Haishool round 7 · a toy universe with corrected rules

From the first minutes to a state.

Each world runs one chain of small simulations on one clock: the light elements, generations of stars, a collapsing cloud, a star and its planets, chemistry in an ocean, replicators, cells, bodies, senses, a language and finally a society. Every number can be recomputed, and every world is the same each time it is run.

One world, era by era

The first six are rare worlds picked out of 300; the rest are worlds 1 to 36 as they come. Ages are years since the Big Bang.

How far 200 worlds get

The highest rung each of worlds 1 to 200 reaches, in ladder order. Grey rungs are where a world stops before life; green from replicators on.

Most worlds stop early: no temperate planet (50), a clump too small to become a star (25), or cells that never become complex (39). Language with grammar appears in 3 of 200, a state in 1.

What was corrected

First minutes

One universe instead of a different one per world: 75.3 % hydrogen, 24.7 % helium, the measured values. Photon heating and the neutron lifetime now set when helium forms.

Stars

"3 % of hydrogen per generation" is gone. Stars now have masses, lifetimes and fates: white dwarfs, supernovae and the elements each returns.

Planets

The temperate zone and the surface temperature are one rule, so "no water" no longer comes from two rules disagreeing. Giants can migrate, mergers conserve angular momentum.

Chemistry

Bond strengths from bond enthalpies, lightning as an energy source, and organic means a carbon–hydrogen bond.

Replicators

No more perpetual motion: spent building blocks, hydrolysis of every bond, copy time grows with length.

The clock

One time axis in years. The star's lifetime ends the world, so a short-lived star leaves no time for evolution.

The worlds of round 6, which the version-5 models were trained on, stay unchanged; the corrections are a switch, checked by 56 frozen tests.

What is still a toy

These points come from an independent review of round 7. They are the next round's work: one population carried up the ladder, grouping and communication paid for in fitness, meanings grounded in the senses, and one continued history instead of best of six.

Training

A model of 33 million parameters was trained on these worlds alone (720,000 lines, 105,000 steps on one RTX 4090): lessons whose answer follows from the question (80 % of the tokens) and the stories of 256 worlds (20 %). Questions it never saw:

Rules: 39 %

Best where the answer is short or a decision: chemistry 63 %, the chain's hand-offs 65 %, replicators 55 %, language 52 %. Weakest where it must do multi-digit arithmetic with powers and roots: gravity 16 %, first minutes 17 %, society 23 %.

Stories: not memorised

On the worlds it was trained on it gets 29 % right, about what it guesses for worlds it never saw (32 %). It did not learn the 256 languages: 0 of 15 sentences said right, 27 of 107 meanings.

The loss was still falling when training stopped (0.88 to 0.50), so a longer run should help. Development and sealed scores agree (39.7 % and 39.4 %), so the rules it learned hold on new inputs.