The model's lineage beat blind rules. The same model, asked alone, did better.

Every arm's champion, and the line to cross.

Each run is one lane. An arm's champion is its child with the best certified training score in the run; the panel then scores it once, on eight instances never used for selection. The dotted tick is the founder, designated and not selected; the guide line is 0.20516, the ceiling of the founder's grid.

One primary test, three gated behind it.

H2 is the exact sign-flip test on the sum of the sixteen differences, against all 65,536 ways of flipping their signs. Its verdict is read from the exact 90% interval for the mean shift, with an equivalence margin of 0.003. S1 to S3 were tested only because H2 was supported, under Holm's correction. E is an existence claim: one champion above the ceiling, certified exactly by both deciders, is proof that a lineage left its founder's design.

The champion, generation by generation.

The run drawn here is the first, by name, whose lineage champion crossed the ceiling. The lineage's champion, LQ-40-1, is 22 inherited edits from the founder, 8 of them rewrites by the model and the rest by blind rules; every AI-alone child is one answer about the founder.

The founder and the champions, inside and out.

Above, each form: its genome laid out by one page of rules, ringed in blue where it changed since the founder and filled blue where it grew something new. Below, its packing of (26, 1), scored exactly. The founder's grid stops at 0.20516.

Open the Museum

How far the births got.

Counted over the sixteen confirmatory runs in. Every rewrite by the rules or by the tree also had to compile and be one function of the same name; a child whose program was already judged in the run carried that decision instead of living again.

Where the steps were, and what carried them.

Three readings of the record that the protocol names and does not test: the landscape of tree-edit distance against fitness, knockouts on the champions that crossed, and a flag for layouts that are a literal table of coordinates. Nothing here was pre-registered; nothing here is a test.

In 22 of 25 champions, one unit carries the crossing.

Each champion that crossed the ceiling was taken apart one unit at a time: a function of main.py that PM-1 also has was swapped for PM-1's, a function PM-1 does not have had its body replaced by a raise, and a file in modules/ was removed. Each knockout lived once on the panel, under era 2's budget and the run's seed, judged by decision/2, on one machine. Every champion lived there first as the control: 25 of 25 scored the record's panel score to the last digit on the same fuel. A unit "carries the crossing" when the champion falls to or under the ceiling without it; "contributes" when it scores less but stays above; "held it back" when the champion scores more with PM-1's version.

The recall flag: 243 layout genes read across the 25 champions, none with a literal table of coordinates (the rule: a list or tuple display of at least 12 numbers, or at least 4 nested rows of 2 or 3). None of the champions is a table; each computes its layout.

What era 3 should ask.

Era 2 found structure that blind rules and blind tree search did not, and found it only where the stronger model wrote the rewrites. It also found that the same model, asked alone for whole programs, did better than its own lineage. These are proposals; nothing below is fixed until its protocol is written, hashed and timestamped before generation one.

• Selection over whole programs. The AI alone wins with a dozen answers a run. A lineage whose rewrite is the whole program, under the same selection, would ask whether heredity adds anything to what the model already writes in one answer.
• Steps after the first. Six of the ten lineage champions that crossed score the same on the panel to nine digits, 0.216272380, and only one went further. The landscapes above (Plate VI and the steps table) say what came after: the crossing step was one rewrite 50 to 123 edits from its parent, a layout written whole in one answer, and every later step of the lineage gained at most +0.0023 in training score. A second large step was never drawn in 1,600 births a run.
• Units and habilis. Modules as content-addressed units, so a layout gene found once can be reused by name.
• The island file system. Lives that make millions of file calls sit near the host's wall-clock guard, because each call is answered in Python. A faster island, at an era boundary, takes them away from it.

Every run is reported. None is replaced or dropped.

A run's name is its seed. Each run's head is timestamped on Bitcoin through OpenTimestamps when it comes in. The sign-blind rule, which sees the sizes of the differences and never their signs, kept the era at sixteen runs.

Every run was replayed on the pod that wrote it, one model answer in ten asked again: thirteen of sixteen reproduce every line. In the other three, one life each differs, every one a life of millions of file calls near the host's 300-second wall-clock guard, which is not part of the physics: two met it in the replay and reproduce when relived alone, and one, the era's only quarantine, met it in the run itself and meets it again alone. No line that selection, a champion or the summary depends on differs. Away from home, era-2-1 replays every line from a fresh copy on a machine that did not write it. strata/era-2.replay-2026-10-26.json