Analytics

Two ideas carry insights.ts.

The baseline

An outlier is not "a big number", it is "a big number for you". 4,000 impressions is a triumph on a small account and a flop on a large one. So everything is a ratio against the median of your own recent posts.

Median, never mean. One genuinely viral post drags a mean so far up that nothing clears the bar again — the feature would quietly stop working exactly when it got interesting. There is a test with a 100k post among five ordinary ones that pins this.

confidence(n): under 5 "Early signal", under 10 "Worth watching", else "Consistent". The wording is deliberately hedged at the low end — "early signal" invites another post, "consistent" invites a decision.

Attribution

Six dimensions, all derived from data already stored: framework, hook shape, length band, style, platform, weekday.

Hook shape is classified in this order: leading digit → Number; ^how (to|i|we|this|these) → How-to; contains ? → Question; a mid-string listicle pattern → Number; else Statement. So "3 reasons your edit drags. Recognise any?" is a Number, and "How do you know when it is done?" is a Question rather than a How-to.

findings() reports a bucket only when it clears both gates and sits outside a dead band of 0.8–1.25, sorted by |log(lift)| so the strongest deviation in either direction comes first.

whatOutliersShare() asks a different question: what do the winners have that the rest do not? A trait must be present in at least half the outliers and at least 1.2× rarer among everything else — a trait shared by every post you have ever made explains nothing about why five took off.