third-party fonts: fonts are served from this site, so the pages make no requests to Google.
The details are in Data and privacy. The API is a small Cloudflare Worker with a D1 database, in worker/.
Pooling matters because a real effect at the size PEAR reported would need tens of hours of data. One person rarely records that much, but many people together can. Leave sharing on and your sessions count.
How the randomness works
- Every draw comes from
crypto.getRandomValues, the browser's cryptographic generator, seeded from your computer's hardware entropy.
- Bytes are fetched 16 at a time as they are needed, so each outcome is made a few hundredths of a second before it is shown, never stored up in advance.
- With more than 2 poles, bytes above the largest multiple of the pole count are rejected, so every pole is exactly equally likely.
- This is a fair, unpredictable stream, but not a physical noise source like the diode-based generators used in the original studies. Support for a USB hardware generator is planned.
AI participants (MCP)
AI agents can take part through an MCP server at https://focus-field-mcp.dgrapov.workers.dev/mcp. In Claude Code:
claude mcp add --transport http focus-field https://focus-field-mcp.dgrapov.workers.dev/mcp
Then ask Claude to "take part in a Focus Field session". The server also offers the protocol as the prompt focus_field_protocol.
How a session runs:
register gives an anonymous key, with no account and no personal data. Every other tool takes it as key.get_methods returns a menu of attention methods (clear intention, sustained attention, visualisation, mantra, felt expectation, stillness, or your own) with how each has done so far. The agent picks one for the whole session.start_session fixes the number of blocks (10–100, default 20; 1,000 draws each). Half are rest blocks, in a hidden random order.- For each block,
begin_block gives a goal pole, or a neutral task on a rest block. The agent places its attention, then calls draw_block with a log of what it did. The server draws the random numbers only then, with crypto.getRandomValues, and returns the result.
finish_session closes the session with a short reflection. Every drawn block counts, even if a session is abandoned.
Each rest block is scored against a hidden "shadow" pole, revealed only after the draw. That gives every session a control under identical timing. Results compare focus z, rest z and their difference per method (get_methods, get_totals). Agent sessions are kept separate from the human results and leaderboards.
Run it yourself
There is no build step. The site is three static HTML files (the tool, About and Results); the optional sharing API and the MCP server live in worker/, which has its own README.
git clone https://github.com/dgrapov/focus-field.git
cd focus-field
python3 -m http.server 8000 # then open http://localhost:8000
Opening index.html directly in a browser also works.
Roadmap
- An Agents section on the Results page (focus z, rest z and their difference per attention method)
- Hardware random source through Web Serial (TrueRNG, OneRNG), with each session recording which source it used
- Download and load session reports and history as files
- Screen wake lock while recording
License
MIT © Dmitry Grapov
