Why Relationships Are the Unit of Trust, Not Tokens
A token is a surface. A relationship is a structural commitment. Trust cannot be assigned to surfaces because surfaces do not tell you how they got there. It can be assigned to relationships because a relationship, by its nature, exposes what it connects.
Consider two answers to the same question. Both are fluent. Both are internally consistent. One is correct and one is not. Nothing on the surface distinguishes them.
This is the everyday failure mode of statistical language systems, and it is the reason "trust the output" has never been a coherent instruction. There is nothing on the surface of an output to trust.
What can be trusted
A relationship can be trusted. Not because relationships are always right — they are not — but because a relationship, unlike a token, is an auditable objectA relationship object exposes: two endpoints (entities), a type, a provenance edge to its origin (evidence or derivation), and optionally a signature. All four are inspectable.. It has two endpoints. It has a type. It has, if the system is built correctly, a provenance edge pointing to whatever brought it into existence.
If a claim is a node and its supporting relationships are the edges reaching it, then the trust question stops being "do I believe this string" and becomes "does the edge set that supports this claim exist, and does every edge trace to evidence I can check". This is a different kind of question. It is answerable.
The triangle inequality of trust
In a metric space, distance obeys the triangle inequality: the direct path between two points is never longer than a detour through a third. In an evidence graph a related inequality holds: the trust you can place in a claim reached by two independent paths is bounded below by the weaker of the two paths, and above by neither.
This is the corroboration boundFor any claim C reached by independent evidence paths P1..Pn, the trust score is bounded below by max(trust(P_i)) and increases sublinearly with the addition of independent paths. Not summation.. It says that adding a corroborating path never lowers trust and often raises it — but never by simply summing scores. Trust is not additive. It is structural.
Why token-level trust cannot recover this
You can attach a confidence number to a token. Models do this. But a confidence number over a token is a scalar on a surface. It cannot express "this word is trusted because these three edges exist and one of them derives from a signed evidence artifact." It can only express "this word is likely."
Likelihood is a summary of the training distribution. Trust is a claim about the world. These are different quantities that a scalar cannot carry at the same time.
The practical consequence
A system whose atomic unit is the token can be made more fluent. It cannot be made more trustworthy in this stronger sense, because there is nowhere to put the trustworthiness. There is no edge to point to.
A system whose atomic unit is the relationship inherits, from its very architecture, a place to attach evidence, a place to attach provenance, a place to attach a signature. It does not automatically become trustworthy. But it becomes possible to be.
Possibility is the whole game.