What validation forbids

No validation check establishes truth. It establishes non-refutation against an expectation, and the expectation is an artifact with an author, a version and a rate of decay.
Per-value checks verify form but never existence. The ceiling is hard and is reached cheaply, which is why coverage numbers in quality tools are systematically misleading.
Absence cannot be detected by any record-level check. So the empty-output failure is invisible to most check suites.
You cannot guarantee completeness and stay available under partial failure. Quarantine and block are the two horns, and there is no third.
No single gate level is correct for several consumers with different cost asymmetries.

The escapes, and their prices

Escape	Price
Move the referent outside (reconciliation, The reconciliation ladder)	moving or hashing the other side's data
Move validation upstream to the producer	highest leverage per check, but costs organizational negotiation instead of compute, which is why it is rare
Sample instead of scanning	detection probability for rare defects. The arithmetic is unkind: a defect rate of 10⁻⁴ needs tens of thousands of sampled rows to see it even once (The reconciliation ladder)
Auto-derive expectations	intent: you get checks nobody can adjudicate
Quarantine	invisible bias
Block	availability, plus permanent headroom

Case, resolvedConstructed case · Ravi at Freshlane

1,000 orders set aside, and revenue 4% low

A new check quarantined 1,000 of 100,000 orders and the batch published with revenue low, and blocking the pipeline on any failure stalled it for six hours and left the warehouse slow for a day. Where it started: Gates, quarantine, stop-the-line →

The quarantine bias. Ravi adds the quarantine rate to the published table's metadata, broken down by region, by app version and by basket size. Anita's dashboard now shows "1.0% of rows quarantined; 4.2% of revenue". He also fixes the parser, so the long-basket orders are accepted. He sets a rule: if quarantined revenue exceeds 1% of the batch, the check escalates from quarantine to block for the finance report only.

One gate, many consumers. He cannot give one gate level to all consumers. He writes a small data contract per dataset: the finance report blocks on failure; the dashboard quarantines and shows the quarantine rate; the recommendation feed warns and keeps serving the previous day's data. He also changes the stop-the-line scope to one partition (one store region, one day), so a bad region does not stop the whole company's data. And he sets autoscaling of the pipeline so that normal utilization is 50–60%, which makes a six-hour stop recoverable in 6 hours (at 50%) to 9 hours (at 60%) instead of 54.