Severity & confidence

How Ghost-hunter ranks hypotheses.

Ghost-hunter ranks hypotheses on two axes: severity (how much spend the hypothesis explains) and confidence (how strongly the evidence supports it).

Severity

LevelMeaning
CriticalExplains ≥ 50% of the unexpected spend.
HighExplains 20–50%.
MediumExplains 5–20%.
LowExplains < 5% but is anomalous.

Severity is computed from the billing CSV before any LLM call — the LLM does not pick severity.

Confidence

Confidence bars come from Opus and reflect how much corroborating evidence exists. A hypothesis starts at low confidence and rises as proposed commands return supporting output.

Hypothesis 1: BigQuery on-demand scan blew up    [████████░░] 0.82
Hypothesis 2: Egress from us-central1 to EU      [████░░░░░░] 0.41
Hypothesis 3: Idle Cloud SQL replica             [██░░░░░░░░] 0.19

Ghost-hunter stops the loop when one hypothesis crosses 0.85 confidence or when no hypothesis has improved in three consecutive rounds.

Why this matters

The dual axis stops Ghost-hunter from chasing a high-confidence-but-low-severity rabbit hole, and stops it from anchoring on a high-severity-but-unsupported guess. The final report leads with the highest-severity hypothesis above the confidence threshold.