Evidence

Evidence is the support-safe record Spinup keeps for each run and setup turn against a Spinup Agent. It explains what the agent used, what changed, and what was saved, without exposing secret values, raw prompts, or unbounded runtime output.

What a run records

For each run, Spinup records:

• the selected harness and model
• the runtime metadata Spinup chose for the run
• the projected state the harness received (capabilities, skill set, secret bindings by name)
• the run output Spinup is allowed to keep
• the evidence trail of what was applied, validated, projected, or rejected during the run

Setup turns leave similar records. When a harness asks Spinup to save a memory or a capability as a proposed change, the receipt the agent gets back is evidence that Spinup accepted, rejected, or held the change for review.

Why evidence is bounded

Evidence is designed to answer "what happened" without becoming a backdoor into the agent. Spinup keeps enough to support an operator review and to explain a failed run, but not the full prompt body, not the raw secret values, and not the unbounded byte stream from the runtime.

That boundary is the contract. If a piece of information is not safe to surface in a support context, it does not enter the evidence trail.

Where to go next

• See where runs sit in the product model in Runs.
• See how proposed memory updates become saved versions in Memory.