AI change summaries

Read the evidence first

An activation request starts a plain-language summary in the requester's selected language. The generated text is stored with the request and included in the approval evidence snapshot. It is labelled AI-generated and advisory. An approver using another language sees the stored summary with its language indicated; opening the evidence does not translate or regenerate it.

The mechanical changes, rationales and fixture results are the authoritative evidence. The summary does not decide whether the change should be approved and its absence never blocks approval.

When generation is unavailable

If summary generation fails or does not settle, the evidence uses a deterministic sentence based on the actual comparison counts: “N steps changed, M fixtures changed result”. This fallback remains available without a successful model answer. The interface distinguishes a model summary from the deterministic fallback.

Untrusted authored text

Step names and rationale text can contain instructions aimed at the model. The summary request treats authored content as delimited, untrusted data, uses instruction/data separation, applies injection heuristics and validates the returned structure. These measures mitigate prompt injection; they do not prevent it. A well-formed summary can still be misleading or wrong.

Review the deterministic evidence before acting. The untrusted-documents reference explains the related limits for extraction; a change summary is advisory and cannot replace that evidence or the approver's judgment.