An update note, for anyone following the collection. On 17 September nine measurements closed, two new chapters appeared, and three older chapters changed numbers — including one whose earlier reading did not hold. Each item points to the chapter where it lives in full.
| chapter | what it measures |
|---|---|
| 18 · Searching better does not fix the energy | Chapter 16 had built an exact search for the energy minimum over all 16,392 vocabulary tokens. This chapter measures what it does to the text: three different searches — the 8 most probable, the policy's nucleus, and the whole vocabulary — land at the same point of heading affinity and perplexity. What decides is not where you search, it is the energy. |
| 19 · The inversion that did not repeat | Chapter 15 had recorded that, with enough data, the world model started paying 3.38%. It was one seed. With more seeds, and then with a validation that stopped swinging, the cost shows up in every round: +1.77% on mini and +4.92% on pequeno. |
| chapter | what changed |
|---|---|
| 15 · What it costs and what it delivers | §5 rewritten. The reading that the world model "stops costing and starts paying" is gone. The table now carries both validation rulers and both sizes: the map costs on every seed. §3 filled in with CPU training speed for pequeno, medio and grande, missing since the chapter was written. |
| 16 · The minimum and the volume | §9 gained the contrastive round (InfoNCE) measured on the state matrix, with the weights kept: it has the deepest energy landscape of the three families and the highest cost to the generator. New §10: how many positions the volume needs to settle, and how much it depends on the size of the sampled points. The "0.00% against 82.49%" of §8 became a ratio of medians, close to 4. |
| 13 · The corrections from the meetings | §12.2 closed: 32-token blocks instead of 8 did not make the ζ latent choose between futures, and the subgoal is still reached within a few tokens, because two consecutive blocks are already too similar. The mask cell was filled in too. |
| 17 · What an ELLM is | The table of measurements swapped two numbers that came from the corrections above: the volume of low energy and the world model's cost on the full corpus. |
| 10 · What JEPA is | Two rows of the concept → code map that were "pending" are now filled: per-dimension anticollapse on level 2, and the distillation family with an EMA target. |
With that, no "pending" cell is left open in the collection: every published table is filled with closed rounds.
| where | what came to exist | why |
|---|---|---|
model/train.py --eval-seed | fixed validation batches at every evaluation | drawn validation swung more than the measured effect, and the "best checkpoint" rewarded the easiest sample (19 §4) |
server/ours.py: nucleo candidates | exact energy search only inside the policy's nucleus, with configurable mass | to test whether a firm plausibility constraint, with an exact search inside it, leaves the heading-versus-perplexity trade-off (18 §3) |
server/ours.py: fix in _todos_por_energia | the vocabulary ordering now uses the whole energy | it ignored the plausibility term, and the text came out at perplexity 2,098 (18 §2) |
model/collapse.py: --only-volume, --volume-seed, --sample-scale | measure only the volume, with another seed and another norm for the points | to separate the shape of the valley from the sampling choice (16 §10) |
Two of this day's changes correct claims this collection had published: the exact search for the energy minimum does not improve the text, and the world model does not start paying with more data. Both were found by measuring again what was already written — with more seeds, with the weights kept, with the ruler fixed. That is what the protocol and the number next to each claim are for: when the next measurement disagrees, it is possible to say exactly what changes.
Update note from 17/09/2026. The measurements are in the chapters cited.