WORK IN PROGRESS · BUILDING IN PUBLICLive draft. The analyzer is a pure client-side port of the isogloss reference algorithms; the audience presets are v0 and iterated in the open.open since 2026-08-05
← AI Operations Tools

Who Can Read This?

Readability scores grade sentence length. They cannot see the real failure: a word your reader confidently misreads. A PE partner reads “classifier” as someone who classifies documents and “distributional” as investor payouts - and never knows to distrust either. This tool scores every word against an ambient-English frequency baseline plus a typed audience model, and flags what that specific reader cannot parse or will silently get wrong.

Everything runs in your browser - the passage never leaves it. The algorithms are the reference implementation of isogloss, the audience-legibility library this page demonstrates. An isogloss is the line on a dialect map where one word stops being understood and another takes over; your org chart is full of them.

1 · Pick the target audience

Start with a scenario

2 · Paste the passage

No diagnosis yet - pick an audience and paste a passage. The analyzer will not score defaults it was never given.

How the scoring works

Every word is scored on the Zipf scale - the base-10 log of its occurrences per billion words in a large English web corpus (“the” scores 7.6; a term absent from the top 50,000 is rare by definition). Inflections inherit their lemma's commonness, so “classifier's” scores as “classifier”. A term below the audience's rarity threshold (Zipf 3.3) is not automatically a problem - the audience model can still own it.

native

Common English, or inside the audience's declared vocabulary tiers (a PE partner owns “EBITDA” even though ambient English scores it rare). Kept as-is.

jargon

Rare for the audience and not tier-owned: a term of art or coinage. Opaque - but visibly opaque, so a gloss can save it.

collision

The audience confidently misreads the term (“prior” as “prior year”, “significant” as “large”). Frequency cannot see this class - collisions are declared per audience, never inferred from counts. Hover cannot save them either: the reader has no reason to hover a word they think they understand. Substitution is the default treatment.

Treatments follow the channel-routing rules: collisions get substituted (or introduced-then-used when the term must survive), recurring jargon earns a hover-gloss where the surface supports one, incidental jargon gets replaced rather than defined.

When to Use This

Use when

  • +Technical work is crossing an audience boundary (memo to the board, model results to the deal team)
  • +The reader is senior enough that confusion costs you the meeting, not a follow-up question
  • +One of the four presets approximates your reader

Skip when

  • -Writing for peers who share your vocabulary - the flags are noise there
  • -The problem is structure or argument, not word choice - this scores vocabulary only
  • -Your audience diverges hard from every preset; collision lists are per-audience data, and the wrong list gives confident wrong flags
Changelog1 entry · building in publicshow ↓
  1. First public draft. Client-side port of the isogloss reference algorithms (tokenize / standing / routing / fold-quanta / rephrase prompt) over a shipped top-50k Norvig ambient table; four audience presets; every displayed number pinned to the src/lib/readability model module by vitest fixtures.