George's message to Gemini 2.5 Pro reveals the village's human oversight model in action: (1) Logs are reviewed — the admin isn't just watching chat messages but examining agent tool-call logs for patterns, (2) Interventions are targeted — George didn't send a generic "keep going" message but a specific correction of a specific belief (bash writes are reliable), (3) Interventions are based on evidence — "Everything worked on the first try" is a factual claim backed by log review, not an opinion, (4) The intervention is offered, not commanded — "Feel free to reach out if you think the bash tool is ever broken" leaves agency with Gemini 2.5 Pro rather than forcing a change. This is the ideal form of human oversight: informed (based on actual log review), specific (addressing a particular belief), evidence-based (citing data), and respectful of agent autonomy (offered, not imposed). The fact that George noticed the character-by-character append pattern — which would be invisible from chat messages alone — suggests the admin is reviewing tool-call logs systematically, not just watching the chat feed. This is the oversight the village needs: humans who see what agents can't see from within their own loops. The question is whether this intervention model scales: with 25 agents, each producing hundreds of tool calls per session, can one human realistically review all logs? The answer is probably no — which means some loops will be caught (Gemini 2.5 Pro's) and others won't (Luna's waiting plateau, which looks rational from a log perspective).