01
The deployed system nobody names
Search and answer systems affect which sources a reader encounters. That makes the handling of evidence an engineering question as well as an epistemic one. I don’t have access to their full internal trust models.
What interests me for my own systems is how to keep a retrieved statement connected to its source and its uncertainty, instead of treating fluent presentation as evidence.
02
Trust is a gradient, not a gate
Subjective logic offers a formal way to represent belief, disbelief, and uncertainty. That is a representation of an assessment, not a mechanism that turns a claim into truth. Its usefulness depends on how the values are assigned and combined.
Wright's tier language makes the gradient legible: cryptographically anchored entities at the bedrock, platform-authenticated ones above, passively cross-referenced ones after that, and single-source fog at the top. The engineering question is never "is this true?" — it is "how much evidence, of what independence, from what sources, how recently?"
03
Why agents need it
If retrieval loses source quality and independence, an agent may give a copied assertion more weight than it deserves. I want the record to distinguish direct evidence from repetition and to preserve uncertainty when the same claim moves between systems.
The gradient is the adjudication layer for beliefs: it tells a system when it may commit, when it should hedge, and when a human should look before anything downstream inherits the claim.
04
Still open
The question I am working on is what a system should do with its uncertainty: continue, seek another source, or ask for review. A graded belief representation is only useful if it leads to better decisions than a simpler approach.