gg-tank-watch — safety method & red-team

Published evidence for an AI-in-the-loop, real-emergency situational-awareness dashboard (archive of the May 2026 Garden Grove chemical-tank incident). Cited as a provenance receipt from gardwyn.com.

• Failure analysis (F1–F12 red-team) — how the eval harness catches each failure mode, what stays partially guarded, and what is unguarded. Honest about limits.
• Eval summary (eval-summary.json) — deterministic export of the behavioral + schema test suite (198/198 green), bound to the source commit. The sha256 of this file backs the gardwyn receipt; fetch it and sha256sum to verify.
• Decision authority (the human keeps the final verdict) — the design boundary that scopes the AI to situational awareness and keeps the irreversible safety call with a human, plus the three failure classes only a human catches.

The dashboard application source is private. This repository holds only the published method artifacts.