Data source

Each collector is a Bright Data Scraper Studio template. When a page's structure drifts, bdheal repairs it in place — the collector ID never changes, so nothing downstream notices.

Controlled benchmark

medicines.org.uk cannot be broken on cue, so the repair loop is measured against a fixture site we control, whose layout can be mutated to order. Its results are kept in a store of their own, deliberately: a benchmark must not be able to disturb the collectors that serve the corpus.

9 mutations over 19 attempts: 8 behaved as expected. 7 of the 8 that should raise a signal did, and 1 that should raise none stayed silent. 1 repaired, 1 verified against the old layout. Open: url_pattern — expected to raise a signal and did not. Every mutation is listed below. A signal chip is filled when that detector fired, outlined when it was expected and stayed silent. Non-regression means the repair still works against the layout it was built for, so it was not overfitted to the break.

Mutation	Signals	Outcome
class_rename	row_count expected but silent skeleton expected and fired	detected, not repaired
column_reorder	skeleton expected and fired	detected, not repaired
date_format	null_rate expected and fired schema expected and fired	detected, not repaired
field_split_merge	skeleton expected and fired	detected, not repaired
link_label	none declared	declared gap · no signal expected
pagination	row_count expected but silent schema fired but never declared skeleton expected but silent	detected, not repaired
table_to_div	row_count expected and fired skeleton expected and fired	detected, not repaired
url_pattern	null_rate expected but silent schema expected but silent	not detected
wrapper_nesting	skeleton expected and fired	repaired · accuracy 1 · non-regression passed