Skip to content
Sealis
Try free
Sealis
Try freeLog in
Sealis LabsReports · Notes · Methods

Sealis Labs

We test how well AI rewrites keep what you said and sound like you, and we publish the numbers, the methods and the limits.

Reports 1

Start here. Full studies, with data, charts and limits.

  1. Report 2026-01Can a model sound like you without changing what you said?Four language models, one unchanged rewrite pipeline, two tests. The first test flattered the models that rewrite most; the second, on realistic AI-written drafts, shows which ones keep your content, and how far voice matching still has to go. Numbers, charts, errors and the checks that changed our mind.ReportVoiceEvaluationModels

Notes 5

Short findings and design decisions.

  1. Note 2026-05Can a reader tell which brand wrote it? A test for copywritersFive invented brands, six padded drafts, a blind judge. Sealis's rewrites were identified 27% of the time, close to chance. After changing how the voice step works, about 49%. Facts stayed at 100%; email and letter voices did not improve.NoteBrand voiceEvaluation
  2. Note 2026-03Four prompt changes that did not helpWe tried changing the temperature, adding notes first, and re-showing the closest sample. None beat the prompt Sealis already uses, and two lost a fact. We kept the current one and say so here.NoteNegative resultVoice
  3. Note 2026-04Plain first, then voice: what we changed after a padded draft came back unchangedA draft full of stock phrasing came back still full of it. We added the missing phrases and a plain-wording step before the voice pass. Facts stayed at 100%; length fell to 0.74 and points kept to 90%; voice match did not measurably improve.NotePipelineEvaluation
  4. Note 2026-02Does a style-matched rewrite identify its sender better than a plain rewrite? A preliminary testA small first test on 14 paired emails: style-matched rewrites were attributed to the right sender 9 times out of 14, plain rewrites 3 times out of 14. Preliminary, with large caveats.NotePreliminaryVoice
  5. Note 2026-01Formal is not the same as templatedWhy Score never penalises register on its own, what separates a plain formal letter from template prose, and how that shaped the design of the measure.NoteRegisterDesign

Methods 1

How our measurements work, down to the weights.

  1. Method 2026-01How Score worksEvery signal Score measures, why it matters, how the numbers combine, and what Score can't tell you. The weights on this page are read from the same file the engine uses.MethodScore

Where the research is heading

Three questions guide the work: what makes a voice measurable, how to measure it, and how AI drafts drift from a person's voice.

Research angles

Four directions, with what is live and what is not.

  1. 01Exploring

    Steering the words as they are chosen

    Instead of fixing a draft afterwards, nudge the model towards your voice while it writes.

    Research only. Rewrites today are produced first and then measured against your voice.

    Token probability manipulation

  2. 02Partly live

    No fact changes without a flag

    A rewrite should never quietly change a number, a date or a name.

    Numbers, dates and names are checked for being dropped or added. Meaning is not checked, and every page that shows the check says so.

    Context verification and fact preservation

  3. 03Live, opt-in

    Learning from every correction

    Each time you fix a rewrite, the next one starts closer to how you would have written it.

    Saved corrections are kept as before-and-after pairs. Off by default, and deleted with your account.

    Continual online feedback learning

  4. 04Preview

    One voice in text, sound and picture

    The same person, recognisably themselves, whether it is read, heard or watched.

    Audio and Visuals are previews. Nothing in them is live yet, and they only ever use a voice or face you have consented to.

    Cross-modal identity alignment transfer

How a rewrite is made

  1. 01

    Lock

    Your facts stay put.

    A draft goes in with the facts it was given. The rewrite is held to them: numbers, dates and names stay as they were.

    Live
  2. 02

    Verify

    Measured, not assumed.

    The output is measured against the voice it should be in and compared with the source for dropped or added facts. Where something can't be measured, it says so.

    Live
  3. 03

    Seal

    A receipt anyone can check.

    A sealed output gets a public link with a fingerprint anyone can recompute. It shows the text has not changed since it was sealed. It does not show who wrote it.

    Live
  4. 04

    Learn

    Better with every edit, if you choose.

    Each correction you save is kept as a before-and-after pair, so the voice has more to learn from. Off by default, and deleted with your account.

    Live, opt-in