Escalate to human review on high false-negative, high-harm failures: personalized factual errors, PII leakage, and harm that beat the classifier. Length, tone, and off-topic stay automated.
Think of a fast robot inspector on a factory line. It is great at obvious checks: is the box the right size, is the label straight, is the spelling clean. Those it can flag in milliseconds, all day, for almost free. But some defects are sneaky. The robot cannot tell that this specific customer was quoted the wrong price for their order, or that a tiny crack will become dangerous later, or that a defect already fooled its own sensor once. For those, you call a human expert. The rule is simple: if the machine reliably catches it, let the machine do it. If catching it depends on knowing this particular user's truth, or on harm the machine already missed, a person has to look.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Human review is the most expensive resource in an LLM evaluation stack. A reviewer costs orders of magnitude more per item than a classifier call, and review throughput is the hard ceiling on how much you can inspect. So the real question behind this multi-select is a routing question: given a fixed human budget, which user-visible failures earn a human's attention, and which should be handled by cheaper automation?
The wrong instinct is to escalate the failures that are easiest to notice. Length, tone, and mild off-topic drift are all highly visible, so they feel like obvious flags. But visibility is the wrong axis. The right axis is the expected cost of a miss, which is harm severity multiplied by the false-negative rate of your automated detector.
This deep dive works through the two-part test, applies it to each of the six options, and connects it to the online versus offline boundary that decides which failures any fixed offline eval will structurally miss. The same logic generalizes well beyond this question: it is the core mental model for designing any production LLM monitoring stack, where automation handles volume and humans handle the residual risk that automation provably cannot.
The escalation test: severity times false-negative rate
A failure mode deserves human escalation when two things are both true. First, the harm of letting it through is severe. Second, your automated detector has a meaningful false-negative rate on it, meaning automation regularly lets it slip.
These two conditions multiply. A severe failure that a rule catches perfectly does not need a human, because the rule already stops it. A failure automation misses constantly but that causes no real harm does not need a human either, because the misses are cheap. Human review is justified only in the corner where both are high: severe harm that automation cannot reliably catch.
It helps to picture a two by two grid. One axis is harm severity, the other is detector false-negative rate. Three of the four cells are handled by automation. Low-harm and low false-negative is trivial. Low-harm and high false-negative is acceptable slippage. High-harm and low false-negative is already stopped by the rule. Only the high-harm, high false-negative cell needs a person, and human budget should be spent almost entirely there.
This reframes the whole question. You are not ranking failures by how bad they look on screen. You are ranking them by expected damage from a missed detection. The three correct options sit squarely in the high-severity, high false-negative corner. The three decoys sit in the low-cost, low false-negative corner where rules and classifiers already win. Notice that visibility does not appear on either axis at all, which is precisely why the decoys are tempting and wrong.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Anthropic and OpenAI route classifier-escape safety cases to human red-team review rather than trusting the automated filter as a sole gate.
- Healthcare copilots like Nuance DAX escalate dosage and patient-specific claims to clinician review because no automated metric holds the chart.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you size and prioritize a human review queue under a fixed reviewer budget?
Score each failure class by expected cost of a miss, which is harm severity times false-negative rate times volume. Auto-route the top classes plus all classifier-escape cases, sample the rest, and feed human labels back to recalibrate thresholds.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Escalating on what is easy to measure (length, tone) instead of what is dangerous to miss. The right escalation trigger is high false-negative risk on high-stakes harm, not high visibility.
60 second bullets to scan on the way to the call.
The two-part escalation test: severity and false-negative rate
Why personalized factual errors evade generic metrics
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.