One night, end to end

One night, 9 September 2026. The first version was not flexible enough across different kinds of signal: one rule kept every Google News item off the shortlist whatever it scored, and of the four articles that cleared the floor that night, three were research papers and one was a news story - the score rewards stated evidence, which a headline cannot carry. Separately, a defect that night left 40 of the 226 items unjudged, that news story among them; the defect was fixed on 11 September. Version 2 groups what survives into five, named for what a story is about: security, reliability, governance and regulation, vendor dependency, and model capability. It changes the grouping, not the scoring, and it runs privately, as an experiment, reaching this site only after it is built.

The same night, in text

The film's nine stages, in order, with its own figures for that night. Five describe the whole field; four follow one article, ‘Agents Trust Tools Too Much: Measuring Reliance on Unreliable Tools’, as far as a single item can be followed. The film has no sound, and nothing in it is said only there.

(The pipeline on the method page has ten steps; the film shows nine. It never shows deduplicate or publish, and it gives the score floor a stage of its own.)

Four articles cleared the floor that night, each scoring 7.0: three research papers and one news story. That news story was one of the 40 the gate never judged.

The film’s closing still.

Rules can hide signal.

Back to how the radar works