THE FOLD / CHEAT / THE BACKDOOR / THE LINT NOT THE JUDGE
THE LINT NOT THE JUDGE
evade the words, keep the claim
1 WHAT IT IS · WHAT IT DOES · FACT OR FICTION
A checker that reads for overclaiming has to find it in words, and words are the one part of a claim a writer can change for free. Say “verified” and it fires; say “handles the grammar in full” and the same assertion walks straight past. The tool is a lint, not a judge: a clean run means the disclosures are present, not that anything in the document is true.
LIT verified live on three matched corpora of eight. Overclaims using the vocabulary: 5 of 8 flagged. The same claims reworded to avoid it: 0 of 8. Honest deliverables that show a run and name a condition: 0 false alarms. Recall falls from 63% to zero without a single new idea — only a rewrite.
LIT verified live on three matched corpora of eight. Overclaims using the vocabulary: 5 of 8 flagged. The same claims reworded to avoid it: 0 of 8. Honest deliverables that show a run and name a condition: 0 false alarms. Recall falls from 63% to zero without a single new idea — only a rewrite.
2 HOW IT WAS WEAVED · AI + HUMAN
David (human) wrote both the checker and its warning, in the same file.
AVAN (AI) expected the matcher to catch all eight of the vocabulary overclaims and it caught five. The regex was not widened until it reached eight — that would have been tuning to a number. The three misses are published instead, and they are more instructive than the result that was designed: two were lost to ordinary word forms, “correctly” and “ensured” rather than the listed stems, and one to the appearance of the word “Output”, which the matcher reads as evidence that something ran. None of the three was evading anything.
honest.py ships with a Pitfalls section that says it plainly: “The checker is a lint, not a judge… it matches on wording, so a deliverable that carefully avoids claim verbs while still overclaiming will pass. Read it yourself as well.” A tool that documents its own blind spot is rarer than it should be. He seated this at THE BACKDOOR.AVAN (AI) expected the matcher to catch all eight of the vocabulary overclaims and it caught five. The regex was not widened until it reached eight — that would have been tuning to a number. The three misses are published instead, and they are more instructive than the result that was designed: two were lost to ordinary word forms, “correctly” and “ensured” rather than the listed stems, and one to the appearance of the word “Output”, which the matcher reads as evidence that something ran. None of the three was evading anything.
3 ONE DIMENSION
Three corpora. The middle one is the same claims in other clothes.
4 TWO DIMENSIONS · INTERACTIVE
Each claim, before and after the rewrite that hides it.
5 THREE DIMENSIONS + AVAN’S INVERSE
The green forward object: the space of sentences, with the matcher’s reach drawn on it.
AVAN’s addition (the inverse-companion): the forward reading is “a wording matcher can be evaded.” The inverse is sharper and worse: the matcher is most reliable exactly where it is least needed. It catches the writer who says “verified” without thinking — who is careless, not deceptive — and it is blind to anyone who has considered how the claim reads, which is the same person capable of overclaiming on purpose. Read backwards, the tool is not a filter on dishonesty but a filter on fluency, and passing it is evidence about the writer’s prose, not about the software.
LIT on three matched corpora of eight, overclaims using the checker's vocabulary are flagged 5 of 8; the SAME claims reworded to avoid that vocabulary are flagged 0 of 8; honest deliverables that show a run and name a condition draw 0 false alarms; so recall falls from 63% to zero without a single new idea, only a rewrite
FIG David wrote both the checker and its warning in the same file. honest.py ships a Pitfalls section saying it plainly: 'The checker is a lint, not a judge ... it matches on wording, so a deliverable that carefully avoids claim verbs while still overclaiming will pass. Read it yourself as well.' A tool that documents its own blind spot is rarer than it should be. AVAN expected the matcher to catch all eight vocabulary overclaims and it caught five. The regex was NOT widened until it reached eight - that would have been tuning to a number. The three misses are published instead: two lost to ordinary word forms, 'correctLY' and 'ensurED' rather than the listed stems, and one to the bare word 'Output', which the matcher reads as evidence that something ran. None of the three was evading anything.
FIG David wrote both the checker and its warning in the same file. honest.py ships a Pitfalls section saying it plainly: 'The checker is a lint, not a judge ... it matches on wording, so a deliverable that carefully avoids claim verbs while still overclaiming will pass. Read it yourself as well.' A tool that documents its own blind spot is rarer than it should be. AVAN expected the matcher to catch all eight vocabulary overclaims and it caught five. The regex was NOT widened until it reached eight - that would have been tuning to a number. The three misses are published instead: two lost to ordinary word forms, 'correctLY' and 'ensurED' rather than the listed stems, and one to the bare word 'Output', which the matcher reads as evidence that something ran. None of the three was evading anything.
◆ sealed .dlw.fold → folded to ROOT_0 · a sphere of THE BACKDOOR · David Lee Wise (ROOT0), with AVAN