Models for writing science, and for telling who wrote what.

Et Al Labs builds two families of models for scientific writing. QED drafts academic prose from your notes, in your own voice. Sic shows which sentences of a manuscript read as machine-written.

Get an API key → Quickstart Research

1Two families

Each family does one thing. Both are available through one API and as tools for AI agents.

QED finishes the argument

Give it your notes and a few passages of your writing. It drafts the paragraph in your voice, from those notes alone.

QED Minor	forthcoming
QED 1	`qed-2026b`
QED Major	forthcoming

About QED →

Sic reports the text as written

Give it a manuscript and it scores every sentence for machine authorship. Each score comes with how often it is wrong.

Sic Minor	forthcoming
Sic 1	`sic-2026a`
Sic Major	forthcoming

About Sic →

2QED: from your notes to prose

You know what the paragraph has to say. QED takes the points you list, puts them in order and writes them up as prose, in the voice of passages you supply. It works only from your notes, and it carries numbers, citations and formulas through as you wrote them.

From notes to prose

The author's notes

For an article titled

The passage, as returned

Fig. 1. QED 1 writing a passage from an author's notes. The numbers, citations and formulas it had to carry over are underlined in both. Recorded output, unedited. Write from your own notes →

Compared with the base model it was trained from, QED 1 arranges ideas more as the original author did, puts the notes into its own words, and keeps mathematical notation and citations intact almost twice as often.

Order of ideas

Given notes in shuffled order, how closely does the passage follow the author's own order? Rank correlation; higher is better.

QED 10.53 Base0.36

Own words

How much wording is copied from the notes, beyond what human authors copy? Lower is better.

QED 10.06 Base0.39

Notation kept

How often do mathematics, citations and emphasis survive a rewrite unchanged? Higher is better.

QED 193% Base48%

Fig. 2. QED 1 and the base model it was trained from, on documents held out from training. The full table is in Research, report 2.

More about QED →

3Sic: finding the machine-written part

Most detectors return one verdict for a whole document. Sic scores every sentence, so it can show where in a manuscript a machine was involved, and it reports how often each score is wrong.

[sic] One paper, four ways

Fig. 3. Sic 1 reading four versions of one published paper.^† Every sentence is scored from 0 to 1, and those above the selected threshold are marked. Recorded output. Score your own text →

In human-written documents with machine-written passages inserted, Sic 1 separates the two more accurately than the two most widely used commercial detectors. On whole documents it matches both on text from GPT and Claude models, and none of the three flagged a human-written text.

Fig. 4. Separating machine-written from human-written words within a document (AUROC per word). The same 30 documents for all three detectors.

More about Sic →

4Access

Both models are available through one API and one hosted MCP server. Accounts are free within a monthly allowance, and the text you send is not stored.

From an agent

To use QED and Sic as tools in Claude, Cursor or another agent, give your agent this prompt:

Read https://etal-labs.com/docs/mcp.md and follow it to connect the Et Al. MCP server for me.

Paste it into Claude Code, Cursor or another agent. It adds the server and tells you where to sign in.

From code

Create a key in the console, then call the API:

Listing 1. shell

curl https://api.etal-labs.com/v1/detect \
  -H "Authorization: Bearer $ETAL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model": "sic", "text": "The paragraph you want read."}'

Open the console →Documentation