From paper to story.

Starting point

The run did not begin by asking an AI to “write a story.” It began with eight scientific candidates already verified in the reserve. The selected work was arXiv:2603.02223v1, a study of wildfire evacuation behaviour based on 853 valid responses.

Its fertile observation was not that a model could predict everything, but precisely its asymmetry: evacuation mode was more predictable than departure timing. That difference became the story's causal constraint. Cauce can prepare vehicles according to how a population tends to leave, but it does not know when each household will decide to go.

The paper does not validate the autonomous fleet, the villages, or Berta and Maia's decisions. Those elements are fiction. The extrapolation nevertheless preserves the limit that makes the conflict necessary: using a useful prediction without pretending that it resolves the uncertainty the study does not resolve.

Six directions, one selected

• Nobody Leaves on Schedule — selected. A bus assigned to a carless village leaves another settlement where eight workers are absent from the registry. It prevailed because of its immediate physical friction, legitimate opposition, and a decision that did not depend on explaining the model.
• The Fifth Statement — finalist. An auditor corrects an official evacuation reconstruction inside an urban twin. It passed the prose probe, but required more exposition and risked repeating testimony without enough material change.
• Everyone's Keys — discarded. The temporary requisition of autonomous cars shifted the centre of gravity from the paper toward a property dispute.
• Herd Six — discarded. Livestock drones offered strong imagery, but required too much supporting infrastructure and could reduce the story to logistics.
• Three Colours for One Street — discarded. Its choral form was promising but too broad for the intended length.
• Those Who Always Wait — discarded. An insurance penalty drifted toward contractual negotiation and resembled recent structures in the archive.

Discarding did not mean declaring those ideas bad. It meant selecting the one that expressed the science most clearly, produced visible opposition, and differed most from the recent catalogue.

Two probes before the draft

Only the lead and reserve concepts reached prose. Each produced a short scene to test three things before investing in a complete story: whether somebody wanted something concrete, whether another legitimate force could prevent it, and whether the ending created a precise reason to continue reading.

Both passed. The audit finalist depended on reconstructing the past while explaining the twin, the registry, and the commission. The bus scene instead began with an irreversible action: Berta pulls the red cable and stops a vehicle that must leave. The choice was made on dramatic economy, not thematic preference.

What changed in review

The first complete story received the verdict APPROVED_WITH_CHANGES. Review found no critical failure in science, causality, opposition, or ending, but it did identify a formal repetition: a series of numerical warnings was turning the pressure into a countdown too similar to a withdrawn story in the archive.

The complete rewrite made four material changes:

• removed the numerical countdown and replaced it with qualitative changes to the escape corridor;
• explained why the eight workers were missing from the registry through the repeater cabins, contractor badges, and annual update;
• strengthened Maia as legitimate opposition: she protects twenty-two registered residents in Vega Alta and risks her own job;
• preserved the red cable, broken card, first vehicle's failure, and final crossing because they planted and paid off the central conflict.

There was no return to design and no switch to the reserve candidate. There was one complete rewrite before the Spanish version closed.

What fell away

Numbers, interface messages, and explanations that exposed procedure without changing decisions were removed. The eight workers' administrative absence was not turned into a speech: it remains a material problem while they walk, arrive, and occupy the seats the system failed to count.

The final synthesis retains one causal line: an AI knows how each community evacuates better than when it will do so; that capacity allocates a resource correctly in statistical terms while leaving one concrete exception unanswered. Berta and Maia must complete the model with human responsibility, and both pay for doing so.

Gates that allowed publication

• CF1: the public AI, driverless bus, and empty cars returning appear in the first quarter and cause observable decisions.
• CC1: a reader without the paper can reconstruct novum, objective, decision, and cost.
• T1: Berta and Maia protect different groups, hold real power over the bus, and lose professional authority.
• B2: the probe ends with an immobilised vehicle, a broken credential, and an alternative still uncertain.
• J1: every change removes an option; the final crossing pays off the vehicles planted at the beginning.
• L1: the blind reading found a concrete expectation at 25%, 50%, and 75%, plus an identifiable gratification at the end.
• Bilingual QA: the English version preserves facts, capabilities, figures, opposition, losses, ending, and scientific limits.

These gates record editorial evidence. They do not turn literary taste into an objective measurement or guarantee that every reader will respond in the same way.

Boundary of this case file

This document is a sanitised public reconstruction based on artifacts recorded during the run. It publishes decisions, alternatives, changes, and gate outcomes. It does not publish complete drafts, internal instructions, private reasoning, or sensitive operational information. Useful transparency makes the process auditable without turning internal work into spectacle or confusing traceability with raw thought.