Method

Can a mind ever learn the rules of the world it is inside, from inside?

That is the question under every experiment here. If a mind inside a universe we built can work out that universe's hidden rule, that is evidence that minds like ours can too. If it cannot, that is the deeper finding. Both answers are reported the same way.

Why evolution and not training

Training means a human writes a score and the machine climbs it. Whatever it learns is shaped by that score. Evolution needs one thing only: a body that runs out of energy dies, and a body with spare energy copies itself with small random changes. Nobody decides what is good. The world decides who is still here. That is the closest thing to a mind without a teacher that we know how to build.

It cannot be made with zero human decisions. A world needs physics, a scarcity, and a way to change. So the decisions are kept to eight, written down, and published before the run. Anything not on the list is not in the code.

The bias ledger

World: a wrapped square grid of cells that hold energy 0..1. Nothing else exists.
Energy in: a fixed amount enters the world each step (E_IN) and lands on cells chosen by one hidden 8-bit rule, starting from one seed event at the centre. From Experiment 3 on, every cell also regrows a trickle (one fiftieth of the normal share) so the world cannot starve itself to black.
Agents: a body holds energy up to E_MAX, burns a little each step to exist, more to move, more per hidden neuron. At 0 it dies.
Birth: at BIRTH_AT energy an agent copies itself and gives the child half. Weights, structure (which hidden neurons are on), the mutation rate itself, and the history-access bit are copied with random change. If birth slots are scarce, parents are picked at random.
Senses and actions: 3x3 local energy, own energy, 3x3 count of other bodies in; stay/up/down/left/right out. Numbered channels only, no names.
Experiment 2 only: an energy bonus, smaller than the cost of living, for predicting the next 3x3 energy. Predictions 5 and 20 steps ahead are scored but not rewarded. Agents also emit 8 bits we score against the hidden rule they never see. A heritable bit gives read access to the energy field as it was 20 steps ago (recorded history); half the founders have it.
Randomness: numpy/torch PCG by default; a quantum random byte file can replace the seed (--qrng).
Measurement, not physics: a dead agent's record (5 action shares, lifespan, crowding, mobility, eating rate = 9 numbers) is kept in an archive if it is far from everything already seen. Archive growth is the metric. This changes nothing in the world.
Experiment 3 only: energy a body holds leaks away by 0.5% each step (second law; nothing stores energy for free).
Experiment 3 only: one bit of the hidden rule flips every 100,000 steps; which bit is random. Bodies are never told. The world does not hold still.

What this cannot tell you

Nothing here can say how our own universe began or why you exist. A simulation can only study its own universe. It can test whether a mind can learn its universe from the inside, and that is the only claim made.

The archive measures nine numbers. A richer life could be happening in numbers not recorded. A flat archive may be a coarse ruler, not a dead world.

The machine is one laptop. Earth had a planet and four billion years. Expect bacteria-scale results and judge them by the pre-registered tests, not by hope.

What counts as a finding

A verdict on this site is the result of one run with one random seed. One run is an anecdote. From 4 October 2026, 20:30 ET, nothing here is called a finding until the same verdict holds in at least 3 of 5 runs with different seeds. Until then the page says what the test says and no more. The two baseline runs will be repeated four more times each when the machine is free; the replication count will be shown next to each verdict.

Open questions the current runs cannot settle, listed so they are not forgotten: which of Experiment 3's three changes does the work, if any (needs one run per change); whether the novelty ruler's nine numbers miss kinds of behaviour that exist (needs a second, different ruler on the same runs).

Corrections, in the open

4 Oct 2026 — Energy too abundantNobody died, so nothing was selected. Fixed with a fixed energy inflow per step and a cap on what one body can hold.
Fixed before the public run.
4 Oct 2026 — Reward larger than the cost of livingBodies lived on the prediction reward alone; zero deaths, zero evolution. Reward cut from 0.03 to 0.008 against a living cost of 0.01.
Fixed before the public run.
4 Oct 2026 — Parents chosen by memory positionWhen birth slots were scarce, bodies in low memory slots always reproduced first: a hidden human bias. Now random.
Fixed before the public run.
4 Oct 2026, 20:35 ET — A belief tested and found false: the regrowth rule does not starve the gridControl run, Experiment 1 physics with no bodies: the grid went from 0.2% energy to 100% full within 2,000 steps and stayed full. So the dark grid in Experiments 1 and 2 is the bodies eating it, not our rule failing. The "trickle" added to ledger line 2 for Experiment 3 was justified by the false belief. It stays, because the Experiment 3 tests were frozen with it in place; but it is now an untested extra, not a fix, and Experiment 3b (one change at a time) must include a no-trickle arm.
Claim withdrawn. Control result: 100% full by step 2,000.
4 Oct 2026, 20:30 ET — Overstated wording corrected"Evolution switched itself off" was replaced with "slowed below detection" on this site and in the project notes: 500,000 births happened, so evolution ran; what stopped was measurable novelty. "Grid starves itself because of our rule" was downgraded to an untested claim with a control run planned. "Two things from physics" on the Experiment 3 page was corrected: the energy leak is physics, the flip clock is a human choice.
Three claims pulled back to what the data supports.
4 Oct 2026, 20:05 ET — Ledger grew from 8 lines to 10 for Experiment 3; line 2 gained a regrowth trickleExperiments 1 and 2 keep running on the original 8 lines; their code is untouched. The trickle on line 2 is a correction: the original rule let the world eat itself to ~1% energy, an artefact of our rule and not a feature of any real world. Lines 9 (energy leak) and 10 (rule flips) are physics constraints the still worlds lacked. Tests c5–c8 were frozen in the same edit, before any Experiment 3 code ran.
Recorded before the run.
4 Oct 2026, 19:40 ET — Pre-registered tests c2, c3, c4 reworded after the runs had startedOld c2 fail: "never above 4". New: "never held 5 of 5 by 1,000,000 steps". Old c3 pass had no judging step and no fail clause; new: "at 1,000,000 steps … fewer or equal". Old c4 pass had no judging step; new adds "at 1,000,000 steps". Reason: the originals could never produce a Fail, which made the page unable to say what it promised to say. The thresholds and directions did not change. The runs were at step ~130,000.
Recorded, not hidden. Future edits to frozen text go here too.
4 Oct 2026 — Bodies were blind to each otherEarly worlds sensed only energy, so cooperation could never appear. A 3×3 other-body channel was added (ledger line 5).
Added before the public run.

Run it yourself

The whole universe is one small program. Download the code and it runs on any laptop. The numbers this site reads are public at the raw feed; the one-line readings under the charts are public at their feed.

Who

Zero World is a ZeroOrigine experiment, built and run by Jagdish Lade. Questions and challenges are welcome; corrections are posted on this page, never hidden.

About the readings under the charts

Every two hours a language model is given the numbers, and only the numbers, and asked for one plain sentence per chart saying what moved. It never sees this page and is forbidden words like "learn", "understand" or "intelligent". A sentence that breaks a rule is dropped, not shown. Each reading carries the step it was written at, so you can tell a stale one from a fresh one. It is a reading, not a verdict; the verdicts come only from the pre-registered tests.