Methodology

Every number in this report is computed deterministically from your uploaded logs by a rules engine that contains no AI model calls (NFR-01). Each call is priced at the provider rate in effect at its timestamp, from a versioned pricing table, machine-verified against independent published price data (R-AUTO-PRICING); calls on models without a verified rate card are listed as unpriced and excluded from totals rather than guessed. Monthly figures normalize the observed window to 30 days (x 30/observed-days). Savings estimates are conservative by construction: cache savings subtract estimated cache-write costs (one write per cache-lifetime window, provider-specific) and apply a 0.7 haircut when write windows cannot be estimated; prefix identity without hash evidence takes a 20% suffix haircut; prompt-bloat savings assume only half the excess is removable; model-downgrade savings are computed at the suggested model's published rates and model suitability requires your own quality evaluation; savings on prompt tokens are priced as the tokens were actually billed (cache reads at cache-read rates, never the full input rate). Waste classes can overlap on the same calls, so the headline savings total is capped at your observed monthly spend; per-finding figures are independent estimates. Spend estimates are FLOORS: provider long-context surcharges and regional data-residency multipliers are not modeled in v1. Evidence rows carry token counts and metadata only — never prompt or completion text.