Catch the AI

Muse Wonder wrote the cards and sentences. Your browser checks their numbers.

The cards and sentences are exactly what Muse Wonder wrote: rounds.json. The brief it got is saved here. Only its disclosure line changed, on Nov 3, from "We asked" to "I asked": this is a solo project, and that line was set word for word in the brief.

Muse Wonder, a Meta model, wrote these sentences from each card. I asked it to slip mistakes into some. Six fixed, made-up examples, not an accuracy test.

The checker reads names and numbers. It doesn't read seasons, units or words like rose and fell.

A pass does not make a sentence true. The checker compares numbers with the card, and some checks do not run on this page.

The code is at github.com/macayaven/ikigai-loop.

A paid study uses a separate version of this page, with its own consent. Other visitors send nothing unless they choose to save their tries.

Every word of a row's name must appear in the sentence, in any letter case, so deploy doesn't match deploys and synonyms don't count.

Numbers are read as written: 1,200, 3.5 and twelve work. Dates and clock times aren't read as numbers. In the plain form, a percent sign is ignored.

The number must equal the record. 3.5 and 3.50 are equal.

A number with no record fails.

A sentence that names more than one record isn't checked until it's split, one record per sentence. A sentence like "We shipped 12 deploys and closed 40 tickets" names two records, so split it into two sentences.

In the plain form, words like rose or fell aren't checked.

A sentence with a negation word fails, and a quantity word in place of digits fails: about, over, nearly, half, a couple, twice.

It can't tell whether a sentence is true, only whether its numbers match what you typed.

To see for yourself that nothing is sent unless you say yes: open your browser's network panel, then use the page. After loading, the page makes no more requests until you choose to save your tries, and then one per try.

This is what the full checker covers:

This check covers parseable numbers, cited sources, labels, direction, negation, quantity forms, and units and periods where you declare them. It does not check that a sentence is true. Without labels it cannot tell which metric a number belongs to, and a cited fact is only checked to exist. A check listed as not checked did not run.

What fails now: a claim with a negation word fails, unless it states a zero right before a word of the cited metric's labels and the cited value is 0 (for example "No deploys failed" with labels deploys and failed); quantity words in place of digits fail (about, over, nearly, half, a couple, twice); with a declared unit, a unit word that disagrees fails; with a declared period, a missing or stray period fails, and so does a value beside the wrong period.

What this checker does not check: it never checks that a sentence is true; a word after a number that is not a unit, a magnitude or a percent form isn't checked; disguised spellings of your private terms aren't caught; advice and instructions in the text aren't caught unless you list advice words.

Disclosed limits: the bare word one as a value, period compounds and comparisons with than are not checked; a number-like or quantity-like word inside a name that a cited fact carries is not skipped, which can only make the check stricter; a unit word that agrees with one cited metric can still fail against another, because the sentence does not say which metric the word belongs to; the Python function verify_claims keeps the earlier, smaller set of checks and verify_and_repair still wraps it.

Every field is in the worked example, and the adapter recipe maps another shape to it.

The exact input the checker ran lands here.