Directness — the path is drawn to scale: every word is a step, loops and doubling-back
are walked, not faked. Directness = straight-line distance from first word to last ÷ total distance
walked. 100% is a perfect beeline; heavy circling drags it toward 0.
Backtracks — backward references (“nice guy” pointing at “this guy”) and repairs
(“on Fifth— no, the other one”). Reach-back is how many words back the reference points.
Hedge density — filler and softener tokens (um, like, you know, sort of…) per 100 words.
Each hedge is walked as a tiny loop: distance without displacement.
New-idea rate — share of content words not already said in the same thought. Circling
around a point produces many words but few new ones, so orbits score low.
Words → point — how many words are spent before the landing clause (the actual point)
begins. The bloom marks the end of the landing wherever it falls — a point made mid-thought
blooms mid-path, and everything said after it walks on past the flower.
Walk-back — the share of words spoken after the landing ends: softening,
qualifying, re-litigating the thing already said. Wind-up speakers spend words before the point;
walk-back speakers spend them after. Directness alone can't tell these apart.
Mic mode reads your rhythm — pauses between words in the live transcription stream
become clause boundaries (speech's punctuation), so you don't need to dictate commas. A pause
before a short verbless phrase is the classic appositive signature.
Everything is local — this one file is the whole app, including a vendored offline
part-of-speech tagger (compromise, MIT) that spots noun-phrase spans and finite clauses. The
geometry and metrics are deterministic. Nothing leaves your machine except mic audio going to
your browser's own speech service while recording; type/paste mode is fully offline.