How these numbers are computed

AIEP measures savings at the evidence-artefact level. Each time a prior reasoning artefact is recalled from the committed KV store instead of being re-derived or re-fetched, that artefact's estimated token cost (≈ 500 tokens) is counted as avoided. Tokens saved = (total artefacts − unique source URLs) × 500.

The reuse ratio is the share of evidence artefacts served from committed prior reasoning. A ratio of 0.72 means 72 of every 100 evidence steps were recalled rather than derived fresh.

All stats are aggregate and approximate. Exact token counts depend on model, prompt structure, and evidence depth — 500 tokens per artefact is a conservative mid-range estimate.