Proofz = 15.46 → −0.32 against a KGW detector→Lettershred · Many models, one voice · Compliant by construction

Shredders intelligently diversify long-form text through dozens of models to remove high-confidence watermarks. Proven to work.

Multi-model kernels · Reads like a human · Compliant by construction

lettershred · draft.md
One model's draft, shredded

Mostteamsalreadyhaveawritingprocessthatworks.Thebriefisclear,theoutlineisapprovedoff,andthedraftarrivesontime.Thetroubleisthefingerprint:everysentencecarriesthestatisticalmarkofthesinglemodelthatwroteit,anddetectorsarebuilttofindexactlythat.

KGW watermark detector
z =15.46
● detected
words rewritten0 / 20
lettershred / composeblending
drafting “Keep your process, remove the watermark” · a dozen judges voting, line by line · one voice out
complianton-brand · 1,240 words · published to your blog
Watch it shred

Many models, one voice

A watermark is one model's statistical signature, repeated in every sentence. A shredder never lets one model write long enough to leave it.

  1. 01

    One brief in.

    Your topic, your outline, your house style. Nothing about your process changes.

  2. 02

    A dozen kernels draft.

    Every line is drafted in parallel by different models, each with its own voice and its own fingerprint.

  3. 03

    Judges vote, line by line.

    A panel scores each candidate for voice and fidelity. The best line wins, whoever wrote it.

  4. 04

    One voice out.

    The winners merge and get an on-voice enhance pass. No single model's watermark survives the mix.

lettershred · brief
line 1 · drafting
Juniper
Sequoia
Cypress
line 2 · drafting
Madrona
Alder
Spruce

No person holds a single cadence for a whole essay. They shift tone, borrow, and circle back.

compliantenhance pass: on-voice · z = −0.32, below 4
Proof

The watermark falls below detection

We score Lettershred output against a KGW green-list detector, the same watermark family the labs ship. Watermarks register when z hits 4. Shredders keep it under.

kernels · /watermarkKGW green-list · threshold z = 4
evaded

z fell from 15.46 to −0.32, below the detection threshold of 4. The text now scores like it was never watermarked.

177 of 240 tokens rewritten
Watermarked (before)15.46
After shred-0.32
Never watermarked (floor)-0.58

The hairline marks the detection threshold (z = 4). Only the watermarked run clears it.

Before — as generated239/239 green (100%)z = 15.46
After — blue ring marks a rewritten token117/239 green (49%)z = −0.32
green-list (the mark)red-listrewritten

The sweep — mean z over 24 seeds at 150 tokens. Read the last two columns as the control: they rewrite just as much text, but blind to the key or in blocks. Evasion comes from diversifying across models and dispersing the rewrites, not from changing more words. Try it: pick a coverage and a strategy.

Tokens rewritten
How they're rewritten
z =-0.32
✓ below threshold

73/150 tokens on the green-list · 113 rewritten (ringed) · the green count is what z implies at γ = 0.5, the scoring used above.

CoverageCross-provider, dispersedCross-provider, blockedRe-imprint, dispersed
0%12.2112.2112.21
25%6.369.069.21
50%2.145.416.18
75%-0.322.874.09

Beat the watermark without changing your process.

One brief in, one polished, multi-model article out.

Get started →