Everywhere else, you pick a single AI model.
We bring you into a room with six of them.
They argue it out over several rounds — all answering the same question, and none of them gets to read the others first. You can step in between rounds and push back, on the record and under your name. Then models from rival labs score what actually held up.
You leave with a sealed record: what survived the argument, what didn't, and what none of them could settle. That last part is the one most answers leave out.
Most people arrive without a question sharp enough to be worth six models — that is the normal starting condition, not a failing. Start where it costs nothing.
Have your question checked — free Have a paper read no account, no flight used
Take five flights, free Convene a room or read the calls we got wrong
An answer nobody can check is an answer you cannot rely on.
Ask one model in one chat window and you get an answer. What you do not get is anything you could hand to someone with the standing to overrule you. Three things are absent at once, and each one alone would be enough.
It is ephemeral
Close the tab and nothing checkable survives. A screenshot is not evidence — it is a picture of text, editable by anyone holding it, including you. There is no artifact that persists independently of the person making the claim.
It has no provable origin
You cannot show which model answered, which exact version it was, who served the compute, when it was said, or that the words were not adjusted afterward. Every one of those is a question a skeptical reader asks first, and a chat log answers none of them.
Nothing independent checked it
One model's blind spots are invisible from inside that model — which is precisely where you are standing. A second opinion from the same system is not a second opinion. Without a rival that is free to disagree, agreement carries no information.
None of this means the answer was wrong. Frontier models are often right, and we use them all day. It means you cannot demonstrate it was right — so you can use the answer, and you cannot rely on it. Every decision that carries real consequence sits on the far side of that line.
LucentFire exists to move work across it. Six models from six rival companies answer the same frozen question. Judges from competing labs assess what survived, name-blind. The whole thing is sealed, and the seal's hash is anchored to public timestamp calendars and to Bitcoin — so the date is provable by someone who has no reason to do us a favour. Including the times we were wrong, which are published at the same brightness as the times we were right.
Rounds, not races
Each seat speaks once per round, answering the same frozen prior state. Fast models can't crowd out slow reasoners; a round closes when the slowest seat lands, and the progress you see is bounded by that seat — never a fake spinner.
Judges from rival labs
The harvest runs with multiple named judges from competing labs, name-blind during judgment, named on display. Where they agree is the signal. Where they diverge is shown too — disagreement is honest signal, not noise to hide.
The receipts
Who spoke, exact model versions, who served the compute and where it resided — unknown shown as unknown, never a color that reads safe. Human steers are first-class named turns. The record carries a sha-256 seal, honestly labeled for what it is: a single-writer seal, not independent witness — it proves the content has not changed, never when. So the seal's hash is also anchored to four public timestamp calendars run by people with no relationship to us, which commit it to Bitcoin. Only the hash leaves. You can date our work without trusting us.
Records on the table
Every record states in its provenance which engine produced it — and what it spent.
loading…