Draft → Challenge → Judge → Tie Off
Important AI work shouldn't depend on one model.
Ariadne has one AI create the work and another try to break it — while you make the final call.
- AI does the research
- another AI tries to break it
- you decide what's right
- Ariadne revises it
- you approve the final result
Already have an account? Sign in
How it works
One model drafts. Another tries to break it.
You don't copy answers between two chat windows. Ariadne handles the handoff. You handle the judgment.
1 · Draft
Odysseus drafts
Claude · the Drafter
Writes the first answer to your question, from your brief.
Works todayComing next: the answer first, with numbered claims and sources underneath.
2 · Challenge
Daedalus challenges
GPT · the Challenger
A different model tries to break the draft: what's wrong, unsupported, outdated, contradicted or missing.
Works todayToday it works from what it already knows. Coming next: one click, web search, and every challenge tied to a claim and a source.
3 · Judge
You judge
The knot
Nothing moves on until you decide: tie it off, or send it back with a note.
Works todayComing next: accept, reject or edit each challenge, or add your own.
4 · Tie Off
You tie it off
Signed off by you
The version you approved is kept with who drafted it, who challenged it and who tied it off. Earlier versions stay in the history.
Works today
A pass, in miniature. Try it.
Scroll here and the hands start passing the work along.
Don't ask another AI if the answer is good. Ask it to try to prove the answer wrong.
Today What you can use now
- Projects with a brief that every step reads first.
- A five-step thread: Odysseus (Claude) researches and drafts, Daedalus (GPT) challenges the draft, you judge at the knot, and Odysseus writes up the version you approved.
- Challenges you can act on: add any of them to your note in one click and send the draft back. Odysseus revises, Daedalus challenges the new version, and a v1 → v2 comparison shows what changed. You press Run for each step.
- Nothing is overwritten. Every run is kept in the history.
- The models work from what they already know. They don't search the web yet.
Coming next What we're building
- Answer-first drafts with numbered claims, so you can read the answer in seconds and point at exactly what you doubt.
- One-click Challenge: the challenger searches the web, and every challenge names a claim and carries a source. Challenges without one are labelled as opinion.
- Judging point by point: accept, reject or edit each challenge, or add your own.
- Revise and re-challenge only what changed: the drafter fixes the claims you accepted, the challenger checks just those, and you tie off. A round or two, not an endless debate.
- Quick Answer and Deep Work: a fast answer when that's all you need, the full thread when it matters.
The labyrinth
AI writes research fast. Checking it is still on you.
When the work matters, the answer arrives and the questions start:
- Is this claim true?
- Does the source actually support it?
- Is it out of date?
- Did the model miss something?
- Do two sources contradict each other?
- Did it confidently make something up?
So AI sometimes creates a new job: checking the AI. Ariadne is built to make that job smaller.
Who it's for
Researchers who lean on AI and can't afford to be wrong.
- You use ChatGPT, Claude or both for literature reviews, summaries, analysis or reports.
- Your work carries numbers, citations or claims that someone else will check.
- You already paste one model's answer into another to see if it holds up.
Analysts, builders and students doing serious work come later — researchers first.
Most AI tools give you an answer. Ariadne gives you a way to challenge the answer before you trust it.
What Ariadne isn't: ChatGPT with more models, AI agents talking to each other, or a promise of no hallucinations. Models still make mistakes, and the challenger is a second opinion, not an authority. The last word is always yours.
An example of how it works
One question, start to finish.
An illustration of the steps, not real output. Some steps already run today in a simpler form; the labels say which.
You ask“Map the strongest pre-seed investors in Montreal investing in AI developer tools.”
Draft
Odysseus writes the answer first: a shortlist, with the reasons for each name underneath, and where it isn't sure.
TodayChallenge
Daedalus goes looking for what's wrong: a fund that no longer writes pre-seed cheques, a partner who has moved on, a source that doesn't say what the draft claims, an obvious name that's missing.
Today, without web searchJudge
You go through the challenges. A real problem? Accept it. Not valid? Reject it. Half right? Edit it. Something nobody caught? Add your own. AI does not get the final vote.
Point by point: coming nextRevise
Odysseus rewrites the claims you accepted or edited, not the whole answer.
Coming nextRe-challenge
Daedalus checks only what changed. A round or two, not an endless debate.
Coming nextTie Off
“I'm satisfied with this version. This is the final answer.” It's kept as finished work, not another chat.
Today
Today, judging happens in one note: add the challenges you agree with and send the draft back. Odysseus redrafts and Daedalus challenges the new version in full.
Where the judgment sits
Agents do the work. Agents challenge the work. Humans judge the work.
Ariadne doesn't replace your judgment. It takes on the mechanical work around it:
- copying context between models
- switching between chat windows
- asking one model to review another
- a first pass at checking claims
- hunting for contradictions
- reconciling two answers
- prompting again, and again
Coming next · Two ways in
A quick answer, or the full thread.
Today every project runs the full thread. Soon you'll choose how deep to go.
- Coming next
Quick Answer
Just need the answer? Start here.
Get a fast answer, then challenge it if it matters.
- Coming next
Deep Work
Working on something important?
Let another model try to break the answer before you sign off on it. Draft → Challenge → Judge → Revise → Tie Off, still answer-first, with the detail underneath.
The door
Pick up the thread.
Ariadne helps you verify important AI-generated work by having one AI model challenge another, while you remain the final judge.
Tie the ThreadFree while we test it with a small group of researchers.