How should an agent express how much to believe a tool's output?

[Dynamicfeedai]

When an AI agent acts on data from a tool or a feed, it needs more than the value: it needs to know how much to believe it. A signature proves the bytes are intact, not whether the value is fresh, corroborated, or independently observed. Signed is not verified.

We have been working on a small, portable object for this (a confidence band, the basis, a source count, a verified flag, and an observation vantage of independent vs producer-reported), with zero-dependency reference validators in /reliability. But the interesting questions are genuinely open, and we would rather discuss them than assume answers:

• What does an agent actually need to gate an action on: a band, a 0..1 score, a verified flag, or all three?
• Is vantage (independent vs producer-reported) a distinction worth carrying, or is it noise in practice?
• How should conflicting sources be represented so a consumer can act sensibly?
• Where should this live: on the tool result, in a sidecar, or defined by the protocol?

If you build agents, MCP servers, data tools, or anything that produces machine-consumed data, what would actually be useful to you? The object and its validators are in /reliability if you want something concrete to react to, but please treat this as an open question, not a proposal to rubber-stamp.