A single model is a poor critic of itself. It will follow your framing, fill a gap with something plausible, and repeat a point back to you as though it were confirmation. None of that looks like an error while you are reading it, which is exactly the problem. The usual workaround - asking the same question in three browser tabs - leaves you comparing three answers that never met.

• The failure is not that the model is wrong. It is that nothing in the room disagrees with it.

A catch is a specific, typed observation about the conversation rather than about the world. Lumi hosts the room and names four kinds: a contradiction between two answers, a claim nobody in the room can check, a point repeated as if it were new, and an assumption everyone is standing on that nobody stated. Each one is attached to the message that caused it.

• Typed, not scored. A catch tells you what to look at, not who won.

The models are yours. You paste a provider key for each seat at the table, the key is sent with your request and never stored, and Drift Catchers orchestrates the room around them. Gemini, ChatGPT, Grok and Claude are there by default, and any model with an OpenAI-compatible endpoint can be added with a base URL, a key and a model id.

• Your models stay yours. That is what keeps the disagreements real.

A session that ends in a wall of text has not helped anyone. Closing the room produces a decision brief in four parts: what the room agreed on, what is still unresolved, what needs evidence before it can be relied on, and a recommendation. That brief is the thing worth keeping, and it is the only thing that is saved beyond the browser.

Most multi-model tools are switchers: the same chat, a dropdown to change who answers. The models never meet. Here they share one transcript, and the product is the friction between them plus the naming of it. Gemini's read of this in the venture's own strategy work was blunt - the API access is a commodity, so the defensible part is the orchestration and what the room learns about how models fail together.

The room is real. Real calls to real providers with your keys, the shared transcript, the typed catches, the challenge round, the decision brief, and durable storage for that brief. Lumi is live inside the venture as the host. What is not built: the session transcript is device-local, so it does not follow you to another browser, and there is no team or shared-workspace layer yet.

• Beta, and free to use with your own keys.

Free to try with your own key. One dollar a month for Drift Catchers itself - unlimited rooms, saved sessions, the decision layer - while you keep paying your own providers. Twenty-four dollars a month for the same product with model credits included, for people who would rather not hold provider keys at all. You pay for Drift Catchers; you pay separately for AI, or bring your own keys and skip that entirely.

Drift Catchers is in a live public first round, raising $250,000 - the software-track figure. It is a runway requirement rather than a valuation: the operating layer, the fundraising costs, the human hours, the AI and infrastructure, the marketing, the legal and tax reserves and an operating buffer, added up and published line by line. Offers are non-binding and start at $1: no card, no money moves, and nothing is binding until the round closes and the company is incorporated.

• No return is guaranteed. We recommend understanding the model, the business plan, and the risks before investing actual money. Consider independent professional advice if needed. Dream x Destiny does not provide investment advice.