AI Run Identity

Defining an AI Run

An AI run is a discrete execution event. It begins when a system assembles a configuration, constructs a context, selects a model, and initiates inference. It ends when the model produces an output and the execution context is released.

Between those two boundaries, several things happen at once. A system prompt is loaded — possibly from a file, a database, or a chain of templates. Retrieved documents may be injected into the context. Tool definitions may be attached. A conversation history may be truncated to fit within a token limit. Parameters — temperature, top-p, stop sequences — are set from configuration that may itself be dynamic.

The run is the totality of this assembly plus the inference that follows. It is not a single API call, though it may contain one. It is not a prompt, though a prompt is one of its components. It is the full composite: every instruction, every parameter, every piece of context that influenced the model's behavior during that execution.

When the run completes, this composite is discarded. The output persists. The assembly does not.

What Makes an AI Run Different From a Standard Computation

Standard computations have stable identities by construction. A function has a name, a signature, and a known implementation. You can read the code. You can trace the execution. Given the same inputs, the same function produces the same result. Identity is implicit in the determinism.

An AI run has none of this. The model's implementation is opaque — billions of parameters that no human can read. The behavior is non-deterministic — the same inputs can produce different outputs. The configuration is assembled at runtime from sources that may change between runs. The context window is a constructed artifact that exists only for the duration of execution.

In traditional computing, identity is a property of the code. In AI execution, there is no stable code to anchor identity to. The model is fixed for a given version, but the model is one component among many. The run's identity — if it had one — would need to encompass the full composition. No existing system captures this.

This distinction matters because every tool designed for traditional computation assumes identity is inherent. Process IDs identify processes. Transaction IDs identify transactions. But an AI run is not a process or a transaction. It is an assembly that exists once and is never recorded.

The Four Properties That Matter for Identity

For an AI run to have an identity, four properties would need to be established. Each addresses a different failure mode. Each is absent from current systems.

Run Identity

The foundational concept. What does it mean for a run to have an identity at all? Not a log entry. Not a trace ID. A record of what the run declared itself to be — its full composition — captured at the moment of execution.

Verifiability

Can a third party confirm that a run executed under the conditions it claimed? Current systems allow the operator to assert what happened. No system allows an external party to verify it independently.

Attestation

When is the identity captured? Logging records events after execution. Attestation would require the run to declare its own composition before inference begins. This distinction — between recording and declaring — does not exist in current practice.

Declarative vs. Inferred Identity

There are two ways to establish what ran. Infer it from logs and outputs after the fact. Or declare it at the point of execution. Current systems do the former, poorly. No system does the latter at all.