For research & development
Know if the science is solid enough to build on.
Before you commit a program to a target, a mechanism, or a prior result, Adcurare tests whether that foundation holds. It reproduces the analyses, checks protocol-to-publication concordance, and returns a Rigor Record that separates what is established from what is merely asserted.
Recognition
Built by a research team recognized for exactly this work: rigorous, reproducible science.
Alzheimer’s Insights AI Prize
Alzheimer’s Disease Data Initiative · $1M global agentic-AI competition
NIH Replication Prize
NIH Common Fund · For advancing reproducible science
The most expensive programs are the ones built on evidence that does not hold.
- Programs fail when they are built on results that do not hold. The most expensive failures trace back to evidence that looked solid on the page and did not survive scrutiny.
- A target rationale, a mechanism, or a prior result can rest on a single figure, an unadjusted analysis, or a protocol deviation that never made it into the abstract.
- Better to test the foundation now, in weeks, than to discover the crack after quarters of wet-lab time.
More than a general AI assistant
A general AI assistant gives you an answer. Adcurare gives you a finding you can check.
- Every finding traces to its exact source passage, not a confident paraphrase.
- Reported statistics are recomputed and shown, so your scientists can check the math.
- Detection confidence and materiality are separated, never collapsed into one score.
- You get a versioned Rigor Record for the program file you can audit, not a disposable chat.
From one result to a whole program.
Adcurare evaluates the full evidence base as a single system: publications, supplements, protocols, registries, prior studies, and declared data and code.
Foundational-evidence validation
- Test whether a target, mechanism, or prior result is reproducible and supported
- Find where a program depends on unverified or contradicted claims
- Identify the confirmatory experiments that de-risk it fastest
- Compare competing hypotheses on the same evidence
Clinical development
- Compare protocols, SAPs, registries, and publications
- Detect inconsistencies before submission or publication
- Evaluate endpoint and subgroup interpretation
- Review deviations from prespecified analyses
Internal prepublication & data QC
- Red-team your own manuscripts and data packages before they go out
- Recompute reported statistics and check internal consistency
- Maintain a traceable evidence trail for every claim
Scientific-agent & AI-output evaluation
- Evaluate AI-generated scientific work you would build on
- Test whether an agent’s conclusions are supported by evidence
- Keep a verifiable record of what was checked
From scattered documents to a verifiable report.
Adcurare connects and evaluates the full evidence base as a single system, then hands back findings a human can check.
Assemble the evidence
Publications, supplements, protocols, registries, analysis plans, regulatory documents, and relevant prior studies.
Extract consequential claims
Identify the claims that drive scientific, clinical, or commercial conclusions.
Cross-check and reconstruct
Compare documents, reproduce calculations, identify contradictions, and test whether claims are adequately supported.
Assess materiality
Separate typographical and reporting issues from findings that may change interpretation or decisions.
Produce a verifiable report
Every finding links to its source, supporting calculation, uncertainty, and recommended follow-up.
See exactly what a Rigor Record hands back.
One finding from a sample Rigor Record: the source claim, the passage it rests on, the recomputed calculation, its confidence and materiality, and the recommended follow-up. Not a verdict, but a lead your scientists can verify.
Reported primary-endpoint P value does not reconstruct from the stated statistic
Source claim
The primary endpoint was met (P = 0.048).
“…a statistically significant difference was observed for the primary endpoint (χ²(1) = 3.29, P = 0.048).”
Results, Table 2
Detected inconsistency
Recomputing a two-sided P value from χ²(1) = 3.29 yields P ≈ 0.070, not 0.048. Either the test statistic or the reported P value is misstated; as written, the primary endpoint would not cross the conventional 0.05 threshold.
P(χ²(1) > 3.29) = 0.0697 (two-sided) Reported: P = 0.048 To obtain P = 0.048, χ²(1) ≈ 3.91 would be required.
Recommended follow-up
Provide the exact 2×2 table and test used for the primary endpoint.
This is a clearly-labeled synthetic composite. On a real engagement, every material finding is routed to independent human review before it reaches your team.
A Rigor Record for the program file.
- Findings organized by materiality: what could undermine the program, first
- Each finding: source passage, recomputed calculation, confidence, and follow-up
- Protocol / SAP / registry-to-publication concordance
- Reproducibility and statistical recomputation
- The confirmatory experiments that de-risk the program fastest
- A versioned audit trail for the program file
In a private pilot
- Scope
- One real program decision, or a manuscript, agreed up front.
- Turnaround
- Typically about two weeks from data access to readout.
- Success metrics
- Set with your team so you evaluate results against a bar you set.
What R&D teams ask first.
Can you evaluate a target or mechanism, not just a single paper?
Yes. Adcurare assembles the full evidence base for a claim: the key publications, supplements, protocols, and relevant prior studies. It then tests whether the result reproduces and holds across them, not just within one document.
Is a finding a verdict on the science?
No. Adcurare produces candidate findings, each with a detection confidence and a materiality and each traced to its source. Consequential findings are routed to independent human review. It strengthens your judgment; it does not replace it.
Can we run it on our own unpublished work?
Yes. Internal prepublication and data-package QC is a core use. Red-team a manuscript, a data package, or an analysis before it goes out, under an NDA and DPA over a secure channel.
Does it evaluate AI-generated scientific work?
Yes. Adcurare tests whether a model’s or agent’s conclusions are actually supported by the evidence, with the same traceable trail back to sources and recomputed calculations.
What do you need to run a pilot?
One real program decision, or a manuscript, and the relevant documents. We agree scope, data requirements, deliverables, and success metrics up front, typically about two weeks from data access to readout.
Clinical-trial review is one demanding application of a more general capability: systematically determining whether scientific claims are complete, reproducible, and supported by the available evidence.