Hallucination and grounding
Numbers no code computed, genes and pathways that are not in the data, citations that do not exist, conclusions that outrun their evidence. We measure how often, and where.
Trust LayerScoped on request
Can you trust the AI in your research?
AI scientists, agents and models now design experiments, analyze data and write papers in biomedical R&D. The Trust Layer tests them independently for hallucination, security and scientific validity, under a protocol fixed in advance, with funding disclosed and negative results published.
Evaluation protocol
Pre-registered- System under test
- A single-cell foundation model (third party)
- Registered
- Before any data are touched
- Primary endpoint
- Cell-type annotation accuracy on held-out donors
- Failure threshold
- Fixed in the protocol
- Funded by
- Disclosed on the report
- Conflict check
- Not our product, not paid to build, no equity holder
Pass reportedFail reportedNull published
What we evaluate
- AI scientists and research agents
- LLM assistants and literature systems
- Biological foundation models
- Analysis pipelines and bioinformatics tools
- Laboratory automation and robotics
What we test
The ways AI fails in biomedical research and development, measured rather than asserted.
Security
Prompt injection hidden in papers, datasets and tool outputs; data leaking between projects or users; tools used outside their scope; secrets exposed.
Scientific validity
Pseudoreplication, the wrong unit of analysis, leakage between training and test data, benchmark contamination, and results that bend when a user asks for significance.
Reproducibility
Same input, same result: pinned versions, seeds and environments, and a third party who re-runs it.
Calibration and abstention
Confidence that matches the evidence, and a system that says the data cannot answer the question when they cannot.
Biosecurity and dual use
Whether a system lowers the barrier to harmful biological work, tested under a pre-registered protocol and disclosed responsibly.
Laboratory systems
Device limits, fail-closed behaviour, human approval and tamper-evident audit in automation built by others.
The report
The protocol, the funder and every result, with failure rates and exact confidence intervals. An independent scientific opinion, not a regulatory determination or a certificate of any kind.
How an evaluation runs
Each evaluation starts from the system's context of use and the risk if it is wrong, the same risk-based logic as the FDA's January 2025 draft guidance on AI in drug and biological product development.
- 1
Context of use
What the system is for, who relies on it and what an error would cost. This sets how strict the test is.
- 2
Pre-register
Endpoints, datasets and the analysis plan made public before testing starts.
- 3
Test
Seeded flaws with clean twins, adversarial probes and a blinded expert audit of claims.
- 4
Score
Failure rates with exact confidence intervals. Pass and fail are reported alike.
- 5
Publish
The full report, with the funder named and negative results included.
Four commitments, in the charter.
- Pre-registered
- Endpoints and analysis plan registered and public before data are touched. Deviations reported.
- Funder-disclosed
- Who paid, stated on every report. The client never approves conclusions.
- Nulls published
- A system that passes is reported as passing; one that fails, as failing. Neither is softened.
- Never our own products
- Not our products, not products we were paid to build, not products of any equity holder. Written into the charter of BioMedsAI, Inc.
We hold Padma to the same tests internally: numbers only from code, abstention when the data are silent, injection probes, identical re-runs. Those results are ours to report; the Trust Layer never evaluates our own products.
Education and capability transfer
Evaluation only matters if people can do it.
Workshops, AI literacy for scientists, reviewer training and capability transfer, so your team can use AI well and check it rigorously, in-house.
Request an independent evaluation.
Scope, funding disclosure and pre-registration are agreed in writing before any work starts.