problem
The AI Run Identity Gap
AI runs execute. They complete. They produce outputs. But no system records what they were. The gap is not in the tooling. The gap is in the category.
What runs is not what is recorded
Every AI run assembles a composition. A model version. A system prompt. Parameters. Retrieved context. Tool definitions. This composition determines the output.
When the run completes, the composition is discarded. What remains is the output, a timestamp, and whatever the logging system chose to capture.
The gap between the full composition and what is recorded is not a logging failure. It is a structural absence. No system is designed to capture run identity because run identity does not exist as a category.
Three capabilities that require identity
Reproduction requires knowing what ran. Without identity, you cannot re-execute a run under the same conditions. You cannot confirm that two runs were the same. You cannot determine why two runs differed.
Verification requires an independent record. Without identity, the only evidence of what ran comes from the system that ran it. No third party can confirm the claim.
Audit requires tracing execution to conditions. Without identity, audit reviews what was reported, not what executed. The audit is of the log, not of the run.
AI systems do not have identity. Every system that depends on reproduction, verification, or audit is operating without a foundation.
What currently fails
Logs record observations. They do not establish what a run was. Two runs with identical logs can be different runs. A log is a description, not an identity.
Outputs are results, not records. An output tells you what a model produced. It does not tell you what composition produced it.
Tracing records execution flow. It does not record execution composition. A trace shows what happened in sequence. It does not define what the run was.
Observability measures system behavior. It does not identify individual runs. Metrics describe populations. They do not define instances.
None of these establish identity. They all describe aspects of execution. The run itself remains unidentified.
This is not a tooling problem
Better logs do not close this gap. More traces do not close this gap. Additional observability does not close this gap.
The gap exists because AI run identity is a missing category. No tool addresses it because no tool was designed for it. The problem is not that existing tools are insufficient. The problem is that the category they would need to address does not exist in current systems.