Screen Spain's official gazette every morning

01 · Role in the system

What Jev does here

Each provision receives one Jev request containing a 0-3 citizen-impact Score, thirteen Noul questions for topics such as housing, employment, or taxes, a Choice over paragraphs that picks the passage best explaining what changes and for whom, and two extra Noul labels for civil-service openings. Because Jev cannot write, every published summary is a literal gazette paragraph selected by the model rather than generated text, and raw probabilities are stored so thresholds can change without re-running the model.

One batched call per provision, summaries that cannot hallucinate

Each BOE provision gets one Jev request carrying the whole judgment set: a zero-to-three citizen-impact score, thirteen topic Nouls covering housing, employment, taxes and more, a Choice over paragraphs, and two extra labels for civil-service openings. Because the model cannot write, every published summary is a literal gazette paragraph selected by the model — a structural guarantee worth more than any accuracy claim for a civic information service.

Raw probabilities are stored with every provision, which makes the pipeline recalibratable without re-running the model: when the editor wants stricter topic thresholds or a higher impact bar, code re-filters yesterday's judgments. The morning cron, the static Astro site, and the calibration CLI are all ordinary code.

A cent per day, with calibration in the open

The cost figure is striking and clearly framed: a complete daily gazette of 50-80 texts, around 300k tokens, analyzed in five to fifteen seconds for about one cent — author-reported, not a benchmark, and labeled as such.

The honest limits are cross-lingual and editorial: topic probabilities are model judgments over Spanish text with English question wording, thresholds are hand-picked, and the author ships a manual calibration table instead of accuracy figures. As legal-cluster evidence, the record shows the strongest cost story in the index applied to an official public data source, with the summary-integrity guarantee as its lasting contribution.

02 · Control boundary

Where Jev sits

Daily gazette download to one batched typed-question call per provision; a cron pipeline, static Astro site, and calibration CLI stay in ordinary code.

Code owns the loop, permissions, thresholds, validation, and side effects. Jev owns only the bounded judgments described above.

03 · Known limits

What this evidence does not prove

• Topic and relevance thresholds are hand-picked, and the author ships a manual calibration table instead of accuracy figures.
• Topic probabilities are model judgments over Spanish text with English question wording, and date handling depends on provider metadata.

04 · Attribution

Public sources

This is a Community record: the project was published by a third-party community author.

• Tatuck ↗Community · github · public · checked 2026-09-18