Facts and views come from people. Round 5 adds a second kind of truth: what follows from rules and can be checked by calculation. Maths, the elements, what things are made of, balanced reactions, the forces, and a toy universe that runs from the first second to molecules that copy themselves. Every answer has a gate: code that knows the truth and judges what the model says.
These are toy simulations with real rules, strongly simplified. They are not a simulation of the real universe, and nobody knows yet how life began. What they give is a world where every statement can be recomputed: mass, charge and atoms are conserved at every level, and a run is the same every time for the same seed.
36 worlds computed by the chain. Pick one; under each era is the line the model is trained on.
Most worlds fail somewhere along the chain, and the place where they fail is itself a checkable fact.
q iron protons. a 2 6. q water molar_mass. a 1 8 point 0 1 5. q balance c 1 h 4 plus o 2 gives c 1 o 2 plus h 2 o 1. a 1 2 1 2. q stronger gravity strong. a strong. q solve 3 x plus 4 equals 1 9. a x 5. q check 1 2 plus 7 equals 2 0. a no.
Numbers are written digit by digit so the model can learn to calculate. Molar masses, mass percents and balances are computed from the formula, never typed in. Three independent checkers went through each table before it was accepted.
Yes, if it is fed differently. Same model, same size; only the diet changed. Questions the model never saw:
| Maths questions never seen | 9,000 lines, each repeated 200 times | 1.5 million lines, each seen 11 times |
|---|---|---|
| All kinds together | about 25 % | 93 % |
| Addition, subtraction, decimals | 98 to 100 % | |
| Solve for x | 17 % | 92 to 100 % |
| Multiplication with the working written out | 98 to 100 % | |
| Multiplication, bare answer | 56 % |
A small file repeated often is learned by heart. An endless stream of fresh, checked examples is learned as a rule. So round 5 trains on a stream from the gates, writes its working out step by step, and lets the gate check every step.
The model on the chat is still version 4b. Four experimental checkpoints and 36 reports have been exported and verified locally; they have not replaced the public model.
Selected checkpoints, evaluated after development-based selection. Each topic contributes 256 sealed questions; six legacy simulation-recall topics are reported separately. Earlier maths-only results above use a different dataset and training objective.
| Measure | 6 layers, width 384 | 8 layers, width 512 |
|---|---|---|
| Sealed exact, eleven topics | 34.48 % | 39.42 % |
| Accepted by each topic's gate | 39.88 % | 44.78 % |
| Old trained facts recalled (1,000 sampled) | 92.9 % | 96.9 % |
| Maths, exact | 63.67 % | 74.61 % |
| Worked maths, exact | 34 / 36 | 36 / 36 |
| Worked forces, exact | 22 / 128 | 37 / 128 |
| Worked substances, exact | 38 / 141 | 49 / 141 |
| World handoffs, exact / accepted | 28.13 / 87.50 % | 38.67 / 97.66 % |
World handoff numbers allow 5 % error; most other input-conditioned gates require the canonical four-significant-digit answer. Acceptance is therefore a mixed-tolerance measure, not uniform exact accuracy or proof of physical understanding.
Feedback improved the wide candidate's development acceptance from 43.79 to 44.32 %. It reduced the small candidate's score from 40.80 to 40.20 %, so the small base checkpoint was selected. Each feedback pass checked 2,816 new questions.
The later production refinements have development reports only: exact / accepted 35.30 / 40.52 % for small and 40.06 / 44.67 % for wide. Their old-fact recall was 90.7 and 95.2 % on a changed sample. They do not inherit the selected checkpoints' sealed scores. The small production run also used correction data despite its feedback candidate having been rejected.
The next upgrade should protect old-fact recall explicitly and teach intermediate calculation steps for weak numeric predictions. The simulation outcomes on this page describe the toy rules, not the real probability of planets or life.