Green checks, zero values: scraper monitoring needs assertions, not liveness

The failure has a shape worth naming: the field still fills, and what fills it coerces to zero ^[5]. A null check has nothing to bite on. Neither does anything else on the standard monitor list, because every one of the five questions it asks answers yes while the incident is running ^[2][18], and zero downloads is a legal reading of the world ^[7]. The only thing that moved is the shape of the values, which is why a collapsing median reads differently from an empty result ^[6][8].

The arithmetic is worth doing. Two bad fields out of eight is a quarter of the column set wrong ^[19]. The six that survived are 75 percent of it, the same figure the run carried as its score while its status read BROKEN ^[20]. The write-up does not say the score is the fraction of healthy fields, so that is arithmetic rather than documentation ^[20]. Either way, a scoring scheme that lands a quarter-wrong run in the seventies is not something to page on.

The part of this architecture worth copying is not the healer. It is packages/health: pure drift detection, rows in and a verdict out, with no I/O, clock, randomness or network ^[14]. The heal path cannot have that signature. It is I/O bound through the Bright Data CLI ^[13], it stops at a human approval gate ^[10], and by the author's own account it exists only because Scraper Studio can heal an existing collector while preserving the Collector ID that downstream systems already depend on ^[12]. Detection is portable. Recovery is purchased.

The diagnosis step is the second transferable idea. The repair instruction is generated from the drift evidence itself, and it names the untouched fields explicitly so the healer leaves them alone ^[11], in a prompt that is pure and capped at a thousand characters ^[16]. Scoping a repair to the two fields that actually moved, rather than regenerating the extractor, is the difference between a small incident and a larger one ^[11].

What the detection half costs is state. The verdict is computed against prior typical values ^[6], so something has to retain them, which rules out a stateless per-row assertion ^[21]. It also rules out the lazy version of the rule: a hard floor of "greater than zero" fires on any genuinely new item, while a median over a window does not ^[7][21].

One caveat about the evidence. This is a single developer's chaos test against his own fixture site, with two numeric fields deliberately relocated ^[3][17]. That is the friendly case, because the whole field collapses and the median hits the floor. The expensive real case is partial: a selector that still matches half the rows, a median that sags by a third instead of going to zero. The write-up reports a verdict and a score, not the rule that produces them ^[4][20].