Five measurements, stated before the run, with their readings listed in advance; none has a threshold called success. Each is read by its own rule and nothing else.
The instrument was exact on every pre-registered check: all 1,319 endpoints reproduced byte for byte on a second machine before any neighbourhood was evaluated, all 141 repeated neighbourhoods agreed exactly, and no placement failed.
X1 · neighbourhood divergence with neutral distance
The dissimilarity between an endpoint's one-mutant neighbourhood and its parent's, measured coarsely on the five classes (D5) and finely on exact signatures (Dsig), by distance target 8, 16, 32, 48, 60. Parents are the unit of inference: three walks summarised within parent by their median, uncertainty from resampling parents. No directional threshold. Pre-listed readings: (a) not distinguishable from zero at every target; (b) above zero at the first target and flat afterwards; (c) rising with neutral travel; (d) at the matched reference by target 60.
Dsig by target (mean over parents of the within-parent median): 8: 0.0013, 16: 0.0026, 32: 0.0046, 48: 0.0065, 60: 0.0077; Spearman's rho over parent medians against target 0.79 (parent bootstrap 0.76 to 0.83). D5 by target: 8: 0.0002, 16: 0.0004, 32: 0.0008, 48: 0.0012, 60: 0.0015; rho 0.64 (0.58 to 0.69). Against realised Hamming, rho 0.74; against Hamming inside the parent's expressed region, 0.52. References: unrelated parents with the identical full signature, Dsig 0.009 (0.008 to 0.010) over 60 parents; unrelated parents matched on lifespan only, 0.402 (0.304 to 0.500).
(c) rising with neutral travel, and still below the matched reference at target 60 Dsig is a lower bound on phenotypic difference: two children with the same signature may differ in ways the signature does not record.
X2 · accessibility without phenotypic change
For each endpoint, the set of signatures reachable in one mutation against its parent's: Jaccard(S) and Dsig. Decision rule fixed in advance: an endpoint's neighbourhood differs from its parent's if Dsig > 0 or Jaccard(S) < 1, provided the reproducibility check passed with every replicate distance exactly zero. The fraction differing, by target, with a parent-level bootstrap. No threshold on the size of the difference.
Reproducibility check: 141 of 141 re-evaluated neighbourhoods agreed exactly, so the rule applies as written. Endpoints whose neighbourhood differs from the parent's: at 8, 100.0% (100.0% to 100.0%); at 16, 100.0% (100.0% to 100.0%); at 32, 100.0% (100.0% to 100.0%); at 48, 100.0% (100.0% to 100.0%); at 60, 100.0% (100.0% to 100.0%). Jaccard(S) by target: 8: 0.982, 16: 0.963, 32: 0.932, 48: 0.902, 60: 0.870.
every endpoint differs from its parent, at every target The rule is sensitive by construction: one differing child among 16,319 is enough, so this fraction records that a difference exists, not how large it is; the size is X1's result.
X3 · accessible variation
Three quantities per genome, as paired differences endpoint minus parent, by target: adjacent variation (the fraction of children whose signature differs from the genome's), phenotypic richness (distinct child signatures), upward accessibility (the fraction of children that outlive the genome on probe A). Pre-listed readings: unchanged; changed in one direction across parents; changed without a consistent direction. No threshold.
Adjacent variation, in percentage points: at 8, -0.003 (-0.009 to +0.000); at 16, -0.008 (-0.017 to -0.001); at 32, +0.002 (-0.007 to +0.011); at 48, +0.004 (-0.012 to +0.020); at 60, -0.005 (-0.024 to +0.014). Richness, in distinct signatures: at 8, -0.1 (-0.4 to +0.3); at 16, -0.2 (-0.9 to +0.4); at 32, -0.0 (-1.1 to +1.0); at 48, -1.9 (-4.0 to -0.1); at 60, -3.2 (-6.0 to -0.7). Upward accessibility, in percentage points: at 8, -0.002 (-0.007 to +0.001); at 16, -0.003 (-0.012 to +0.004); at 32, -0.012 (-0.034 to +0.008); at 48, -0.003 (-0.025 to +0.019); at 60, +0.005 (-0.017 to +0.027).
adjacent variation approximately unchanged (the interval excludes zero only at target 16, by under a hundredth of a percentage point, with no consistent trajectory); richness down at targets 48, 60; upward accessibility unchanged at every target The word evolvability is not used.
X4 · exact against aggregate
For each endpoint–parent pair, the Jaccard overlap of the exact sets of non-neutral (position, value) substitutions, reported as the difference from its exact random-subset expectation (hypergeometric), beside D5 and Dsig, by target. The question is whether exact mutational overlap declines along the targets while the aggregate accessibility distances stay comparatively small, read beside the matched references. Reported as it appears; no binary classification.
Observed Jaccard(M) by target: 8: 0.996, 16: 0.992, 32: 0.985, 48: 0.971, 60: 0.925. Random-subset expectation for independent sets of the same sizes: about 0.024 at every target. Excess above that expectation: 8: 0.972, 16: 0.968, 32: 0.961, 48: 0.947, 60: 0.901. Observed exact overlap remains far above random expectation after sixty bytes of neutral travel; Jaccard is intersection over union and is not converted into a fraction retained. Beside it, D5 8: 0.0002, 16: 0.0004, 32: 0.0008, 48: 0.0012, 60: 0.0015 and Dsig 8: 0.0013, 16: 0.0026, 32: 0.0046, 48: 0.0065, 60: 0.0077.
a large conserved structure with a smaller component rearranging: observed overlap 0.996 at 8 to 0.925 at 60, about 0.024 by chance; D5 at most 0.0015, Dsig at most 0.0077 A Jaccard(M) far above its expectation means the two programs share their exact fragile positions; near it, they share only as much as strangers would.
X5 · nothing new, observation only
E2's detectors, unchanged, over every child, plus self-inspection (CODESIZE, CODECOPY, ADDRESS, EXTCODEHASH of self executed). Incidence reported. Nothing predicted, nothing rewarded. child_has_code and calls that enter code are replicated three times before being reported as more than candidates.
Children reaching each event, over 1,407 neighbourhoods: self_inspection 2,142,809, clean_halt 270,447, returns_data 219, create_executed 195, write_survives 147. Genomes with a child that left code behind: 0. Genomes with a child whose call entered code: 0.
nothing new The scorecard moves only as E2 §9 specified.
A program's neutral descendants, which still do exactly what it does under the frozen measurement, do not keep exactly its immediate futures: their one-mutant neighbourhoods diverge from its own, by a little at eight bytes and by more at every target after, in step with neutral distance.
The amount of adjacent variation stays approximately constant with the travel: the share of children that behave differently, and the share that outlive the genome, are the same sixty bytes away as at home, while the exact mutations that matter slowly rearrange, their overlap still high at sixty bytes but fallen, a Jaccard of 0.996 at eight bytes and 0.925 at sixty.
Two programs can do the same measured thing without being in the same place: silent history does not add or remove possibility so much as rearrange which possibilities are one mutation away, changing the shape of the next step without changing the present.
To check any number: fetch the snapshot or the per-genome summaries, run scripts/e3_analysis.py, and compare with results.json; to check the picture, run scripts/e3_positions.py, scripts/e3_rows.py and scripts/e3_somewhere.py on the record. The root hash above is the sha256 of the sorted sha256 manifest of every file in the record.
the research edition
SOMEWHERE ELSE (E3) is the research edition for this experiment: one token, the picture above at print resolution, pinned to Arweave with metadata carrying the record root, both anchors, and the hash of the script that drew it. Its rules were fixed and hashed before the first attempt, and the artist saw only a synthetic mock before signing, so the bars are inked by the record's own fractions: in every row, five programs that do the same measured thing, each beneath a bar of the mutations that would change it. Minted by the keeper on the research editions contract, offered here by reserve auction. Proceeds go to the artist and fund the work. It confers no rights over any program, release, receipt or experiment.
Reserve 0.05 ETH. The auction runs 24 hours from the first bid, 10% minimum increments, extended by five minutes for a bid in the last five. Settled by Manifold's marketplace contract; the keeper finalizes it when it ends. Contract 0x55540e5bcd1b0a1e2c00de4ae55ef007bde35664, token 3, listing 20956. Bid here, or on the same listing at Manifold.
The catalogue states facts. The lab records what executed. The readings interpret. This page is the lab's. readings · catalogue · as text: e3.md · llms.txt
