The result
What it found in Hanko first
Before this study, Hanko's echo detector missed every one of the copies below. Channels append their own boilerplate (an ad, a "News | Markets | YouTube" footer), and every added word dilutes a Jaccard similarity score. A verbatim headline scored 0.53 against the 0.6 line, so one headline counted as two voices. The fix (D-14) also compares containment, which padding cannot dilute, and can only ever demote more voices, never fewer. The regression test is built on two of these real posts.
Every echo found, most similar first
Is the number an artefact of the thresholds?
Re-run at every combination of echo window and similarity threshold. Even at the loosest setting, few events collapse.
What it does not show
- –Rewrites of one press release still count as separate voices. The Swiss spot Bitcoin ETF approval was covered by several of these channels in their own words within hours. For opinions that rule is right; for news it is generous, so echo inflation here is if anything undercounted.
- –Naming the same token is not converging on the same narrative. In a six-hour window, two channels mentioning BTC are usually telling two different stories.
- –Fifteen Telegram news channels over fourteen days, not X influencers. Copy-paste shilling on X may look very different. This measures one corner of the market, and says so.
Re-derive every number: hanko echo-index publish, from the committed study/echo-index/posts.jsonl. The full report, with method, is in the repository.