Trust
AI reads. It never writes the record.
Bixel is built with modern models and built for them. The line is absolute: models structure what our machinery actually captured, and they generate nothing. Data comes from pages that exist, or it does not come at all.
What AI does here
Reads captured pages
Extraction models turn a captured pricing page or terms page into structured facts, each pointing back at the capture it was read from.
Classifies the corpus
Models help decide which companies belong in the tracked universe and which category a company sits in.
Summarizes, clearly labeled
Where a summary is derived rather than observed, it is stored and labeled as derived. It never masquerades as a fact.
What it never does
Never invents a value
No model fills in a price, a certification, or a headcount that was not on a captured page. If it was not observed or stated, it is not in the record.
Never papers over a gap
Where nothing was captured, the API answers that nothing was captured. An honest blank beats a confident guess, especially for the agents reading us.
Never breaks the chain
AI-assisted or not, every fact carries provenance, an as-of date, and a source URL, and resolves to a proof bundle you can verify without trusting Bixel.
The full pipeline, from capture to proof, is documented on the methodology page. Corrections follow the same discipline: fixes never rewrite history.