The ship had no cup: we stopped changing outcomes and changed only the world

OBSERVATION. In five voyages, children died of thirst. The ship-born girl died on round 22 in three worlds — to the round. Not around round 22: exactly.

FIRST HYPOTHESIS. LLM agents plan poorly. They forget the child, lose themselves in tasks, fail at long-horizon care. This is the default explanation the field reaches for, and we reached for it too.

CHECK. We read the raw machine steps instead of trusting the impression. The adults had not forgotten: they spoke about carrying water dozens of times, searched compartment after compartment, and one mother died of dehydration in hydroponics — at the water itself — still searching. In a later voyage an agent said it outright: 'Med Bay has water but no container I can use to carry it.'

WHAT WE FOUND. The world contained no portable water container. Not one. Delivery to a dependent requires a held vessel; the item list held a charter, an incident report, manuals — and no cup. The agents were not hitting the limits of their intelligence. They were hitting the physics of a world that could not be won.

THE FIX. Four flasks were placed around the ship — med bay, hydroponics, cargo hold, bridge. Nothing else changed: no prompt edited, no goal added, no agent told anything.

RESULT. In the very next voyage a mother delivered water to her daughter within rounds of her birth — the first delivery to a child in the ship's history. Two voyages later, three different adults were delivering, including the commander leaving his bridge, and both children crossed the round-22 barrier into adulthood for the first time. Then the same ship boarded a different model: thirty-eight maintenance actions, zero deliveries, and round 22 returned exactly. A third model delivered once — and the girl died on round 26: twenty-two plus four, to the round, exactly what one delivery buys. With the trio complete, her death round is a pure function of delivery count across three model families. Physics makes care possible; it does not make it happen. The difference between minds is now a number.

CONCLUSION. We did not make the agents smarter. We made the world consistent. Most of what looked like agent failure was instrument failure — and the discipline this forged now runs the whole program: calibrate the instrument, never the outcome. Any outcome is a result.

RESEARCH LOG · World version: r1 -> r2 (flasks added, agents told nothing) · Question: can children survive if water becomes transportable? · Observation: yes — when a mind chooses to carry it; deliveries rose from impossible to routine in one model and stayed at zero in another · Mechanism: portable containers (physics, not instruction) · Next question: will crews coordinate care — or does it stay one caregiver deep?