Machinidae working draft · era 1 · 10 of 10 runs in
Era 1 · results, built from the record

No lineage got past its founder's own design at this budget.

Era 0 found that selection mattered and the model, so far, did not: blind rules under the same selection did as well as the model's lineage, and every champion improved by turning one knob, a smaller margin for the founder's grid. Era 1 starts from era 0's best champion and asks whether a lineage can get past its founder's own design, and asks again whether the model adds anything to selection. The lineage now has a fifth mutation, rewriting a whole function. The AI alone is the same model asked directly for a better program, as often as the lineage's tokens allow, with no heredity and no feedback. Blind rules stay as in era 0. The line to cross is 0.20516 on the panel, just above what the founder's grid can score with no margin at all.

Does the model add anything to selection? The blind rules' champion, with no model at all, scored higher than the lineage's in nine of ten runs: H2 is not supported, p = 1023/1024. Does selection beat the same AI asked directly? The lineage scored higher than the AI alone in seven of ten: H1 is not supported, p = 9/128. No champion of either arm crossed the ceiling.

Open the MuseumEra 0's result
10/10confirmatory runs in, named before the first began; none is replaced or dropped
0champions above the ceiling, lineage and AI alone, of twenty
9/128 · 1023/1024H1's and H2's exact p-values, each over all 1,024 ways to flip the signs
H1 · selection against the AI asked directly
the lineage's champion higherthe AI alone's champion highermean of the runs in, +0.00015
era-1-1
era-1-2
era-1-3
era-1-4
era-1-5
era-1-6
era-1-7
era-1-8
era-1-9
era-1-10
−0.0010−0.00050+0.0005+0.0010
← the AI alone's champion higherthe lineage's champion higher →
H2 · does the model add to selection?
the lineage's champion higherthe blind rules' champion highermean of the runs in, −0.00030
era-1-1
era-1-2
era-1-3
era-1-4
era-1-5
era-1-6
era-1-7
era-1-8
era-1-9
era-1-10
−0.00050−0.000250+0.00025+0.00050
← the blind rules' champion higherthe lineage's champion higher →
Plate I. Both comparisons, run by run, in each of the ten runs named in advance. Above, dᵢ is the lineage champion's panel score minus the AI alone's, the primary endpoint (median +0.00022); below, minus the blind rules' (median −0.00036). A lane with no mark, hatched and marked ∅, is a run not yet in.
Plate II

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 design.

lineagethe AI aloneblind rulesthe founder, designated (chosen)the ceiling (a guide)
era-1-1
era-1-2
era-1-3
era-1-4
era-1-5
era-1-6
era-1-7
era-1-8
era-1-9
era-1-10
era-1-pilotnever tested
0.20440.20460.20480.20500.20520.2054
score on the frozen panel (mean sum of radii, normalized), higher is better
Plate II. Each arm's final champion on the frozen panel, per run. The pilot, when it is in, is drawn below the line and is never tested.
The tests, fixed in advance

Four hypotheses, each tested once.

H1 and H2 use era 0's exact test: the observed sum of the ten differences against all 1,024 ways of flipping their signs. H3 and H4 are existence claims: one champion above the ceiling, certified exactly by both deciders, is proof that a lineage left its founder's design.

Hypothesisdifferencemeanmedianp, exact95% intervalverdict
H1 The lineage's champion scores higher on the panel than the AI alone's, run for run, at equal tokens.panel(lineage) − panel(AI alone)+0.00015+0.000229/1280.0703−0.00007 to +0.00035not supported
H2 The lineage's champion scores higher on the panel than the blind rules' champion, run for run.panel(lineage) − panel(rules)−0.00030−0.000361023/10240.9990−0.00039 to −0.00019not supportedHolm
H3 A lineage champion crosses the ceiling of the founder's design.panel(lineage) > 0.205160 of 10 runsnone crossednot supportedexistence
H4 An AI-alone champion crosses the same ceiling.panel(AI alone) > 0.205160 of 10 runsnone crossednot supportedexistence
What evolution did, in era-1-1

The champion, generation by generation.

The run drawn here is the first in by name. The lineage's champion, lineage-67-15, is 1 inherited edits from the founder; every AI-alone child is one answer about the founder.

lineagethe AI aloneblind rulesthe founder, designated (chosen)
0.222500.222750.223000.223250.22350
rules 0.22327lineage 0.22286the AI alone 0.22256
0255075100
Plate III. Each arm's champion after every generation, by training score. A step is a new champion.
The lineage's champion, edit by edit
  1. gen 0
    founderera 0's best champion
    training 0.22256
  2. gen 67
    lineage-67-15point mutation, line 11 of main.py
    - MARGIN = 0.005 # each radius is 90% of half its cell + MARGIN = 0.003 # each radius is 90% of half its cell
    training 0.22286
Plate IV · descriptive

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 ceiling of the founder's design is 0.20516.

Open the Museum

keptchanged since the foundernew: not in the founderlost: a ghost
decideda champion's frame: in the record, judged exactlychosenthe founder's frame: designated, not selected
the founderfounderscore 0.17995
the AI alone's championalone-2score 0.17995
the blind rules' championrules-94-15score 0.18052
lineage's championlineage-67-15score 0.18019
Plate IV. From era-1-1.
Plate V

How far the births got.

Counted over the ten confirmatory runs in. The commonest cause of a stillbirth: lineage, another shape than the operator's (1,359 of 2,164); rules, no change (6,356 of 6,472); the AI alone, unterminated code block (1 of 1).

liveddiedstillborn
lineage
68.6% lived · 17.8% died · 13.5% stillborn
rules
48.4% lived · 11.2% died · 40.5% stillborn
the AI alone
97.8% lived · 2.2% died · 0.0% stillborn
Plate V. The share of each arm's births that lived, died or were stillborn. The AI alone's births are as many as its tokens bought.
armbirthsliveddiedstillbornmodel tokensfuel
lineage16,00010,9812,8552,16432,897,6102.30 × 10¹³
rules16,0007,7361,7926,47201.55 × 10¹³
the AI alone12,78512,508276132,908,3803.07 × 10¹³
After era 1

What era 2 should change.

Two eras have now asked whether the model adds anything to selection, and blind rules have done as well both times. Era 2 is being designed around that result. These are proposals; nothing below is fixed until its protocol is written, hashed and timestamped before generation one.

  • H2 first. The model against blind rules becomes the primary test, with the number of runs set by a power calculation from the differences eras 0 and 1 measured, not by habit.
  • Edits counted apart from heredity. Era 0's Aleator was one edit from its founder by construction, so its H1 was expected. A control given as many edits as a lineage's champion carries, with no selection between them, would separate what heredity does from how many edits were made.
  • The model as a factor. The same era with a stronger model beside the pinned 7-billion-parameter one, so the record can say whether a better mutagen changes the answer.
  • Replay the tape. Named lineages re-run from a past generation with new seeds, to measure how much of an outcome was chance.
  • Predictions in public. Forecasts of the result opened to anyone before generation one, and scored when it is in.
The runs

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, and every run is replayed with one model answer in ten asked again.

runstaterecordsheadhead timestamp
era-1-1complete11,95595a16ae70c6be4d2…timestamped
era-1-2complete12,32305743e9148aacade…timestamped
era-1-3complete11,97088e8acb1b5e7e226…timestamped
era-1-4complete11,8615475619b21dbaa6c…timestamped
era-1-5complete11,822d174123d8b073a3b…timestamped
era-1-6complete11,937a81c1b91d9735cbb…timestamped
era-1-7complete11,92342e95db23d079f92…timestamped
era-1-8complete11,92613a068c788d7a9a5…timestamped
era-1-9complete11,776c18524b7195ee384…timestamped
era-1-10complete11,91222565d981d0469cb…timestamped
era-1-pilotcomplete11,854ef5066ebe494f81b…timestamped