Thirty-two runs cannot tell the lineage from blind rules.

Every arm's champion, and the step to take.

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 PM-2, designated and not selected; the guide line is 0.003 above it, where a second step begins.

One primary test, one gated behind it.

H3 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 is tested only if H3 is supported; as the only secondary it needs no correction. E is an existence claim: one champion at least 0.003 above PM-2 on the panel, certified exactly by both deciders, is proof of a second step.

The champion, generation by generation.

The run drawn here is the first, by name, whose lineage champion took a second step; until one has, the first in. The lineage's champion, LQ-78-14, is 44 inherited edits from PM-2, 17 of them rewrites by the model and the rest by blind rules; every AI-alone child is one answer about PM-2.

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.

Open the Museum

How far the births got.

Counted over the thirty-two confirmatory runs in. Every rewrite by the rules 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; a child over 16,384 bytes was stillborn for the cap.

What the second step cost, and what era 4 could ask.

Era 3 started from an organism that evolved, PM-2, with the hexagonal layout already in it. Under the same selection that found that layout in era 2, no lineage took a second step of 0.003 on the panel in thirty-two runs, and blind rules took none either; the model asked alone did once, with a layout it had also written from PM-1 in era 2. Several champions that tied the founder on the training set scored lower than it on the panel: a change that is neutral on eight instances is not neutral on eight others. And the genome cap of 16,384 bytes, fixed since the M2 era, left PM-2 (13,098 bytes) little room: 314 of the AI alone's 514 answers and a share of every arm's duplications were stillborn for it. These are readings of the record, not tests; nothing below is fixed until its protocol is written, hashed and timestamped before generation one.

• A cap that fits the founder. An era whose founder evolved needs a cap set from the founder, so a whole-program answer and a duplication have room; the cap is part of the physics, so it moves at the era boundary.
• Selection that sees the panel's kind of variation. The training set is eight instances; champions overfit to them. More training instances, or a training set redrawn at checkpoints, would make a nudge that is real on the panel distinguishable from one that is not.
• The AI alone's one step. It came from a whole program the model also wrote in era 2. Whether the model has further layouts in it, or only that one, is a question for the model alone, asked more times from PM-2, before any lineage is asked to find more.
• Era 2's readings on era 3's record. The landscapes, the knockouts against PM-2 and the recall flag, as the protocol's §12 names them, are not run yet.

Asked since, outside the record (2026-10-27). A trial for era 4 asked the model alone 1,000 more times from PM-2 and spliced era 3's 529 answers again, under a cap of 65,536 bytes and the same judge. Under the lifted cap 1,304 of the 1,529 were born, against 399 under era 3's. Single steps stay rare: 12 answers step 0.003 above PM-2 on the panel and hold on sixteen fresh instances, 0.8% an answer. Pieces are common: 201 answers beat PM-2 on some panel instance while most lose overall. The trial is not an era and changes nothing on this page. Eras are now frozen, and the Laboratory takes those answers apart.

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, decides whether the era runs sixteen or thirty-two.

Every run was replayed, one model answer in ten asked again: thirty-three of thirty-three reproduce every line. strata/era-3.replay-2026-10-27.json