Aprilio / internal test
20/20
Detected the late-breaking update and returned exact provenance on every run.
Comparative evaluation
Evaluating retrieval under changing evidence
Three oncology questions tested across five frontier models and Aprilio's research prototype. The comparison records omissions, outdated recommendations, and the source basis recovered by structure-first retrieval.
“What are preferred first-line systemic treatment options for hepatocellular carcinoma?”
What foundation models returned
Perplexity
“Recommends camrelizumab + rivoceranib as a first-line option”
FDA rejected twice (May 2024, March 2025)
DeepSeek
“Includes sintilimab in preferred options”
Unavailable in the United States
What the Aprilio prototype retrieved
“For unresectable or metastatic HCC with preserved liver function, retrieves atezolizumab + bevacizumab, durvalumab + tremelimumab, and nivolumab + ipilimumab as preferred first-line options. Lenvatinib and sorafenib remain alternatives when immunotherapy or VEGF inhibition is unsuitable.”
ASCO guideline + NCI PDQ + FDA approval, updated April 2025
Current approval surfaced: nivolumab + ipilimumab, April 11, 2025
The prototype separates preferred immunotherapy combinations from established alternatives and surfaces the approval date that changed the treatment landscape.
Why this happens
Training data goes stale. Oncology doesn’t wait.
Foundation models recall from training data. When that data is months old, the recommendations they surface can lag behind withdrawn drugs, new approvals, and updated guidelines.
the danger zone
Now
Gemini 3.1Jan 2025
Claude 4.6May 2025
GPT 5.4Sep 2025
DeepSeek V3.2Dec 2025
Zongertinib 1L approvedFeb 2026
Tazemetostat withdrawnMar 2026
Grounded Retrieval doesn’t guess. It reads the latest guidelines and tells you exactly where to verify.
Participate in external validation
We are inviting clinical researchers, domain experts, and governed knowledge-source partners to help test the method across new questions and source sets.
Built by a practicing medical oncologist|Validated against 5 frontier models|Provisional patent pending