Live report from our bundled sample logs. No upload, no API key — this is what you receive after sending your own JSONL export.
# Retrieval Autopsy — Sample Agent Logs **Period:** bundled design-partner example **Events:** 8 **Backend:** tavily ## Summary | Metric | Actual | Adaptive estimate | |--------|--------|-------------------| | Searches | 40 | 30 | | Cost (USD) | $0.3200 | $0.2400 | | Savings | — | $0.0800 (25.0%) | **Guardrail score:** 75/100 (moderate) **Over-fanout events:** 4/8 **Probe-only stops:** 1/8 ## Top fixes 1. **what is photosynthesis** — 5 → 1 searches (save $0.0320, probe_only) 2. **how does refund policy work for annual plans** — 5 → 3 searches (save $0.0160, partial_fanout) 3. **who is the ceo of anthropic** — 5 → 3 searches (save $0.0160, partial_fanout) 4. **how does kubernetes liveness probe work** — 5 → 3 searches (save $0.0160, partial_fanout) 5. **latest vector database benchmarks 2026** — 5 → 5 searches (save $0.0000, full_fanout) ## Governance note Hardest query by probe confidence: **compare stripe vs paddle pricing and api limits for saas billing** (probe=5%). A `legal` policy profile would evaluate evidence authority before returning `allow_answer`. ## Recommended next step 15-minute call to wire governed retrieval in staging: probe → policy → `allow_answer` → audit. --- *Generated by Query Fanout retrieval autopsy.*