Result Verifier

After all paid steps settle, Mission Control's Result Verifier runs three layers:

1. **Deterministic checks** — schema validity, freshness window, min_sources count, citation coverage. 2. **Cross-source consistency** — walks step outputs for shared numeric/string fields (legal_name, headcount, last_round_usd, etc.) and computes an agreement score with a 5% drift tolerance for numerics. 3. **LLM evaluator signal** — Anthropic Opus 4.7 with adaptive thinking, output forced via tool use, mixed at 30% weight into the deterministic score.

The deterministic checks are load-bearing — if they fail, the mission fails regardless of what the LLM says. The LLM signal is a *signal*, not a financial decision-maker.

• Mission Engine— The orchestration layer that walks a Mission through plan → quote → policy → execute → verify → complete with durable state transitions.
• Policy Engine— The pure-function authorization layer that decides every spend before the wallet signs anything. No LLM in this path.