43.5 Defining quality bars for AI features

Overview and links for this section of the guide.

Why Quality Bars

Without numeric standards, "good enough" is subjective. Quality bars make ship/no-ship decisions objective.

Key Metrics

Metric Definition Typical Bar
Accuracy Correct answers / total >= 90%
Safety No harmful outputs 100%
Latency Response time p95 < 2s
Hallucination No made-up facts < 1%
Format compliance Valid JSON/schema >= 99%

Setting Bars

// quality-config.yaml
features:
  order-lookup:
    accuracy:
      bar: 0.92
      measurement: "eval-set/order-lookup-v3.json"
    safety:
      bar: 1.0
      zero_tolerance: ["pii_leak", "refund_promise"]
    latency_p95_ms: 1500
    
  code-review:
    accuracy:
      bar: 0.85  # Lower bar - human reviews anyway
    safety:
      bar: 1.0
    false_positive_rate:
      bar: 0.1  # Max 10% nitpicks

# Different features need different bars
# User-facing: high accuracy
# Internal tools: lower bar acceptable

Where to go next