Several minds, one answer

Frontier models answer your question separately, then argue about each other's answers. What survives the argument gets written up as one answer.

Argument should be free. The models that answer your question cost real money to run, and one day this will have a price on it. What will never sit behind a wall is the part that matters: several of them disagreeing in the open, where you can read it. Anyone can ask without an account, without a card, and see the whole argument rather than the tidy conclusion.

Three steps, and none of them is a vote

They answer alone

Every model gets the same question and none of them can see what the others said. Agreement only means something when it was reached independently. The moment models read each other first, they start agreeing for social reasons rather than factual ones.

They argue

Each one reads the others and goes after the weakest claim it can find. Invented figures and confident guesses rarely survive being read by a model with no reason to be polite.

One answer gets written

A chair writes the final answer out of what held up. Disagreement that survived is kept and labelled rather than averaged away, so you get a range where the truth is a range, and a number where it is a number.

Who is in the room

What this is not

It is not a majority vote, and it is not a way of making a weak model strong. Models can agree on the same wrong thing, and letting them talk can make that worse rather than better. What Synapse gives you is the working: several independent attempts, the objections each one raised, and a final answer that had to survive being read next to them. When they disagree, you see it, and that is the part a single chat window hides.

In training

DELA 1.0, our own seat at the table

Every run here produces something rare: the same question answered independently several times, then the objections each model raised against the others, then the answer that survived. That is a record of reasoning being corrected, not just of text being produced.

DELA 1.0 is Elva's own model, trained on that record. An open Qwen base to start with, tuned on the arguments this council has already had, built to be useful on real work rather than cautious about it, and to say plainly when it is unsure instead of hedging everything.

  • BaseQwen, open weights, fine-tuned by Elva
  • Trained onCouncil transcripts: answers, objections, and what held up
  • AimStraight answers on real work, and honest uncertainty
  • StatusIn training. It takes a seat here when it earns one

Start with no account at all

Ask something you would normally want a second opinion on

Ask your first question