A 405-article corpus
Findings
The numbers on this page come from the research run, computed on GPU fp16. The live simulation on the Simulation page runs on CPU fp32. Both were tested on the same slice and came back with zero difference across three axes — extraction fidelity, numeric precision, and batch composition. The coverage of that test is stated plainly: 1 document of 263, and 1,520 of 92,713 causal sentences, which is 1.6% of the corpus. A full-corpus CPU pass has not been run, so whether 76 stays 76 on CPU is still open.
Measured, not claimed
Accuracy against the pharmacist's annotations
Measured on the random stratum only — the active stratum was chosen by uncertainty and is therefore not representative.
Precision
0.94
Recall
0.73
F1
0.82
Correct / wrong / missed
16 / 1 / 6
95% confidence interval for F1: 0.667 – 0.939. The interval is wide because there are few observations in this scope. Reporting F1 without it would hide how little data supports it.
On the same scope, measured per pair, the baseline path without meaning filters reaches only precision 0.4286 with 20 errors. That gap is the measured basis for separating the operational table from the candidate table.
The headline figure above is measured per sentence, while this baseline row is per pair. On genuinely matched units, pair-level, the advanced path reaches precision 0.5385 against 0.4286 for the baseline. Both are floors rather than estimates: the gold set samples sentences, so a pair is called wrong only because its few sampled sentences happen not to assert it. A per-sentence baseline row was not computed in this study.
Three questions, three diagrams
Sankey diagrams
Three diagrams answer different questions. One is not enough: statistical evidence and clinical priority do not point the same way, because the association measure is lowest exactly on the most frequent and most fundamental relations.
Filters76 of 76 relations
Filters
76 of 76 relations
≥ 0.108
opioidopioids · 16 side effects
morphineopioids · 10 side effects
oxycodoneopioids · 9 side effects
fentanylopioids · 4 side effects
methadoneopioids · 3 side effects
tapentadolopioids · 3 side effects
tramadolopioids · 3 side effects
codeineopioids · 1 side effects
hydrocodoneopioids · 1 side effects
hydromorphoneopioids · 1 side effects
dexamethasonecorticosteroids · 1 side effects
olanzapineantipsychotics · 3 side effects
haloperidolantipsychotics · 1 side effects
benzodiazepineother · 8 side effects
naloxoneother · 4 side effects
methylnaltrexoneother · 3 side effects
cannabidiolother · 2 side effects
naldemedineother · 2 side effects
midazolamother · 1 side effects
Choose a relation on the map, the diagram or the table to see its care plan and its evidence.
Candidate tier — 160 pairs
Raw co-occurrence, unfiltered and unevaluated. Kept deliberately for expert review. It still contains relations that are clinically inverted.
Their identities are not shown on this site until expert validation is complete — publishing an unreviewed list is the same as circulating it.
Coverage of the required drug classes
| Class | Operational | Candidate |
|---|---|---|
| opioids | 51 | 63 |
| antidepressantscandidate only | 0 | 10 |
| corticosteroids | 1 | 7 |
| antipsychotics | 4 | 13 |
| other | 20 | 67 |
The coverage is uneven, and that is reported as it stands: the corpus is genuinely thin for classes other than opioids.