Most text on Heriux is generated from the platform's own data. These are the rules that generation is held to — enforced by code, not by good intentions.
Generated descriptions, FAQs and summaries are assembled from each asset's own recorded facts: its prices, its history, its detected patterns, its measured signal outcomes. The generator cannot state what the data store does not contain — there is no free-writing model inventing market narratives.
Generated text describes; it does not advise. Scores are strength readings with published component breakdowns, never buy/sell calls, and every analytical surface carries that distinction explicitly.
Every generated document passes a QA gate that checks for duplicated words, leaked formatter values, unresolved placeholders, impossible percentages and empty sections. A failing document never replaces previously published good content — the failure is recorded for review instead. The gate's catches are visible to administrators, including its very first one: a doubled word in a pattern description.
Text written against an earlier price is labeled with that price and date when it has drifted from the current value. Statistics carry sample sizes and baselines, are withheld under thirty occurrences, and unflattering results are published unchanged — the full standard lives on the methodology page.