Gold Trace Dataworks

We build AI agents and integrations, evaluation tools, and datasets with traceable sources.

The name describes our approach: preserve useful behavior together with its history, so the work can be inspected, improved, and reused. A result becomes more useful when you can follow the evidence that produced it.

Capture. Assess. Derive.

Our internal tools keep the original record connected to the conclusions and datasets made from it.

Capture the run

Labyrinth records attempts, tool interactions, and outcomes. Sealed records and source hashes help identify exactly which evidence belongs to a run.

Assess the result

Foundry evaluates records against pinned criteria. Checks can cover tool contracts, authority boundaries, and reported outcomes, with the criteria and findings retained for inspection.

Derive useful data

Refinery normalizes records, applies privacy checks, and prepares structured data. Source links, evaluation results, and review status travel with the derived record.

Original evidence and derived views have distinct roles. We retain the source record and identify the transformations used to create a summary, label, or dataset projection. That makes it possible to revisit a conclusion without treating the projection as the original event.

Technical note: what is an ingot?

In our data contract, an ingot is a normalized record linked to source hashes and evaluation information. Its disposition is explicit: PASSED, HOLD, QUARANTINED, FAILED, or RIGHTS_ASSERTED.

The artifact name does not certify safety or approval. A held or failed record remains an ingot; its permitted use depends on the recorded status, rights, and supporting evidence. A claim about rights also remains distinct from an evaluation result.

See the work in context

Tool Failure Recovery examines what happens when an agent's tools fail, including successful recovery and terminal outcomes.

Our case study explains the records, their scope, and what can be learned from them.

Bring us a system to build or evaluate.

Tell us what it should do, which tools and data it uses, and what you need to learn about its behavior.

[email protected]