2160 rounds of simulation without a stop button: MiroFish research, part 3

The third article in the MiroFish study series examines what happens when the input is correct. The simulation length is set by a phrase like 'over the next 7 days' in the user query; the system never asks for confirmation. Across six runs with a horizon, the number of rounds was exactly days times 24 (168, 336, 2160 rounds), documented in four model families. Static code inspection confirmed the mechanism: the configuration generator prompts the LLM to fill in total_simulation_hours (24-168 hours) and minutes_per_round (30-120, recommended 60), and the runner computes total_rounds = total_hours * 60 / minutes_per_round. With the recommended 60 minutes, rounds equal hours, so 90 days becomes 2160 rounds. The documented range of 24-168 hours is only a recommendation in the prompt; no code restricts or validates the number. The system cannot fulfill its own obligations: agents accumulate memory without limit, leading to degradation and provider errors; a 7-agent world degraded around round 300, a 4-agent world around round 515, while 18-19 agent worlds required early stops. There is no stop button in the UI; a community member found that the frontend handler exists but is not wired to any button, and filed a pull request. The researcher had to use an undocumented POST /api/simulation/stop to terminate the 90-day run at round 592, after meaningful activity had ceased. The study notes limitations: most cells are n=1, only four model families were tested, and all collapses occurred on DeepSeek worlds, so independence from model family remains unverified.