The error that passes every review.
This research finds it.
ContextRx is an independent research project exploring how NLP and machine learning can detect contextual word errors in pharmaceutical documents and drug labels — correctly-spelled words used in the wrong clinical context. Eight years in development across five architectural iterations, now operating as a production system scanning FDA-approved drug labels and surfacing errors that pass through existing pharmaceutical quality systems, including regulatory review.
● Corpus: 34,201 FDA-approved drug label documents · Research use only
Advise to report any signs of infection immediately.
Titrate based on renal function and patient response.
In development, as an independent research effort
Architectural iterations, from rule-based to learned context models
FDA-approved drug label documents in the research corpus
vs. Grammarly's 7/10 on a curated pharma context-error benchmark*
* Internal benchmark across curated pharmaceutical context errors; methodology available on request. Results are research findings, not a validated regulatory control.
The errors that slip through every review
These aren't typos. They're correctly-spelled words in the wrong clinical context — invisible to spellcheckers, undetectable by text-comparison tools, and consequential when they reach a published label. The examples below are representative of the error class this research targets.
Advise patents to report any signs of infection immediately.
patents → patients
Patient instruction error — FDA deficiency letter risk
Titrate does based on patient renal function and weight.
does → dose
Dosing instruction corrupted — prescribing error risk
Contraindicated in patients with hepatic inflammation (Child-Pugh B or C).
inflammation → impairment
Wrong contraindication category — patient safety impact
Store at controlled room temperate (20°–25°C / 68°–77°F).
temperate → temperature
FDA-defined storage condition absent — recall risk
Contraindicated in patients with placenta group in clinical trials.
placenta → placebo
Clinical trial exclusion criterion corrupted
Administer 10 mg orally one daily with or without food.
one → once
Dosing frequency ambiguous — patient harm risk
Modeling pharmaceutical context, not spelling
The research question: can a model learn what a word should mean in a regulated clinical passage well enough to flag a real word that is wrong there? Five architectural iterations over eight years — from rule and n-gram baselines to learned contextual representations fitted on label text — converged on the pipeline below.
Ingest the document
Any pharmaceutical document — Package Insert, SmPC, Patient Leaflet, submission text. Processing runs locally; no document content leaves the machine it runs on.
Model the context
The NLP pipeline, built over five architectural iterations and fitted on a corpus of 34,201 FDA-approved labels, scores every token for semantic fit in its pharmaceutical context — clinical meaning, not spelling.
Surface candidate errors
Each flag is reported with the substitution candidate, the clinical risk category, and the surrounding context, so a human reviewer can adjudicate the finding.
Errors found in live, FDA-approved labels
The system currently runs in production against published drug labels. It identifies contextual errors that have already passed through pharmaceutical quality systems, including regulatory review — which is the central empirical finding of this project. Examples below are anonymized; findings are reported for research purposes, not as regulatory determinations.
Adults: take one tablet by mouth once every six ours, not to exceed 4 doses in 24 hours.
→ hours
Live in market. Surfaced by a research scan of the DailyMed corpus.
Avoid in patients with severe hepatic inflammation (Child-Pugh C).
→ impairment
Live in market. Surfaced by a research scan of the DailyMed corpus.
Store reconstituted suspension at controlled room temperate.
→ temperature
Live in market. Surfaced by a research scan of the DailyMed corpus.
Run the system on sample text
Paste a passage of label or regulatory text to see the context scan behave. Please use non-confidential text — this page is a hosted sandbox, not the offline research build.
Hosted sandbox. The research build runs fully offline with zero data egress.
Get in touch about the research.
If you work on labeling, regulatory writing, pharmacovigilance, or NLP and want to discuss the method, review the findings, or explore collaboration, send a note. Please don't include confidential or unpublished document content.
What to expect
- 1This is a research inquiry, not a sales conversation — nothing is being sold
- 2Reply typically within a few days; the project runs outside working hours
- 3Happy to share methodology, benchmark design, and error taxonomy
- 4Open to collaboration, peer review, replication, and dataset discussion
An independent project, run on its own time
ContextRx is a personal research project conducted independently of any employer, client, or sponsor. It is not a commercial product, it is not for sale, and its findings carry no regulatory weight. The architecture is offline-first by design so that experiments can run on sensitive text without any data leaving the machine. Findings are shared to invite scrutiny of the error class, not to make claims about any specific product or manufacturer.
Where existing tools stop, and this research begins
This project does not aim to replace any existing tool. It isolates a single error class that general-purpose spelling, grammar, and proofreading systems are not designed to detect, and asks whether it can be modeled.
| Capability | Microsoft Word | Grammarly | ContextRx (research) |
|---|---|---|---|
| Misspelled words | ✓ | ✓ | ✓ Included |
| Grammar & style | ✓ | ✓ | Not the focus |
| Version-to-version text comparison | ✗ | ✗ | Out of scope |
| Artwork & barcode verification | ✗ | ✗ | Out of scope |
| Pharma context errors (patents→patients) | ✗ Missed | ✗ Missed | ✓ Research focus |
| Fitted on FDA label text | ✗ | ✗ | ✓ 34,201 docs |
| Runs fully offline | ✓ | ✗ Cloud-only | ✓ Offline-first |
| Logged, reviewable findings | ✗ | ✗ | ✓ Full audit log |
| Benchmark context-error score* | n/a | 7/10 | 10/10 |
An open question, eight years in.
Contextual word errors reach published, FDA-approved labels today. This project exists to measure how often, and to test whether language models can catch them. Scrutiny, replication, and collaboration are all welcome.
Independent development, five architectural iterations
FDA-approved label documents in the research corpus
Offline research build — no document content leaves the machine