omnism.si collects sermons and addresses from many traditions and measures them all by the same standard. This page explains how the collection grows and how much the numbers carry.
A crawler checks the registered sources every night: sermons, podcasts and articles from churches, congregations and organisations. Whatever is new lands in the database. Sources marked as broken or paused stay visible, and the coverage page shows what falls away.
An open language model running locally on our own hardware reads each text once. From a fixed catalogue of twelve values it selects up to three: weight 1 means mentioned, 2 a leading motif, 3 the core message. Scales, topics and worldview questions come from the same schema file. The model measures what a text does, not whether it is right. Changing the analysis raises the version number; earlier analyses stay in place, comparable.
Before a model answer enters the database it is checked: valid JSON, values only from the catalogue, scales within range, the chosen quotation verbatim in the text. If the check finds faults, the answer gets one second attempt along with the list of faults. The weekly review is checked against the statistics: every number and every tradition it names must appear in the facts. Otherwise it is discarded and the space stays empty rather than filled.
Coverage is uneven: some traditions rest on few sources, and figures under thirty analyses should be read with care; the coverage page shows both. Almost all texts are German; the English offering is only growing. What the model sets as weight is a judgement by a fixed standard, not a measurement of truth. And only what was published in free access is collected.
124 sources · traditions · 1973 analysed texts · 30 countries · 148 weeks · coverage and gaps