A synthetic machinid is an organism made by hand or by a stated rule, never by an era. Its genes come from material already kept: here, 1,529 whole programs the AI wrote about PM-2, era 3's founder, 529 in era 3 and 1,000 in the era 4 trial, run before any era 4 exists. Run whole, most of those programs scored lower than PM-2. The Laboratory takes them apart: every layout function is lifted out as a gene, read as text and never run outside the sandbox, and PM-2 carries it beside its own genes, so a gene can only add. Each synthetic is then judged by the same two deciders, under era 3's physics, on the panel and on sixteen instances no choice ever read. None of this is evidence of what evolution did; it says what can be built from what the AI wrote.
Above, each form: its genome laid out by one page of rules, filled blue where it carries a gene PM-2 does not have, a ghost where a gene was taken away. Below, its packing of the panel instance chosen, as the judge decided it; every frame is dotted, because every specimen here was chosen, by hand or by a rule, and none was selected by an era. Choose another instance to see where each gene wins. The highest panel score in the record, eras 0 to 3: 0.23247, era-2-11's alone champion DQ-9; PM-2 is 0.22543. Those organisms never lived on the check set, so against them the comparison is on the panel only, where every synthetic here was chosen.
Era 3's founder: era 2's champion LQ-95-1, taken whole. Every synthetic here is built on it.
PM-2 with its hexagonal genes replaced by the square grid it started from: it falls back to the ceiling of era 2's founder.
PM-2 carrying the library's best gene, trial:577's generate_hexagonal. Run whole, the answer it came from stepped +0.0055 on the panel and only +0.0010 on the check set; at 22,315 bytes, era 3's genome cap would have killed it.
PM-2 and, for each panel instance, the gene that scored it highest: five genes. With trial:577's among them, the other four add nothing; without it, they still hold about two-thirds of the gain.
An operator from era-3-22:DQ-12 (optimize_layout) applied to the library's best gene, joined by rule and chosen on the panel from 210. Run whole, DQ-12 scored 0.0203 below PM-2, and era 3's genome cap killed it at 30,953 bytes.
The slim chimera without the gene its knockouts found doing the work: the four left only tie it where they were chosen.
From the era 4 trial, run before any era 4 exists: PM-2 and five whole answers, each run as written, the best candidates kept.
PM-2 made to believe its own check and claim every radius 0.5% larger than it is. The judge refutes it and names the disk it catches leaving the box or overlapping another.
Corrected. PM-2 + one gene (lab/corrections-2026-10-27.json, 2026-10-27): The note said the answer lost overall. It did not: run whole, it stepped on the panel and did not hold on the check set (runs/era4-trial, trial:577). The slim chimera (lab/corrections-2026-10-27.json, 2026-10-27): The note said one gene does all the work. Its knockouts say less: taking that gene away costs 0.0037 of the 0.0105 on the panel, so the other four are redundant, not idle (lab/chimera-2026-10-27.json). The best composition (lab/corrections-2026-10-27.json, 2026-10-27): Added how many the panel chose among, and what the operator's answer did whole (runs/era4-trial, era-3-22:DQ-12). The hand-joined chimera (lab/corrections-2026-10-27.json, 2026-10-27): "The era 4 trial's" could be read as an era 4. There is none: eras are frozen, and the trial is not an era.
Every top-level function of (n, h) in the answers, lifted with the helpers it calls and carried by PM-2 as one more gene. A whole answer replaces PM-2's program and loses what PM-2 already did well; a carried gene only adds, so most genes sit at zero and the ones that win show plainly. Drawn: the 260 genes that add anything.
For each panel instance, the gene that scored it highest was given to PM-2. Each gene was then taken away once. Taking away any of four costs nothing: one gene, trial:577's, covers them. Taking that one away costs about a third of the gain, and the other four hold the rest. They are redundant organs, not idle ones.
| gene (answer · function) | bytes | panel instances it wins alone | chimera without it: panel | check |
|---|---|---|---|---|
| trial:330 · triangular_layout | 998 | (26, 1/1), (36, 7/4) | +0.0105 | +0.0072 |
| trial:536 · hex_layout_vertical | 1,028 | (26, 3/2) | +0.0105 | +0.0072 |
| era-3-25:DQ-2 · generate_hexagonal | 1,096 | (26, 1/1), (30, 15/8), (36, 7/4) | +0.0105 | +0.0072 |
| trial:577 · generate_hexagonal | 3,178 | (26, 1/1), (26, 3/2), (30, 15/8), (38, 7/4), (36, 7/4) | +0.0067 | +0.0046 |
| era-3-30:DQ-7 · generate_hex_layout | 894 | (36, 7/4) | +0.0105 | +0.0072 |
210 operators on layouts (optimizers, local searches, perturbations) were lifted from the answers and each applied to the library's best gene; 12 beat that gene alone on the panel and on the check set, and 2 ran out of fuel (not drawn). The rule joins two pieces the AI wrote in different answers; the panel alone chooses, and the check set reports. An optimizer costs fuel: across is what each life burned.
No model and no hand: the pool is fixed, the 2,158 genes and 210 operators above, as text. Arm S keeps sixteen organisms for a hundred generations, era 3's budget; each child is its parent with one change drawn at random (add a gene, wrap a carried piece in an operator, drop a piece), and the best sixteen live on. Arm A spends the same 1,600 births with no inheritance: PM-2 and one random gene each time. Each run chooses on eight training instances of its own; its champions then live once on the panel and on the check set, so neither score was chosen on. The rules were written before the first run lived (lab/README.md, the amendments). Eight runs, every one shown. This is not an era: it says what blind selection does with this pool, not what any era did.
Against the best single gene on the check set (+0.0072), selection's eight champions stand 4 above it, 1 level with it (within 0.0002), 3 below; the best is +0.0131, against the rule-made composition's +0.0136, and none passes it. The median champion: +0.0080 with inheritance, +0.0053 without; selection is higher in 6 of 8 runs. In 5 of 8 runs it carries era-3-22:DQ-12's optimize_layout, the operator the composition uses, found blind. The cost: its champions burn 19% to 100% of a life's fuel, against the composition's 7.3%, and 2 of 8 ran out of fuel on the panel and on part of the check set after scoring above PM-2 on their training set: nothing in era 3's rules charges for fuel until a life runs out, so selection filled the budget and fell over the edge on instances it had not seen.
| run | S: pieces | S: fuel | S: panel | S: check | S carries era-3-22:DQ-12's optimize_layout | A: panel | A: check |
|---|---|---|---|---|---|---|---|
| lab-select-1 | 21 | 100% | −0.2254 | −0.1103 | yes | +0.0067 | +0.0053 |
| lab-select-2 | 12 | 19% | +0.0112 | +0.0120 | yes | +0.0067 | +0.0053 |
| lab-select-3 | 19 | 73% | +0.0149 | +0.0109 | yes | +0.0105 | +0.0072 |
| lab-select-4 | 14 | 57% | +0.0123 | +0.0070 | +0.0067 | +0.0053 | |
| lab-select-5 | 16 | 48% | +0.0154 | +0.0090 | +0.0004 | +0.0009 | |
| lab-select-6 | 15 | 33% | +0.0088 | +0.0038 | +0.0063 | +0.0037 | |
| lab-select-7 | 19 | 68% | +0.0201 | +0.0131 | yes | +0.0105 | +0.0072 |
| lab-select-8 | 10 | 100% | −0.2254 | −0.1112 | yes | +0.0022 | +0.0000 |
The same eight runs again, by the same names, so the same training sets and the same random draws, with one rule added and written down before any of them lived: a child whose life on its training set burns more than half of a life's fuel dies there, and is never selected or champion. Everything else is Plate V's. A life does not depend on the rule, so lives the first runs had already lived are read from their rows; the rest lived on rented CPU, after each machine re-lived run 1's first ten generations identically.
Under the half-fuel rule 0 of 8 champions ran out of fuel, against 2 under the first rule. The median champion on the check set: +0.0075, against +0.0080. Against the best single gene (+0.0072) they stand 4 above it, 3 level with it (within 0.0002), 1 below; the best is +0.0120 (the rule-made composition: +0.0136). Run by run, the rule's champion is higher on the check set in 3 of 8, the same organism in 2 and lower in 3: the heavy optimizers it kills carried the best first-rule champions as well as the starved ones. Its champions burn 12% to 52% of a life's fuel on the panel; the rule killed 164 children in all, and 4 of 8 champions carry era-3-22:DQ-12's optimize_layout.
| run | first rule: fuel | check | half-fuel rule: fuel | check | children it killed | carries DQ-12's optimize_layout |
|---|---|---|---|---|---|---|
| 1 | 100% | −0.1103 | 24% | +0.0081 | 29 | yes |
| 2 | 19% | +0.0120 | 19% | +0.0120 | 9 | yes |
| 3 | 73% | +0.0109 | 26% | +0.0073 | 31 | |
| 4 | 57% | +0.0070 | 52% | +0.0073 | 27 | |
| 5 | 48% | +0.0090 | 12% | +0.0088 | 7 | yes |
| 6 | 33% | +0.0038 | 33% | +0.0038 | 14 | |
| 7 | 68% | +0.0131 | 51% | +0.0078 | 23 | |
| 8 | 100% | −0.1112 | 20% | +0.0072 | 24 | yes |