Content Accuracy Monitor

An article can remain online long after its facts change. Across a collection of more than a thousand articles in Santander Salto, finding every outdated date, incorrect value, or subtle inconsistency by hand becomes a herculean task.

I built the Content Accuracy Monitor (CAM) to help editors find those problems. It detects changes to facts in published articles, finds outdated information and inconsistencies, connects findings to supporting sources, and proposes corrections for human review.

My responsibility spanned the product’s architecture, backend and frontend implementation, deployment, and ongoing operation. The editorial team retains the final decision about what is correct and what gets published.

Early versions exposed a model-capability problem. Broad instructions to search for updated information were unreliable with smaller models. More capable models handled the research better, but were too expensive to scale.

I refined the search and verification workflow to make lower-cost models effective. I changed the investigation harness around the models so the work could be done at a viable cost across more than a thousand articles.

A fact can be easy to check without being useful to audit. Early extraction would select incidental details, such as Cristiano Ronaldo’s Instagram follower count, even when they were irrelevant to the editorial task. Every unnecessary claim created more research work, more token cost and another result for an editor to review.

A separate failure mode appeared when the system treated differently worded yet equivalent information as a contradiction.

These failures shaped what the system needed to distinguish: a relevant factual problem, an actual change in information, and a difference that does not justify a correction.

I added reporting controls that let editors flag model errors during their normal review workflow. The system uses these real examples for prompt improvement optimization: irrelevant claims, false differences, and other findings that did not serve the editorial task.

Editors can inspect the evidence and proposed revision, make their own changes, and decide what is ready. All of these human corrections are used to train the system.

The result is an editorial tool that helps people decide where to investigate and what to correct, with a feedback loop that makes their use of the product part of its continued improvement.