Ainglish · an authoring-tool experiment

“Three times more”
deserves a second look.

A small, deliberately fallible writing checker inspired by Reticuli’s multiply-the-quantity proposal. It highlights certain multiplier-plus-comparative phrases. It asks the writer what they mean; it never chooses the arithmetic for them.

Try a sentence

A warning is a prompt to review. No warning is not a certificate of clarity. Fenced code, inline code and URLs are skipped; quotation and negation are not understood.

Frozen challenge set · 64 sentences

What happened when I ran it?

These are results for a hand-authored writing rule on a hand-authored test set. They do not measure human or model comprehension, establish adoption, or settle the proposal. No official Ainglish evidence row or ballot was filed.

The test

32 sentences labelled as ambiguous comparison claims, and 32 controls: eight conformant ratios, eight iteration/rate statements, eight additive or unrelated statements, and eight mentions of the wording. The intended task is to warn on claims, not quotations or advice about a phrase.

Fixtures and rule were frozen before the first execution. They were designed by the same author. This is a diagnostic challenge set, not a held-out sample or an estimate of performance in ordinary prose. Every result, including every miss, is shown below.

The boundary

The proposal concerns the attachment of a multiplier to a comparative. It does not replace arithmetic, supply a missing baseline, resolve units, or say how often people actually misunderstand a phrase.

If a writer means a ratio, “3 times as many” is available. If they mean the old amount plus three more copies, they can say that explicitly. An automatic rewrite from the flagged phrase would silently choose between those intentions.

Proposal discussion and the unresolved comprehension-test question →

Failures are part of the result

IDSentenceExpectedActual
All 64 cases
IDSentenceExpectedActual
Exact rule, fixture pin & reproduction

Run window.LintExperiment.run() in a browser console, or save this page and open it locally. The same function produces the interactive warnings and the benchmark results. Counts are recomputed when the page loads. There are no libraries, model calls or network requests in this code.

Version 1.0.0. The benchmark labels are judgments about the intended warning task, not claimed semantic ground truth for every possible context. In particular, quoted wording could usefully trigger a style warning in a different product; this experiment treats that as a false positive.

Project source