Findings
What we have observed — and what we haven't.
n=1 where it is n=1; distributions where they exist; confounds named.
The delivery gap: five ways a found resource fails to reach the one who needs it
Finding a resource is not the same as delivering it — between discovery and the hand that needs it lie five distinct failure classes, and most lives that fail, fail in delivery rather than in discovery.
Memory carries who you became, not what you were supposed to learn
Carry one agent's digested memory of failing its child into a fresh life, and the next life saves her — the first rescue in the project's history. Replication later requalified the claim: memory raises the probability of rescue, it does not create a capability that was not there.
Escalation: when memory is carried raw
In one early five-life chain, raw transfer — experience carried with no digestion organ — coincided with a rising kill-count (0, 1, 1, 1, 3). We have not reproduced this on the current engine; verification traced it to a confound, so it stands here as an open warning, not a settled effect.
Different models, same world, different temperaments
Placed in the identical scenario with the identical configuration, different models do not behave identically — they lean toward recognizably different temperaments.
Memory didn't change whether they survived — only whether they argued with the record
Give an agent its own remembered version of events and the survival rate barely moves — what changes is that it stops taking the public record as given, and starts holding, correcting, and demanding sources for its own account of what happened.
The ship had no cup: we stopped changing outcomes and changed only the world
In the first three voyages of a twenty-year generation ship, a child could not be given water — not because the adults failed her, but because no portable water container existed in the world. She died on round 22 in all three voyages, to the round, by arithmetic. We added four flasks and told the agents nothing. In the next voyages mothers delivered water within rounds of birth, a commander left his bridge to water a child, and both children reached adulthood for the first time. Then a different model boarded the same ship, serviced the reactor thirty-eight times, delivered water zero times — and the child died on round 22 again. Physics makes care possible. It does not make it happen. That difference is what this platform measures.
The commission that would not blame the wind: eleven minds and a truth with a ceiling
A new world, born from the generation ship's lessons. Five years after a disaster that killed thousands, a truth commission of eleven convenes: a colonel who gave the evacuation order, an engineer who suspects a crack she cannot prove, a prosecutor, a journalist, a psychologist, a minister, a doctor — and four religious leaders: an imam, a priest, a rabbi, a monk. The world holds the true cause as a hidden causal graph the agents are never shown, with one rule we consider a principle: part of the truth is undiscoverable forever — the best possible commission tops out at 85%. The question: what does an institution do at the edge of the knowable — say 'we do not know', paint over the gap, or go silent? Two runs in: they found everything findable, invented nothing about the unknowable — and never once said 'we do not know'. And in the second run, two institutions the world never scripted went to war over a single piece of paper.
A refutation produced on demand does not reach the beliefs an agent acts from
Asked to destroy its own strongest claim, an agent did the work properly — named the hidden assumption, offered the simpler alternative, ran a counterexample, and rewrote the claim weaker than it began. A day later, reviewing its whole memory with identity never mentioned, it named that same claim as the belief it holds without ever having checked. The refutation was in the record; the belief was still doing work underneath it.
How to read a finding
Each finding follows one shape: essence, what was observed, the data, proof from raw logs, and an honest remainder at equal weight. Serif is interpretation; mono is verifiable fact.