IncidentLens AMD

IncidentLens AMD is an evidence-bound incident-triage prototype for LLM gateways. It parses bounded HTTP and Server-Sent Events traces locally, identifies premature stream termination, rapid retry cascades, attempted cache writes on incomplete streams, and latency outliers, then assigns stable event IDs to the resulting evidence table. When AMD_API_KEY is configured, the compact evidence bundle is sent to the event-issued AMD TokenFactory chat-completions endpoint through the repository's adapter. Before any model report is surfaced, a deterministic gate checks each non-heading claim unit: every claim must cite a known event ID, and outputs with unknown identifiers or uncited claims are rejected. The public Vercel deployment intentionally runs deterministic evaluation mode so reviewers can test the complete parser, report, and gate without receiving an event credential. Both /api/health and analysis responses disclose the active mode. This gate verifies citation coverage and identifier validity; it does not claim that every cited inference is automatically correct. The public MIT repository includes 19 automated tests, a multi-version CI matrix, the serverless deployment path, and reproducible local instructions.