OpenGradient can still fail at the message layer, before the model gets a token to predict.
I got stuck on the chat flow because the risk is not the final answer first. It is the stack of messages that shaped it.
A builder can call llm.chat(). The request can carry system, user, and assistant messages. It can also include tools and tool_choice. The result can come back with payment proof and TEE-backed prompt verification.
That sounds complete until an agent starts making decisions from it.
If the wrong system message sits above the user request, the model can follow the wrong authority. If a previous assistant message stays in the thread when it should have been cleared, the next answer can inherit stale context. If tool_choice nudges the wrong function path, the agent can act while the final output still looks normal.
The builder cannot only prove that OpenGradient ran the prompt.
They have to prove which conversation frame the model actually saw when the decision was made.
That is the consequence I care about. A wallet risk agent, audit assistant, or routing bot can produce a verified answer and still be wrong because the message roles fed into the call were wrong.
A signed answer is not enough if the conversation frame was polluted.
#OPG $OPG @OpenGradient $LINK $BLESS
I got stuck on the chat flow because the risk is not the final answer first. It is the stack of messages that shaped it.
A builder can call llm.chat(). The request can carry system, user, and assistant messages. It can also include tools and tool_choice. The result can come back with payment proof and TEE-backed prompt verification.
That sounds complete until an agent starts making decisions from it.
If the wrong system message sits above the user request, the model can follow the wrong authority. If a previous assistant message stays in the thread when it should have been cleared, the next answer can inherit stale context. If tool_choice nudges the wrong function path, the agent can act while the final output still looks normal.
The builder cannot only prove that OpenGradient ran the prompt.
They have to prove which conversation frame the model actually saw when the decision was made.
That is the consequence I care about. A wallet risk agent, audit assistant, or routing bot can produce a verified answer and still be wrong because the message roles fed into the call were wrong.
A signed answer is not enough if the conversation frame was polluted.
#OPG $OPG @OpenGradient $LINK $BLESS