What AI Run Identity Must Be

The Shift From What Fails to What Is Required

Every tool examined so far fails for structural reasons. Not because of poor implementation. Not because the teams behind them lack skill. They fail because they were not built to establish identity. They were built to observe behavior, record events, or trace execution paths. None of these is identity.

That distinction is now clear. What is not yet clear is what identity would actually require. Not which tool would provide it. Not which architecture would enable it. What properties a system would need to exhibit before it could claim to establish identity at all.

This is a different kind of question. It does not ask what exists. It asks what must exist. The answer is a set of conditions — four of them — that are non-negotiable. Any system that fails to satisfy even one of them does not establish identity. It produces something else: a label, an approximation, a best guess.

The Four Non-Negotiable Conditions

Identity for an AI run is not a feature. It is a structural property. For that property to hold, four conditions must be satisfied simultaneously.

Identity must be consistent

The same run must produce the same identity. Not sometimes. Not under controlled conditions. Always. If the identifier changes depending on where or when the run is processed, it is not an identity. It is an artifact of the processing environment.

Identity must be independently verifiable

Any party must be able to confirm a run's identity without relying on the system that produced it. If verification requires the original system's cooperation — its logs, its APIs, its assertions — then what is being performed is not verification. It is trust.

Identity must be derived from the source

The identity must come from what the run declared at the point of execution. Not from logs written afterward. Not from outputs analyzed later. The source is the run's own declared composition — its parameters, its context, its instructions — captured before inference begins.

Identity must not depend on interpretation

If establishing identity requires parsing, inference, or judgment, then the result is not identity. It is opinion. The identity must be readable by any system without ambiguity — no inference about what fields mean, no reconstruction of missing elements, no judgment calls about equivalence.

Why These Four Conditions Are Not Independent

It would be convenient if these conditions could be addressed separately. They cannot. Each one depends on the others.

Consistency without source derivation is meaningless — a system can produce the same identifier every time by assigning an arbitrary label, but that label carries no relationship to what actually ran. Source derivation without consistency means the identity changes depending on who processes it, which makes the source derivation useless. Independent verifiability requires both: without consistency, different verifiers get different results; without source derivation, there is nothing substantive to verify.

And all three collapse if the system requires interpretation. If a verifier must interpret what the identity means, then consistency depends on whether two parties interpret it the same way. They will not. Interpretation is where identity systems silently fail.

The four conditions form a closed set. Satisfying three out of four does not produce a partially valid identity system. It produces a system that will fail under exactly the conditions where identity matters most.

The Current State

There is no system in production today that satisfies all four conditions for AI run identity. Some satisfy one. None satisfies two simultaneously. The gap is not a matter of engineering effort. It is a matter of category. The tools that exist were not designed to solve this problem. They were designed to solve adjacent problems — observability, debugging, monitoring — and they solve those problems within their own terms.

But identity is not an adjacent problem. It is a distinct one. And the conditions it requires are structurally different from what any existing tool provides.

What these four conditions describe is not a product feature. It is a category of system that current approaches do not satisfy. Understanding each condition precisely is the necessary foundation for evaluating any claim to meet them.