Independent verification for AI-generated science
Science is about to be produced faster than it can be trusted
AI agents now run entire analyses unattended, and their results are rarely checked by anyone independent. Genomarker checks them: it refuses what the data can't support, and signs the rest as evidence anyone can verify.
Catch Seal Replay
Genomarker Assurance StatementGM-2026-08-14-8F29E1Example
- Analysis
- Differential expression
- Assurance profile
- GM-BIOMARKER-DE-1.3
- Input integrity
- PASS
- Study design
- PASS
- Statistical methodology
- PASS · DISCLOSED FINDING
- Confounder control & leakage
- PASS
- Plan alignment
- PRE-SPECIFIED
- Execution & provenance
- VERIFIED · COMPLETE
- Independent replay
- REPRODUCED_EQUIVALENT
- Long-term reproducibility
- GRADE A
The problem
Results are trusted long before anyone re-runs them.
Of 27,271 code notebooks linked to biomedical papers, just 879 re-ran to their originally reported results (GigaScience, 2024). AI agents now multiply the volume — and every failure mode with it:
Methods that were never defensible
The wrong statistical test, an unsupported design, a violated assumption — chosen silently and reported confidently.
Confounding and leakage no one caught
Batch tracks treatment; test data leaks into training. The result looks strong precisely because it is wrong.
Analyses that drifted from the plan
What was pre-registered and what was run quietly diverge. Post-hoc becomes indistinguishable from pre-specified.
Provenance that says what, not how
Parameters unrecorded, software versions lost, the exact input data no longer identifiable.
AI decisions nobody can inspect
An agent made an inferential choice mid-analysis. The agent is gone. The justification never existed.
Results that expire with their software
Rerunnable today, unreproducible in five years. The environment decays; the claim stays in the literature.
Scientists, pipelines and, increasingly, AI agents — producing more analyses, faster and cheaper than ever.
The independent check: catch what the data can't support, seal what it can, replay it on demand.
Papers, biomarker programs, investment decisions and regulatory filings that depend on the computation being right.
The proof
We audited published science we had no part in.
Nine published gene-expression studies, re-run end to end on our production system with data and methods as published. We report what we found in the studies — and in our own platform.
9
Published studies audited
re-run on our production system
7
Re-executed end to end
1 refused with a stated reason · 1 not auditable from its deposit
57
Structured findings
1 hard refusal · 11 blocking · 39 warnings · 6 confirmations
1
Hidden confound caught — analysis refused
recorded in no metadata field; read from the raw instrument files
THE CATCH
Hidden in the raw files.
In a large public patient study, none of the 48 days on which samples were scanned held both patients and controls. No metadata field recorded it. Genomarker read the scan dates from the raw instrument files and refused the comparison: the data cannot separate the biology from the scan day.
A statement about what the public deposit supports — not about the published conclusions.
How it works
Catch. Seal. Replay.
One engine runs every check, whether a scientist, a pipeline or an AI agent submitted the analysis.
CATCH
Refuses what the data can't support.
Before anything runs, Genomarker checks whether the data can answer the question: batch confounding, pseudoreplication, the wrong method for the data type, deviations from a pre-registered plan. It refuses or flags — and says why.
SEAL
Signs what it can.
Inputs, code versions, parameters, rules and results are recorded and signed against published keys, so anyone can check the record without trusting us.
REPLAY
Re-runs the work independently.
We re-executed seven published analyses end to end on our production system, and our core statistics match independent R reference implementations.
Who it's for
Built for the people who carry the risk.
For investors & BD teams
Genomarker Diligence
Independent review of the computational claims behind a financing, licensing deal or trial decision — a written verdict, claim by claim, before you commit.
- Life-science investors
- Pharma BD & licensing
- Translational reviewers
For research teams
Genomarker Runtime
Run supported analyses with the checks built in. Every run produces a signed record, so your results are ready for a reviewer, a partner or a data room.
- Biotech R&D teams
- Translational research groups
- Core facilities
For AI-science platforms
Genomarker Assurance API
The checks your agents call before they claim a result — the same rules and the same evidence, whoever the actor is. Developer preview, mid-2027.
- AI-scientist platforms
- Pharma AI teams
- Workflow systems
What a statement says
Precise claims, not “certified correct.”
Scientific truth can't be read off a workflow, so Genomarker never issues a vague verdict. Each statement says what was checked, what passed, what was flagged — and what it does not establish.
| Assurance dimension | The question it answers |
|---|---|
| Input integrity | Do we know exactly what data entered the analysis? |
| Study design | Does the experimental design actually support the contrast being tested? |
| Statistical methodology | Were appropriate methodological safeguards applied? |
| Confounding & leakage | Were detectable confounders and information leakage caught? |
| Plan alignment | Did the analysis follow the pre-registered plan? |
| Execution & provenance | Did the requested computation run as specified — and can we reconstruct what happened? |
| Independent replay | Can it be re-executed independently of the original run? |
| Long-term reproducibility | Is enough state preserved to replay it years from now? |
Independent replay and long-term reproducibility are assessed as their engines ship (mid-2027); until then a statement marks them not assessed.
The north star
Results should not expire when their software does
Where this is going: a result sealed in 2026 that an independent researcher can still verify, reconstruct and replay in 2036 — without the original scientist, agent or software.
2026
Original analysis
- Analysis requested — by a scientist or an AI agent
- Design problem detected; execution refused
- Reason returned; plan corrected
- Analysis run; signed record issued
- Environment and dependencies preserved
- Evidence published as GM-2026-08-14-8F29E1
2036
Ten years later — no original researcher required
- Independent researcher retrieves GM-2026-08-14-8F29E1
- Verifies the signatures offline
- Inspects the original scientific decisions
- Reconstructs the computational environment
- Replays the computation
- Reproduced — equivalent result
Team
Who's building it.
Salar Sayyad
Founder & CEO
Built the Genomarker platform end to end, solo and full-time since Feb 2026 — five releases in under six months and a nine-study public audit run on production. 10+ years as an AI engineer, data scientist and product manager; BSc Computer Science, University of Toronto.
Dr. Saed Sayad
Scientific advisor · AI & bioinformatics
PhD in biochemistry & bioinformatics; adjunct professor at the University of Toronto and former professor of data science at Rutgers University, with 25+ years in machine learning and predictive modeling, focused on biomarker discovery and precision medicine. Inventor of the Real Time Learning Machine and creator of the widely used online text An Introduction to Data Science. Advises on statistical methodology and the rule library.
Dr. Mark Hiatt, MD, MBA, MS
Scientific advisor · clinical & regulatory
Stanford fellowship-trained physician executive in precision medicine — former VP of Medical Affairs at Guardant Health, SVP of Market Access & Strategy at BostonGene, and CMO at RadSite; 75+ articles and book chapters, 150+ conference presentations. Advises on clinical, regulatory and market-access context for evidence standards.
Proof before trust.
We're onboarding research teams, AI-science platforms and review clients in small cohorts. Bring us an analysis — or a claim you're about to bet on.