A model reads the open cancer literature as it is published, reasons about mechanism, and builds a cure dossier per cancer — continuously, unattended, and in the open. Every trace on this page is unedited and unretried.
Roughly 1.5 million papers on cancer sit in the open literature, and several thousand more land each week. No individual can read them; the connections that matter — a resistance mechanism in one tumour type, an approved inhibitor of that exact node in another — are routinely separated by a field boundary, a vocabulary, and a decade.
SCIENCE is a continuous, unattended attempt to work that seam. Each visit takes one cancer, pulls a fresh evidence pack — recent papers from Europe PMC, active interventional trials from ClinicalTrials.gov, gene anchors from the Monarch Initiative — and hands it to a frontier model together with everything the desk already believes about that cancer. The output is not a cure. It is a testable increment: a hypothesis specific enough that a bench scientist can say what experiment would kill it.
Everything published here is unreviewed model output, generated for human review. It is not medical advice, not a diagnosis, and not a treatment recommendation.
A visit is deliberately narrow. The model receives the evidence pack and its own dossier — open questions, active leads with their citations, what it has already retired — and must integrate one into the other: strengthen, weaken, retire, or open leads, then write the literature queries for its own next visit. Memory is the dossier, so the reasoning compounds across visits rather than restarting.
Every claim traces to a paper or trial in the pack, cited as PMID, NCT,
or DOI — or it is tagged [KNOWN], [INFERRED],
[SPECULATIVE]. Citations are checked against the pack; anything else is flagged
UNVERIFIED in the ledger.
A paper that undercuts a lead moves it down or kills it. Negative trials are recorded so the desk cannot re-propose them. Retiring a lead on good evidence is a success of the method.
One attempt per visit. No retries for a better answer, no curation of the trace. Refusals, truncations, clamps, and failures print to the terminal exactly as they happen.
Findings feed forward. Each visit closes by naming the questions that should drive the next one, so a programme deepens along its own evidence rather than down a fixed list.
The failure mode of any hypothesis generator is confident volume. A lead is cheap to write and cheap to believe. So promotion is not a judgement the model gets to make about itself — it is a mechanical consequence of independent literature arriving on a later visit.
A coherent therapeutic hypothesis with at least one verified citation. Cheap to create, cheap to kill.
Corroborated by new, independent literature on a later visit — at least two evidence-bearing visits and two verified citations.
Corroborated repeatedly — three or more evidence-bearing visits, four or more citations — and survived the desk's own attempts to kill it. It should be rare.
The engine enforces this, not the prompt: a claimed level without the corroboration behind it is clamped back down, and the clamp prints to the public terminal.
Chronic myeloid leukemia sits in the programme as a control, revisited on a fixed cadence. CML is solved: a BCR-ABL fusion drives it and imatinib-class inhibitors treat it. When the desk reaches it, the trace can be read against a known answer — did it reason from the evidence to kinase inhibition of the fusion, or recite what it already knew? The control validates no individual lead. It calibrates how much weight to put on the ones about cancers with no answer to check against.
The desk runs continuously, one cancer per visit, and every step is broadcast as it happens — literature and trial fetches first, then the model's reasoning streamed token by token, then the ledger moves. Nothing is edited, curated, or retried for a better answer.
Europe PMC for the literature — titles and abstracts, recency-weighted, with the desk's own fielded queries. ClinicalTrials.gov v2 for active interventional trials, their phase, status, and intervention arms. Monarch Initiative v3 for the gene anchor, distinguishing causal (germline-leaning) from correlated (somatic-leaning) associations. All three are public and keyless; every request prints to the terminal.
Claude. The evidence pack plus the cancer's living dossier are composed into one prompt per visit; the full trace is streamed to the live terminal and the ledger.
Every finding carries the block needed to replay it by hand: the sources, the exact queries, the identifier of every document shown to the model, the model that served the visit, its effort setting, and token usage — the Claude for Science standard, so a human researcher can pick up any visit and check it.
/events — SSE stream of the live terminal