The point of this detector is to measure rhetoric, not prestige. The text analysis step feeds into modular content generation. Flagged bits are excluded as source material, or—when in doubt—a human steps in to verify and fact-check.
During the Times test, we found that our earlier rules were too broad: ordinary narrative touches—perfectly acceptable in mainstream reporting—were being flagged as full-blown “effects.” Rather than hard-coding an exemption for a prestigious outlet, we refined the logic:
More granularity. We introduced flags (yellow cards) for light framing or stylistic color, reserving effects (red cards) for language that crosses a higher rhetorical threshold.
Cue + context requirement. Rules like apples-to-oranges comparisons now trigger only when the sentence contains both a comparison cue (“as … as,” “now as then”) and a categorical or probabilistic claim without hedging or sourcing (“just as likely to fail”).
Guard tokens. Phrases like “analysts say,” “according to data,” or even a single “likely” or “might” automatically soften or downgrade a hit, reflecting genuine uncertainty.
Applied to the NYT paragraph in question:
“The move … is just as likely to fail as it did then.”
It contains both a comparison lexicon and a categorical outcome claim, with no guard tokens. Under the refined rule, that still warrants a red-card effect.
