The Overflow Lab

The four parts, and why they are not strategies

These are not four kinds of person. They are four parts every person has, in different proportions, in different weeks. The agents here are pure types only because a model needs handles — the interesting version is the mix inside one life.

The Reciprocator is not a quadrant — it is the scaffolding. Keeping a ledger is what you do while trust is still being established. Reciprocity is what overflow looks like before anyone trusts anyone; overflow is what reciprocity becomes when the ledger stops being necessary. It is in the model because the literature says it should win, and overflow needs a real opponent rather than a straw one.

Sprint 1 log — what broke, and the one thing that held

The three failures

• Build 1 — the world was too generous. Baseline production exceeded metabolism, so nobody died and everyone hit the energy ceiling. No scarcity means giving buys nothing and hoarding costs nothing, so no strategy could be distinguished from any other. A model without scarcity cannot say anything about overflow.
• Build 2 — the reproduction rule decided the outcome. Agents reproduced on accumulated energy. Overflowers give away everything above their threshold, so they could never reach it. Overflow was structurally forbidden from propagating. The result looked like a finding about overflow; it was a finding about my rule. Replaced with cultural imitation — which is what the book is actually about, since it argues about orientations people pick, not genes.
• Build 3 — the wider success signal was a constant. "How is the whole doing" was computed as a global average, identical for every agent, so it could not discriminate between them and both signals collapsed into "who has the most." Now measured over each candidate's own neighbourhood. And without an exploration term the imitation dynamic drifts to fixation, so a single run reports noise — mutation added, but replication with a seeded generator is still missing.

The one result that survived every build

Duty-givers go extinct in every condition tested — all values of non-rival share, both success signals, every starting mix, in all three builds. They give on schedule regardless of their own state, and under a rationed commons that is fatal. Hoarders survive them; so does everyone else.

A gift given from depletion costs more than it delivers.

That is H3 from the design document, and it is the prediction that most directly concerns a reader's actual life. It is also the least surprising, which is worth saying out loud.

What sprint 2 needs

• A seeded random generator and results averaged over many runs. Single runs of an imitation model are noise, and reporting one is how you fool yourself.
• Weaker imitation relative to selection, so composition tracks survival rather than drifting to whoever got lucky first.
• A larger population. 260 agents is small enough that drift outruns selection.
• Then, and only then, the non-rival crossover and the density threshold — the two numbers the chapter actually wants.