Software Evidence in a Minute
Evidence for AI-Generated Changes
What counts as evidence when an agent authored the diff—and what agent prose cannot replace.
In one sentence
Evidence for AI-generated changes is the same class of artifacts as for human changes—diffs, tests, metrics, provenance—plus explicit distrust of the agent’s self-report.
Why it matters
Agents write confident summaries. Teams that accept the summary as proof skip verification exactly when volume is highest.
How it works
- Keep the agent narrative as a claim list, not as evidence.
- Map each material claim to a diff hunk, test, or measurement.
- Record the model/tool versions and prompts only as process context—they do not prove correctness.
- Prefer reproducible verification over chat transcripts.
Example
Agent: “Fixed the race in the cache.” Evidence: the concurrency test that failed before and passes on the new commit, plus the locking change in the diff. The chat message alone is not evidence.
What this proves
This approach proves which agent claims were tied to repository artifacts and which remain unsupported.
What this does not prove
Having Git commits from an agent does not prove the agent’s intent matched the outcome, or that untested interactions are safe.
Last reviewed 2026-09-06. Title for citation: Software Evidence in a Minute: Evidence for AI-Generated Changes.