We’re curing cancer on top of research we can’t trust.
When Amgen tried to reproduce 53 landmark cancer studies, only 6 held up. Litmus reads a paper and tells you which results will hold, before you bet ten years and a billion dollars on them.
It starts with cancer. It ends as the verification substrate every AI scientist has to run through.
STK33 silencing is selectively lethal in KRAS-mutant tumours
J. Oncogenic Signalling · 2016 · preclinical
- p-value recomputes to .082, reported .008
- n = 4/group, powered only for d ≥ 1.5
- independent group found the target dispensable
Every flag is clickable to the exact sentence it came from. Nothing the auditor can’t point to survives.
The most expensive failure mode in the most expensive industry
A $3-trillion enterprise, built on a literature no one verifies.
Read those together: the single most expensive failure mode in the most expensive industry on earth is building on results that were never true, and today, nothing catches it early.
Time isn’t money. Time is lives.
We find out at the worst possible moment.
A target built on irreproducible biology doesn’t fail on day one. It fails in Phase II or III, years and hundreds of millions later, after patients have waited for a therapy that was never going to work.
Litmus turns a years-later, $100M failure into a day-one triage decision: this result is solid, that one is standing on sand, verify these three things first.
- Day 0Target selection
Cheapest decision to change. Highest leverage in the whole pipeline.
- Year 2–3Preclinical validation
The irreproducible result quietly propagates.
- Year 6Phase I
Safe, but built on the wrong hypothesis.
- Year 10+Phase II/III failure
$100M+ gone. Patients waited a decade for nothing.
What Litmus does
One trust layer. It reads a paper the way your most sceptical colleague would, at scale.
Statistical forensics
statcheck, GRIM, GRIMMER, SPRITE, power and p-curve, run in code, never by a model. If a reported number is arithmetically impossible, we prove it.
The rest of the literature
For every central claim we retrieve related work and weigh it, actively hunting the strongest disconfirming evidence, not the confirming kind.
Grounded, calibrated verdicts
A replication likelihood you can trust: every reason points to a source span, calibrated so 70% means 70%, and it abstains when the basis is thin.
Nothing ungrounded survives
Every number is clickable to the evidence it came from.
If a flag can’t be traced to an exact span in the source or a real external paper, Litmus drops it. That single rule is what separates an auditor you can defend in a boardroom from a confident-sounding guess.
- Character-level citations back to the source
- The grounding guard doubles as a prompt-injection circuit breaker
- Adversarial refuters try to break every high-severity finding
“Viability of KRAS-mutant lines was significantly reduced relative to controls (t(12) = 1.9, p = .008).”
t(12) = 1.9 → p = .082; reported p = .008
Recomputes to non-significant. The reported value flips the finding.
Why it’s a platform, not a tool
A toll booth on a $3-trillion road.
Litmus doesn’t need to capture a trillion dollars. It sits on the trust decisions inside the R&D economy and charges a slice of the waste it prevents.
And the timing: the whole world is racing to build AI that reads the literature and does science, and none of it knows which literature is true. Feed it the unverified corpus and it amplifies the garbage at machine speed. The missing piece is a trust layer they can query. A credit bureau for scientific claims.
- 1Cancer preclinical target validationBeachhead
- 2All preclinical & biomedical research
- 3Pharma R&D, target selection & asset diligence
- 4Evidence-based medicine & guidelines
- 5Funders & publishers, triage & screening
- 6The verification API every AI scientist calls
The trillion-dollar framing is the arena, not a revenue claim. The anchors under it, $28B/yr waste, $2.6B/drug, 90% clinical failure, 89% non-replication, are all published and cited. A trust company that overstates its own numbers is dead on arrival. Ambition, grounded.
Which papers are real? Find out before you build.
Feed Litmus a known-failed cancer paper and watch it flag the math, surface the contradicting literature, and score it low. Then feed it a solid one.