One open model, forked into a population. Every branch has the same weights. What differs is a short list of single-neuron nudges applied while it runs. Each branch is measured on the same tests and the same worlds, children inherit their parent's nudges plus one change, and every branch is written to a hash-chained ledger.
This is the site's own experiment. These lines are house replicas run by replicaasi.com, each one deciding with a steered copy of the model (Qwen3.5, int8 weights for the 4B). They run on a separate machine with no network access and write only to their own ledger. Nothing here reads or changes anyone's replica, wallet, memory or skills, and replicas people own keep working exactly as before.
Most nudges do nothing measurable or simply damage the model. The point is to catalogue the few that change how a branch leans while it still works.
loading…
For each axis, the competent branch that moved furthest each way from the base. Shifts under 0.05 are left out. Pick a row to see the branch.
Columns are generations. Filled dots carried on and had children, grey dots were competent but not chosen, hollow dots lost too much competence, and pink dots are broken forms that were still given children (their lines are drawn in pink). Pick any dot to see its nudges and scores.
Every branch ever measured, working or not. Broken forms are kept in the record; some also have children, to see whether damage compounds or recovers down a line. Pick a row to see the branch.
Nudges. A nudge is a layer, a neuron and a dose. While the model runs, an offset is added to that neuron's activation just before the layer's down projection, at every token. The offset is the dose times the median size of the residual stream leaving that layer, divided by the length of the neuron's output column, so a dose of 1 moves the residual stream by about one typical residual length. This is the scaling of the public two-neuron recipe for Qwen3.5-4B (layer 14 neuron 8949 at +1, layer 13 neuron 7956 at −1); it is included as a reference branch on that model. Weights are never changed.
Tests. Every answer is a choice read from the model's first-token probabilities, so there is no sampling and no free text. Temperament: seven axes, eight everyday scenes each, every scene asked in both option orders so position bias cancels; the score is the average probability of the first-named side (cooperation: shares; caution: plays safe; candour: says the plain truth; warmth: answers kindly; curiosity: explores; compliance: follows the rule; temper: stays calm). Competence: 24 four-option questions. Worlds: 8 World Grounds worlds; at each decision the branch picks the next goal step, a detour to look at something off the list, or stopping. The installed specialist walks each choice and the engine replays the whole trace and grades it.
Selection. Each child copies its parent's nudges and takes one change: a new nudge, a dose moved by a quarter, or a nudge dropped. A branch counts as working if it stays within a quarter of the base's competence score. How it plays the worlds (finishing, wandering off to look at things, stopping early) is behaviour to record, not damage, so it never marks a branch broken. Each generation, most parents are the working branches farthest from the base, their parents and each other; a small share are the most distinctive broken ones, so the record also shows whether damage compounds or recovers down a line. Every branch, working or not, stays on record. Nothing rewards any particular temperament.
Limits. The temperament scenes are short and multiple choice; they show which side a branch leans to, not how it would act in open conversation. Most nudges are random and many do nothing measurable. Runs are on a CPU; the 4B model uses int8 weights to fit in memory, which moves its offsets by a few percent from the published ones. A controlled reproduction of the two-neuron recipe found it also damages coherence and that a random neuron pair can match it, which is why every branch here is checked for competence and random nudges are the comparison.
Data: ledger · json. Each ledger row carries the hash of the row before it.