AI-gateway project and resource tagging

How tagging attributes spend to a build

A gateway sits between the application and the model providers, so every request already flows through one control point. Tagging uses that point to stamp each request with the identifier of the build it belongs to: a project key, a resource id, or a work-item reference. Because the stamp is applied at source rather than reconstructed later, the attribution is contemporaneous, which is exactly what makes it credible as evidence.

The identifiers to capture

• A stable project identifier that maps one-to-one to a recognised or in-progress asset.
• A phase marker distinguishing research, development and run, so only development-phase tokens flow into the asset.
• An environment or purpose flag separating fine-tuning and evaluation from production serving.

The discipline an auditor accepts

Tagging is only as good as its coverage and stability. An auditor will look for consistent application across the build, an identifier that does not change mid-project, and a governance rule that untagged spend is expensed by default rather than swept into the asset. A default-to-expense rule for untagged tokens protects the amount, because it means every capitalised token was affirmatively attributed rather than included by omission ASC 350-40.

Where tagged data lands

Tagged, phase-marked rows are the token ledger. From there, development-phase rows for a period are aggregated and posted as an addition to the intangible, while research and run rows fall to the profit and loss account. The tag is the join key across the whole bridge, so its quality upstream determines how defensible the carrying amount is downstream.