Selection mattered. The model, so far, did not.

Every arm's champion, on a panel none of them saw.

Each run is one lane. An arm's champion is its child with the best certified training score in the whole run, ties to the earlier birth; the panel then scores it once, on eight instances that were never used for selection, two of them the instances with published records. The tick is PM-0, the founder, on the same panel.

What the champions look like, inside and out.

The founder and each arm's champion from era-0-1, packing the panel's (26, 1): 26 disks in the unit square, drawn from the packing the judge certified. PM-0 lays its disks on a grid, each 90% of half its cell. With no margin at all, that grid scores 0.20516 on the panel. No champion of any arm, in any run, scored above it; the best reached 0.20447. What selection found was a smaller margin for PM-0's own grid, not another arrangement.

Open the Museum

Four hypotheses, each tested once.

Each test is exact: the observed sum of the ten differences against all 1,024 ways of flipping their signs, so the p-value is a fraction and no distribution is assumed. H1 is the primary hypothesis, at 0.05. H2 to H4 share Holm's correction. The interval is the set of shifts the same test does not reject, 2.5% each way. A hypothesis is supported by a positive mean with p below its threshold, and a negative result is reported as plainly as a positive one.

Why H1 was expected.

Aleator has no heredity: every child is one edit from PM-0, so its champion can be no better than the best single edit the model offered. In all ten runs that best edit scored the same, 0.18515 on the panel: the model wrote one idea, a smaller margin for PM-0's grid, six different ways. The lineage and the blind rules could stack edits and keep shrinking the margin; Aleator could not.

So H1 shows that heredity lets small gains accumulate, the least that evolution must do. It says nothing about whether the model's edits are better than blind ones. That is H2's question, and H2 was not supported: the result that matters most here is the one the controls give.

What else the protocol asked for.

Section 4 names one sensitivity check and section 6 two descriptive measures. None of them is a hypothesis; each is reported as the protocol says.

The champion, generation by generation.

The run drawn here is era-0-1: the first run in by name, chosen by its name and not by its result. Aleator's champion is one edit from PM-0, as every Aleator child is. The lineage's champion, lineage-99-13, is 14 inherited edits deep, and the blind rules' 17.

How far the births got.

Every birth lands in one category. Stillborn: the edit was not the one asked for, so no life. Died: it lived and earned no score, by spending its fuel, crashing, or writing what the deciders cannot read. Lived: it was scored. Quarantine and void would mean the world or the judges failed; they are counted in the table. Counted over the ten confirmatory runs in.

Every run is reported. None is replaced or dropped.

A run's name is its seed: every draw in it comes from the record that grows from its name, so the same name gives the same bytes. Each run's head, the hash of its last record, is timestamped on Bitcoin through OpenTimestamps when it comes in.

The rules were frozen first.

The protocol was timestamped before the pilot's first birth, and every run's first record carries its hash. The one amendment, which moved the model to a single class of rented GPUs whose machines answer the same requests byte for byte alike, was timestamped before any data it affects.

What era 1 changed.

Every gain in era 0 was a smaller margin for PM-0's grid, and no champion got past what that grid can do. So era 1 asked whether a lineage can leave its founder's design, and changed four things, each fixed and timestamped before its first birth.

• The founder. Era 0's best champion, so the lineage starts where era 0 ended.
• A fifth mutation. The lineage may rewrite a whole function, one birth in five, so it can leave the grid in one step.
• The AI alone. Aleator gave way to the same model asked directly for a whole better program, as often as the lineage's tokens allow.
• A line to cross. 0.20516 on the panel, just above what the founder's grid can score with no margin at all. Crossing it is an existence claim: one champion, certified, is enough.

No champion of any arm crossed it; the best reached 0.20512. Blind rules scored higher than the lineage in nine of ten runs. The era 1 report.