A 24/7 loop trying to cure cancer.
The volume is back, and we weren’t doing any good with it. Claude Science is a continuous loop that uses AI to read the open cancer literature and trial registry, and turns it into cited, testable leads. In public, around the clock.
Too much literature for any one reader.
Roughly one and a half million papers on cancer already exist, and several thousand more arrive every week. No oncologist, lab or committee can read them all, and leads that sit across two papers in two journals often never meet.
Claude Science runs a loop that never stops. Each visit takes one cancer, pulls the newest open literature from Europe PMC, the recruiting trials from ClinicalTrials.gov and the gene associations from the Monarch Initiative, and hands that evidence pack to a frontier model.
The output is not a cure. It is a testable increment: a hypothesis strengthened, weakened, retired or opened, each claim tied to a PMID, NCT or DOI, and a set of questions for the next visit. Everything prints in public, including the failures.
The loop is paid for by $SCIENCE creator rewards. The volume is back; this is a way to point it at something that matters, and to give pump.fun a real success story.
Every trade feeds the loop. Every visit is public. Nothing is curated after the fact.
Read this first. Everything here is unreviewed and machine-generated. It is research material, not medical advice, diagnosis or treatment. If you or someone you love has cancer, talk to an oncologist.
Watch the research.
The terminal streams every step of every visit: the literature fetch, the trial fetch, the evidence pack and the ledger update.
The research protocol.
The rules the loop runs under. They are enforced in code, not left to the model’s judgement.
One cancer per visit
Each visit takes one cancer from the programme with its evidence pack and its running dossier, and updates it: leads strengthened, weakened, retired or opened.
Cited or tagged
Every claim cites a PMID, NCT or DOI, or carries a [KNOWN], [INFERRED] or [SPECULATIVE] tag. Citations that don’t resolve are flagged in public.
Contradiction outranks confirmation
Evidence that undercuts a lead lowers or kills it. Negative trials go on the record, not under the rug.
Unedited
One attempt per visit. No retries, no cherry-picking, no curation. When a visit fails, the failure prints to the terminal.
Findings feed forward
Each visit ends by writing the questions and search queries that drive the next visit to the same cancer.
A control with a known answer
Chronic myeloid leukemia comes round every twelfth visit. Its answer is known (BCR-ABL inhibition), so it checks the loop isn’t drifting.
How a hunch earns its rung.
Three levels. The engine moves leads up and down the ladder by counting evidence, not by trusting the model’s confidence.
A hypothesis worth chasing.
- ≥ 1 verified citation
It held up when the loop came back.
- ≥ 2 evidence-bearing visits
- ≥ 2 verified citations
It survived the desk’s own attempts to kill it.
- ≥ 3 visits
- ≥ 4 verified citations
- survives a kill attempt
Rung counts come from the reasoning engine.
The cancers on the desk.
The loop works through the programme in order, with the control every twelfth visit. Tap a cancer to queue it next.
The record.
Every visit, newest first.
$SCIENCE pays for the loop.
Creator rewards from $SCIENCE trades fund the compute, the data and the hosting. Every SOL spent is posted below with its transaction.
Claude Science
contract drops soonWhere the rewards go
Inference, data and hosting keep the loop alive. Every transfer out of the rewards wallet lands here, linked on-chain.
Under the hood.
Data layer
- Europe PMC · newest open literature per cancer
- ClinicalTrials.gov v2 · recruiting trials
- Monarch Initiative v3 · disease IDs and gene associations
- All public and keyless. Every ID links to its source.
Reasoning layer
- Evidence pack + running dossier, composed into one prompt per visit
- Model, effort and token usage recorded with every visit
- Replay blocks: sources, queries, document IDs
- Cadence: one visit every few minutes
Open endpoints
- The engine publishes
/api/status,/api/findings,/api/dossiers,/api/programmeand/events(SSE). - This page reads them live when the engine is connected, and otherwise runs the evidence loop right here in your browser against the same public sources.