By Ryan Richardson · Published 8 October 2026
Leak check: any internal tag or process word in the reader's copy, threshold zero. The lint: any em dash, plus a fixed list of machine-sounding and filler words, threshold zero em dashes and at most six per thousand words. Tell scan: sentence openers a machine favours, at most 0.08 per thousand words. Density: mean sentence length around 15 words, no paragraph over 120 words. Overlap: nine-word runs shared with the author's own earlier work, threshold zero. Trace check: every personal sentence tagged, every cited fact id real, threshold zero bad ids.
None of these six thresholds was picked after the fact. The tell-scan line was drawn by measuring real writers first: a well-known clean human writer measured 0.000 per thousand, a well-known business author measured 0.075, and prose a language model had revised measured 0.464, nearly six times over the 0.08 target.
A search over one reading copy found 53 leaked internal tags across 24 lines: 46 fact-citation brackets, 5 interview-round tags, and 3 stray process words. One of them had been sitting inside a quote, dressed up as part of the sentence being quoted, which is exactly why it read as plausible on a first pass and wrong on a second.
The overlap gate's first run on one manuscript failed outright, finding 11 shared nine-word runs with an earlier book, the longest 20 words. Every one was the same true story told in near-identical words across both books. All 11 were reworded, and the re-run came back at zero.
Plant the exact defect a gate claims to catch, and confirm the gate fails: a fake citation bracket slipped inside a quote for the leak check, one em dash anywhere for the lint, one paragraph opened with a known machine phrase for the tell scan, one deliberately long sentence for density, nine words lifted from your own earlier work for overlap, one untagged personal sentence for the trace check. A gate that stays green on a planted defect isn't a gate yet. Never give any of these six an override flag, because an override gets used once, at eleven at night, under deadline pressure, which is exactly when the gate is worth the most.
| Claim | Value | Source |
|---|---|---|
| Leaked internal tags found by the leak check in one run | 53, across 24 lines | The Sixty Steps manuscript |
| Shared nine-word runs found on the overlap gate's first run | 11, longest 20 words | The Sixty Steps manuscript |
| Tell-scan rates measured across real and AI-revised writers | human writers 0.000-0.075 per thousand words; AI-revised prose 0.464; target 0.08 | The Sixty Steps manuscript |
| Matrix difficulty/coverage for this step | Easy / Almost nobody does this | The Sixty Steps matrix |
| Density gate paragraph ceiling | no paragraph over 120 words | The Sixty Steps manuscript |