superthesis
/ in a world filled with noise

FIND THE
SIGNAL.

An AI that fact-checks any claim by arguing both sides and scoring the evidence — every verdict a calibrated 0–1 reading, every source cited.

/ every verdict traceable to its source — no black box
What it is

Thesis argues with itself. An Antithesis attacks your claim, a Thesis defends it, a Synthesis scores the evidence. The output isn't an opinion — it's a calibrated reading from 0 to 1, every source cited.

Use Thesis to debate your claims, answer your questions, and draft your plans.

Built for analysts, researchers, and anyone who can't afford to be confidently wrong.
✓ No credit card✓ Calibrated confidence✓ Every claim sourced
Sample reading
0.73
falsecontestedtrue
strong · leans true
01 / decompose

Split the claim

Broken into the specific, checkable sub-claims it depends on.

02 / argue

Both sides, sourced

Antithesis attacks, Thesis defends — neither gets the last word.

03 / adjudicate

Score the evidence

Synthesis rules on quality alone, then opens the full audit trail.

/ try it · no signupinteractive demo · prepared examples
/ don't take our word for it

Watch a belief get tested.

Run one of the prepared examples and watch Thesis argue both sides and resolve a calibrated signal. Live runs on your own claims open after sign-up.

CLAIM ›
Try
0.00
SIGNAL
resolving…
0 · FALSECONTESTEDTRUE · 1.0
Thesis
Antithesis
Synthesis
prepared examples · live runs cite real sources
/ one engine · three jobsthe same adversarial core
/ what it does

Point it at whatever you need settled.

Debate your claims

Test a belief

Give it something that can be true or false. Thesis splits it into checkable parts, argues each, and returns one calibrated 0–1 signal — reasoning, not opinion.

Answer your questions

Resolve a question

Ask anything open-ended. Thesis researches every side, weighs the competing evidence, and synthesizes it into one clear, sourced answer — the full debate, resolved.

Draft your plans

Build a plan

Name a goal. Thesis researches the path and lays out a time- and cost-optimized plan — the high-level arc down to concrete steps, with estimated hours and dollars.

/ benchmarksmeasured in public
83.4%
Accuracy on the AVeriTeC dev set — vs. 71% prior published best (Apr 2026). Methodology →
70 = 70
Calibrated confidence: when Thesis says 0.70, it means 0.70 — expected calibration error just 0.03. Real uncertainty, not false precision. Reliability curve →
100%
Of verdicts open to the exact passage behind them — auditability, not just a citation count.
We don't pretend raw accuracy is our edge — debate trades a little of it for calibration and stability you can trust.
See the full benchmark →

The everyone-knows test

/ the everyone-knows testyour call · 0 / 0
/ did they really say it?
your call —
0 · refutedcontestedverified · 1.0
0.00attribution signal

/ why adversarial debate beats a single model

A single model answers. Three agents that disagree are forced to show their work — and that difference is everything when you need to be right.

Built for those who can’t afford to be wrong.

Run your thesis free — no credit card, no limits. Everything's on us while The Field is in early access.