current HITL implementations solve the pause/resume problem but not the trust problem.
but there’s no way for the agent to verify the approval actually came from an authorized human vs an injected signal.
HDP (human delegation provenance) fixes this at the protocol layer. native CrewAI support is implemented and awaiting review: 
feat(examples): HDP delegation provenance integration by asiridalugoda · Pull Request #5135 · crewAIInc/crewAI · GitHub
happy to answer questions.
This is a useful distinction. How are you thinking about the boundary between proving who approved an action and proving what they approved? For a payment, the amount, merchant, time window, and agent identity all seem like they need to be bound into the authorization, otherwise a valid approval could be reused in a different context.
We are exploring this problem at Credian, so I am especially curious whether HDP records enough context for an auditor to reconstruct the decision later. More context on what we are researching is at https://credian.io