Performance
Every number here comes from writers Pessoa never trained on. It saw two of their texts and nothing else. These are our own evaluations, with the bar for success set before each test, and they haven't been peer reviewed yet.
Case Studies
Each draft was written by an open AI model from a real person's notes. Highlighted words are Pessoa's own; the rest is carried over from the draft.
The author is surprised by the judgmental responses to steak preferences, emphasizing that taste is subjective, much like preferences for caviar or coffee. Initially, they only ate well-done steak due to family habits but later preferred medium-rare. Some claim they can't tell the difference in well-done steak, but the author disagrees. For well-done, choose cuts with a loose texture and fat, like skirt or flank. Expensive cuts like ribeye can still be good if cooked properly, while lean cuts like filet mignon and sirloin are poor choices. Cuts with lots of connective tissue, like chuck, should be avoided. Marinating helps add moisture, and flipping the steak often during cooking helps retain juices. Avoid flipping once and butterflying, as they can dry out the meat. Cooking at lower heat and checking internal temperature is important. Letting the steak rest after cooking helps it reach the desired doneness. Cooking well-done steak requires more skill than medium-rare.
I'm surprised at the judgmental attitudes being expressed here about steak preferences. There is no right or wrong answer. The taste is what it is, just like many other foods. Some people like caviar, some don't. Some people like coffee, some don't. Some people like steak cooked rare, some don't. End of story. I was a well-done-only eater when I was a kid, and my family ate that way, but when I learned to cook, I became a medium-rare person. I think many people who say they can't tell the difference between a medium-rare and a well-done steak are lying to themselves. I think the actual reason they like well-done is because they don't like the texture of rare. If you want a well-done steak, the key is to choose a cut that is going to have more marbling or loose texture. Skirt and flank are among the best cuts. You can even get a good steak well-done from a ribeye if you have a thick one, but this is a hard technique to master. Filet mignon and sirloin are really poor cuts for a well-done steak. Cuts with lots of connective tissue (like chuck, round, etc.) are a really bad idea. Always marinate to add moisture. (Especially if you're cooking to well-done.) Always flip the steak often. (Never flip it once and butterflying it, as the latter will dry out the steak as it cooks.) Always cook at a lower heat to keep the temperature inside the steak more even and to avoid overcooking the outside. Always check the internal temperature to make sure you don't overcook the steak, and let it rest for a minute or two after you take it off the heat to let the juices redistribute and the center finish cooking to the desired doneness. Cooking a well-done steak is more of an art than a science and requires more skill than cooking a medium-rare steak.
Claims kept and added come from claim-level judges whose scores on a perfect copy are 0.85 kept and 0.11 added, so read anything near those as the ceiling. Quality is judged out of 5.
Open-Source Detectors
On 48 writers Pessoa never saw, we rewrote AI drafts made from each writer's notes. Each detector is set so it flags only 5% of real human writing. The chart shows the mean of six open detectors before and after rewriting, next to the writers' own texts.
Detectors: desklib, MELD, fakespot, SuperAnnotate, Binoculars and our own. 395 drafts from 48 held-out writers. Judged quality: 4.45 of 5 for rewrites, 4.79 for the writers' originals.
Pangram Evaluation
Pangram is the commercial detector built to catch humanized text, and it flags none of our writers' original texts. It reads Pessoa as human when the ideas come from the person. When the AI supplied the ideas too, it usually doesn't.
469 rewrites, up to 40 per test, scanned by hand in Pangram 4.1 in October 2026. A rewrite can sit in several tests.
Comparison with Published Methods
Same 332 rewriting tasks for every method. Prompting keeps the content, but almost every AI draft is still flagged. TinyStyler evades detectors only by breaking the text.
| Method | AI drafts still flagged | Meaning kept | Quality, of 5 |
|---|---|---|---|
| Pessoa | 35.5% | 0.75 | 4.51 |
| Few-shot prompting (STYLL), your samples as examples | 96.6% | 0.80 | 4.68 |
| Register-analysis prompting | 94.5% | 0.73 | 4.48 |
| TinyStyler, the best open style-transfer model | 49.3% | 0.40 | 1.28 |
| AI draft, unchanged | 97.8% | – | 4.66 |
| What people use today | Examples | What you get |
|---|---|---|
| Editors and humanizers | Grammarly, QuillBot, Wordtune | Everyone nudged toward the same polished style |
| Brand-voice tools | Jasper, Writer | One voice for a whole company |
| Style profilers | My Writing Twin | A description of your style, with no rewriting in it |
| Prompting with your samples | ChatGPT or Claude, examples pasted in | Meaning kept, but the chat-model tone stays: 97% of drafts still flagged |
| Personalect | Pessoa | Your own voice from two samples, with every claim checked |
Generalization
Pessoa kept recognizing the target writer and kept quality near 4.5 to 4.8 across fiction, Q&A, essays, a model family it never trained on, and drafts of 600 to 2,500 words.
| Slice | Texts | Writer recognized (AUC) | Claims kept | Quality |
|---|---|---|---|---|
| Kinds of writing it trained on (reference) | 748 | 0.78 | 0.76 | 4.50 |
| Fiction | 124 | 0.77 | 0.71 | 4.78 |
| Q&A answers | 145 | 0.88 | 0.69 | 4.62 |
| Essays on ideas newer than its training | 116 | 0.81 | 0.76 | 4.44 |
| From telegraphic notes | 185 | 0.87 | 0.79 | 4.48 |
| From bullet notes | 200 | 0.78 | 0.65 | 4.75 |
| Drafts by Granite 3.3 8B | 90 | 0.85 | 0.70 | 4.76 |
| Drafts by OLMo-2 13B (never trained on) | 78 | 0.79 | 0.63 | 4.73 |
| Short drafts (~150 words) | 86 | 0.84 | 0.85 | 4.47 |
| Full-length drafts (600–2,500 words) | 88 | 0.77 | 0.69 | 4.71 |
| Voice from two 300-word excerpts | 88 | 0.73 | 0.71 | 4.77 |
On a public Reddit style-transfer benchmark, Pessoa matches the best open model and clearly beats GPT-4 (0.45 against 0.31), with the highest meaning similarity of any system.
Rewrites share 0.07% of their eight-word runs with the training data. Pessoa writes new sentences rather than recalling old ones.
Going from one to four sample documents barely changes the result, so a voice works from two short samples.
Architectural Philosophy
Each request runs start to finish on our own model. No outside AI service sees your text.
Request Pipeline
- InputWriting samplesTwo texts by youInputAI draftFrom any model
- ModelPessoaReads your voice card and rewrites the draft as you
- CandidatesSeveral versionsFacts fixed, voice free
- SelectionClaim checkNothing dropped, nothing added; closest to your voice wins
- OutputYour rewriteWith an invisible, keyed watermark
Design Principles
Person-Specific Targeting
A person's habits are a narrower, measurable target than "sounding human", so every rewrite is pulled toward yours.
Usually: everyone is pushed toward one generic style.
Detector-Independent Training
No AI detector shaped Pessoa. We test it on detectors it never saw, because gaming one only lasts until it retrains.
Usually: humanizers iterate against detectors, and the text reads badly.
Meaning Preservation
A rewrite that changes what you said is unusable. Every claim is checked, and quality has to clear 4 out of 5.
Usually: paraphrasers rarely measure meaning at all.
Zero-Shot Personalization
Your voice card and two short excerpts are all it needs, so a new voice works the moment you add it.
Usually: per-person fine-tuning, which is slow and didn't help in our tests.
Base-Model Foundation
Chat-model tone is the biggest giveaway, so Pessoa starts from a model that never learned it.
Usually: a wrapper prompts ChatGPT or Claude, and their tone comes along.
Content–Style Separation
The words that carry your facts stay steady. The words that carry style are free to become yours.
Usually: one setting for every word, trading meaning against naturalness.
Built-In Provenance
A watermark and our own detector, so we can show when Pessoa wrote a text.
Usually: evasion only, which rules out schools and platforms as customers.
Out-of-Distribution Evaluation
New writers, new kinds of writing and new AI models, with the bar for success set before each test.
Usually: hand-picked demos, or tests on the training data.