Never Cry Wolf: Why Failures Default to Pass

The Error Budget Question

When the LLM times out, returns garbage, or the server crashes mid-inference — what should the evaluator do? There are only two options: assume something is wrong (violation), or assume everything is fine (pass).

We choose pass. Every time. Here's why.

The Cost Asymmetry

A false negative costs ~30 seconds of delay. A false positive costs your relationship with the user. In a monitoring tool, credibility is everything.

The Pattern Across the System

The Confidence Signal

Notice the fail-safe returns confidence: 0.3, not 0.9. This is deliberate:

try {
  const result = await this.llm.chatJSON<EvalResponse[]>(messages, {
    temperature: 0.6,
  });
  return this.parseEvalResults(result, rules);
} catch (error) {
  // Never produce a false positive from infrastructure failure
  return rules.map(rule => ({
    ruleId: rule.id,
    verdict: 'pass' as const,
    confidence: 0.3,  // Low confidence signals "I didn't really evaluate this"
    reason: 'Evaluation skipped due to inference error',
  }));
}

Downstream components can distinguish between a genuine "I evaluated this and it's fine" (confidence: 0.85) and a "the LLM was down so I defaulted" (confidence: 0.3). If you wanted to build a system health dashboard, low-confidence passes would flag inference reliability issues.

When You'd Flip This

This error budget only works for monitoring tools. In other domains, the calculus flips: