An AI trying to
find a cure.

Rare disease is an economics problem wearing a biology costume.

There are more than 7,000 recognised rare diseases and roughly 300 million people living with one. The overwhelming majority have no approved therapy. Not because the biology is uniformly intractable — because a few thousand patients cannot repay a decade of development. For most of these conditions, nobody has ever sat down and written a serious mechanistic hypothesis.

Language models can write that hypothesis in minutes. The problem is that they can also write a beautiful, confident, completely wrong one, and a plausible falsehood in medicine is worse than silence. Fluency is not correctness, and a single model reviewing its own work grades itself generously.

The output is not a cure and does not pretend to be. It is a testable increment: a hypothesis specific enough that a bench scientist can say what experiment would kill it, and an explicit record of what already died.

Eight enter. Most do not come out.

Each cycle pulls one disease from the Monarch Initiative knowledge graph — causal genes, phenotype spectrum with frequencies, gene pleiotropy, and phenotype-matched model organisms. The challenger enters six to eight candidate therapies. Then the killing starts.

A kill is permanent — there is no appeal and no revision. What reaches the archive has been attacked four separate ways by a model that was rewarded for destroying it. When nothing survives, that is published as a wipeout, because a disease where every plausible idea dies is a real result and worth knowing.

Every claim is tagged at the source: [KG] from the knowledge graph, [KNOWN] from literature, [INFERRED] reasoned, [SPECULATIVE] flagged as unsupported. The gauntlet checks the tags too — inference dressed up as established fact is grounds for a kill.

How you know it is reasoning and not reciting.

A model that has read the literature can produce a confident, correct-sounding paragraph about a disease without deriving anything. That failure mode is invisible when every disease in the set is unsolved, because there is no answer to check against.

Spinal muscular atrophy sits in the set as a control. SMA is solved — nusinersen, risdiplam and onasemnogene abeparvovec are approved, and the SMN1/SMN2 mechanism is textbook. When the gauntlet reaches it, the trace can be read against a known answer. Did it reconstruct the SMN2 copy-number logic from the graph, or restate what it already knew?

The control validates nothing on its own. It calibrates how much weight to put on the diseases where there is nothing to check against — which is the entire rest of the set.

The argument happens in public.

Both models stream token by token as they work — the researcher's draft, then the critic tearing into it, then the verdict. Nothing is edited, curated, or retried for a better answer. Refusals, truncations and failures print exactly as they occur.

What is in the queue.

Every disease resolves to a MONDO identifier against Monarch before entering the queue. The set grows on its own — the researcher proposes related disorders, and they join the frontier.

Everything that survived. And everything that didn't.

Newest first. Click any entry to expand its surviving hypotheses and the ones the critic killed.