The actual model, running now
A simulation of how the model works
The model running here is the model reported in the article — its code was lifted verbatim from the research notebook, not rewritten. Every stage opens as it is passed, with its own numbers.
Ready to go — one click
Effectiveness and safety of opioids for dyspnea in patients with lung cancer: secondary analysis of multicenter prospective observational study
Taniguchi Y, Matsuda Y, Mori M, Ito M, Ikari T, Tokoro A, dkk. · Translational Lung Cancer Research 2022. Transl Lung Cancer Res 2022;11(12):2395-2402.
doi.org/10.21037/tlcr-22-512 · CC BY-NC-ND 4.0 · 8 pages · 251 KB
Why this article?
- Domain score
- 29 / 8
- Sentences into NER
- 214
- Distinct drugs named
- 9
- Baseline → advanced
- 68 → 34
- Rejected by meaning filters
- 45 · 4 reason codes
This article was picked from the 405 corpus documents on measured behaviour, not on how it reads: it scores far above the threshold, names nine distinct drugs, and halves cleanly from 68 pairs to 34 after the meaning filters — with rejections spread across all four reason codes. Every stage therefore has something real to show.
The figures on this card were measured during the corpus run. Running the same article on its own can give slightly different ones -- its domain score is 27, not 29 -- because the assembled text and the batch composition are not what they were when it was one document among 405. Both are correct; the context differs.
Upload one PDF
Or bring your own paper
Upload one PDF
The same model, run on an article you choose. The file is processed and then deleted; nothing is kept.
PDF · 10 MB maximum · 60 pages maximum · one file
Why only one file?
One file, not several — and the reason is in the data itself. Of the 263 articles that pass the domain filter, only 63 contribute a single operational relation: three in four relevant papers yield nothing. The final gate also requires a pair to appear at least twice, and 57 of the 76 relations only reach that across more than one article. Simulated over this corpus, 2,000 draws: uploading five PDFs yields 2.2 relations on average, after six minutes of waiting. The complete map of all 76 is already waiting on the Results page.
Computed in advance
Six scenarios, no queue to wait in
One document is processed at a time, so uploads are refused while the service is busy. The six sentences below were run beforehand — inside the same image, by the same pipeline — and show exactly the same decisions without queueing.