Incorrect key. Try again.

What it does

• Looks up a service member's readiness (GO / NO-GO) by DoD ID through the readiness process API.
• If they are NO-GO, finds and ranks members who can take over their sortie. Ranking is computed by code, not by the model.
• After you pick a replacement, a confirmation card appears. Only your click on Confirm swap lets the swap be written to Excalibur.

Tools

• A LangGraph agent (GPT-4o mini) can call only get_member_readiness, find_replacements and swap_sortie_member.
• Answers come from the tool results, not the model's own guess. Diagnoses are never passed to the model.

Hardening

• The replacement is re-validated (GO, qualified, medical and currency valid, no conflicting sortie) before the confirmation card and again after you confirm.
• Off-topic requests are rejected before reaching the model. Message length, history and agent steps are bounded.
• A login key gates every API call. Per-IP rate limiting and a concurrency cap protect against abuse.
• Demo data only: the APIs behind this hold mock records.

Observability

• When enabled, each message and each swap confirmation is traced to Langfuse: the model calls, every tool call, latency and cost. A whole conversation is grouped as one session.
• Each trace is scored: in_scope, swap_approved (did a person click Confirm) and swap_executed (did Excalibur confirm the new assignment).
• Tracing is a separate, optional layer; it doesn't change what the agent is allowed to do.