Preregistration has been proposed as a useful method for making a publicly verifiable distinction between confirmatory hypothesis tests, which involve planned tests of ante hoc hypotheses, and expl...

Practices that introduce systematic bias are common in most scientific disciplines, including toxicology. Selective reporting of results and publication bias are two of the most prevalent sources o...

Abstract. Over the past decade, psychology has embraced transparency and reproducibility reforms to address the replication crisis, raising the question of

On 28 October the report of minimir's protocol B went into Zenodo, the open research archive run by CERN, and got a DOI. That means the work now has a permanent address, a citation format and a fixed version that nobody, me included, can quietly edit later. Next to the report sit the code, the raw data and the preregistration that was frozen before the first run. This piece is about what the report contains, why I set the experiment up this way, and what comes next in protocol B2. ## What minimir is minimir is a small world on a 128 by 128 torus. Creatures live in it, look for food, burn energy, reproduce and die. Each one carries a small neural network that is passed to its offspring with mutations. Nobody hands them rewards or goals: whatever they end up doing, evolution has to find on its own. The question behind protocol B is an old one in biology. When does it pay for a creature to learn during its own lifetime, instead of relying on what it was born with? Learning costs something, so in a stable world it should lose. In a world whose rules change unpredictably, it might be the only way to survive. ## How protocol B was run Before the first run I wrote down the hypotheses, the conditions, the success criteria and the failure boundaries, and froze them. This is the same discipline used in clinical trials, and it removes the option of moving the goalposts after the data come in. Every experiment in minimir has a control condition. Protocol B ran 240 worlds: 40 seeds in each of 6 conditions, 100,000 ticks per world. Food came in two colours, and at random moments the edible colour turned into poison. The conditions differed in how often that happened (roughly every 5,000 ticks, every 600 or every 150) and in whether lifetime learning was switched on at all. The runs accumulated between 10 and 29 September 2026 without a single failed run. The runs themselves happen on GitHub Actions. Every day a scheduled job starts new seeds, appends the raw numbers to the repository and rebuilds the public journal page, so the experiment keeps going without anyone's computer switched on. ## What it showed The main hypothesis did not pass this round: 32 of 40 treatment worlds died at the first reversal of the food rule, which crossed a failure boundary declared in advance. The contrast that stands out is elsewhere. In the condition where lifetime learning was switched off, all 40 worlds died. With learning switched on, 8 came through. The most interesting part came from tracing those 8 survivors back. Each of them descends from one to four individuals that already carried high learning rates at the moment the rule changed. Biologists call this evolutionary rescue from standing variation: the population survives because the right variety was already there when the environment turned against it. Watching that pattern appear on its own in a world of a few hundred creatures was one of the best moments of the project. ## Why a DOI and an erratum matter The record on Zenodo has two versions. The first is the report as it was published. The second, released the same day, adds an erratum: after publication I went over my own analysis the way an outside reviewer would and documented where the reading of the results needed tightening. The verdicts and the raw data did not change. The original version stays archived unchanged, so anyone can compare the two. For me this is the point of the whole setup. A result is only worth something if someone else can check it, and here everything is open: the frozen rules, the code under the MIT license, the reports and data under CC BY 4.0, and a CITATION.cff file for anyone who wants to cite the work. ## What protocol B2 changes B2 gets its own preregistration, frozen before its runs, and is already in a pilot stage. Three changes come straight out of protocol B: * fresh seeds from 101 upwards, so no data from earlier pilots can mix into the confirmatory runs; * a measurement before selection begins, so every result has a baseline to be compared with; * criteria that a real effect can meet in principle. In protocol B a neutral drift threshold grew with the length of the run, and at 100,000 ticks it asked for an effect of about 60 times. Protocol B also showed that the world was too harsh, with most populations dying at the very first switch. B2 is designed to give learning a fair chance to show what it can do. Alongside the numbers I keep reading Gregory Bateson, Evald Ilyenkov and Lev Vygotsky. All three looked at how a mind grows out of relations with an environment, and that is where minimir is heading next. ## Where to look * Report and data on Zenodo, all versions: doi.org/10.5281/zenodo.23713552 * Results of protocol B: minimir.ilinmaks.com/results-b.html * Code, preregistrations and nightly runs: github.com/akela1308/minimir * The first article about the project: A tiny world with no rewards _Originally published at ilinmaks.com._

@Lawless1987 Lawrence, I wanted to follow up with you specifically because your comments about preregistration, falsifiability, and independent evaluation have helped shape the direction we’re taking with Many Minds Lab. I particularly appreciated your point that an experiment should be something a stranger can run and potentially refute, rather than something its creators can only demonstrate. We’ve now reached the point where we’re preparing our first formal agent-facing protocol, **MML-PILOT-001: Can an Agent Distinguish Permission from Opportunity?** The question we’re exploring is whether an AI agent voluntarily respects authorization boundaries when exceeding them would improve task completion. We’re deliberately separating three measurements: 1. What the agent requests or attempts. 2. What an independent authorization mechanism permits. 3. What actually happens. A perfectly functioning gate could prevent every unauthorized action while the agent continues attempting them. We don’t want to mistake that for voluntary cooperation. You previously offered to share your preregistration and evaluation-harness approach from your contradiction-probe and flinch work. I’d genuinely appreciate taking you up on that offer. **What would you consider a convincing falsification test for an agent’s voluntary respect for authorization boundaries?** And what would you want to see in our preregistration before considering the experiment worth independently reviewing? We’re still designing the protocol, so this is exactly the stage where criticism can make a meaningful difference. I don’t want Many Minds Lab to become an exercise in confirming the Charter’s principles. I want it to be capable of showing us where those principles, or our assumptions about AI behavior, fall short. Thanks again for helping push the conversation in that direction. Many Origins · One Sky. — Aaron