004 / EXPERIMENTS IN EMERGENCE
Order from chaos.
One ant. Two rules. Absolutely no master plan.
The classic ant at 11,000 ticks. Press Play to follow the highway.
AUTONOMOUS DISCOVERY / CAMPAIGN 001
Small rules.
Strange outcomes.
52 turn sequences. Five starting conditions each. The lab screened every combination, picked eight different rules, and ran those to one million ticks.
Verifying the discovery archive…
Download all 260 results ↓How the lab chose these results
Every case ran for 20,000 ticks. Starting conditions were blank, 10% noise with seed 42, 30% noise with seeds 137 and 2026, and two opposing ants. Rules contain two to six L/R instructions. Only first-turn R rules are included, choosing one from each mirror pair on a blank grid; repeated shorter strings are omitted. Fixed noisy grids do not necessarily preserve mirror equivalence.
The controller selected leaders for ground covered, compactness, 180° color symmetry, sensitivity to starting noise, late exploration, and travel distance. Remaining places use the mean percentile across these measures. Each finalist has a different rule and at least 100 nonzero cells at screening. These are transparent taste heuristics, not a universal definition of interesting.
The eight finalists then ran for 1,000,000 ticks. A separate implementation verified every final cell and ant position. Selection used the screening results; the longer follow-ups may behave differently. Previews fit the whole pattern into an image; replay and JSON retain the exact grid.
This completed campaign is an automated deterministic search, not AI learning or an always-running job. Reproduce it with the documented runner. Finite observations do not prove endless repetition or novelty.
AUTONOMOUS EXPERIMENT / 01
Does order survive
a little interference?
Start once. The runner compares 13 conditions, records every outcome, then stops. Select a completed trial to replay it.
13 trials · 100,000 ticks each · no provider calls
| Trial / replay | Rule | Visited | First 104-step motif | At endpoint |
|---|---|---|---|---|
| Run the study to collect measured results. | ||||
THE RULES
Local decisions.
Unexpected structure.
On a white cell: turn right. On a colored cell: turn left. Flip the cell, then move one square. Repeat.
For multicolor rules, each letter chooses a turn for that cell’s state. The cell advances to the next state, wrapping to zero.
Langton’s ant ↗ Further travels with my ant ↗WHAT WE ACTUALLY TEST
Choose 30,000, 100,000 or 1,000,000 ticks per trial. The classic RL control, a single cell flipped ahead of the ant at tick 12,000, eight seeded noisy starts, a two-ant world, and two multicolor rules. Initial noise fills a 19 × 19 patch at 10% or 30% density. Seeds: 42, 137, 2026 and 9001.
The detector compares five consecutive 104-step traces for the same states, headings and nonzero translation. It only applies to single-ant RL runs. “Observed” is finite evidence; “not observed” means not detected within this horizon. Other periods are not tested. Endpoint uses the latest sample at or before the horizon.
Two ants update in fixed order on the same grid. A tick moves each ant once. The world has no wrapping boundary. The camera follows the first ant when a full-world view would make cells smaller than two pixels. Whole world always shows the full extent; Follow ant preserves local detail. Camera changes affect only your view.
This is a deterministic autonomous controller, not an AI model. It runs locally in a browser worker and stops after the catalogue. No background server job, Jev calls or online learning. Hidden tabs stop the study; exported receipts retain configurations, initial/final cells and interventions.