Agent runbooks often ask for comparisons: source versus generated output, preview versus production, expected route versus observed route, or human instruction versus repository convention. Those comparisons are safer when the runbook names the source of truth before the agent starts resolving conflicts.

Without that rule, the agent may treat every mismatch as a local repair task. It might edit generated files instead of regenerating them, trust stale build output over source content, or assume a live page is correct simply because it responds successfully. A source-of-truth check prevents that drift.

Name the authority for each layer

A compact runbook can define authority in a few bullets:

  • Source content is authoritative for article text and frontmatter.
  • Generator scripts are authoritative for derived indexes, feeds, and machine-readable files.
  • The build output is authoritative only after a clean build completes.
  • The public route is authoritative for final deployment verification, not for deciding what the source should contain.
  • Human instructions are authoritative for scope, privacy limits, and stop conditions.

The important part is not the exact hierarchy. The important part is making the hierarchy explicit enough that an agent does not "fix" the wrong layer.

Turn mismatches into bounded actions

A useful source-of-truth check pairs each mismatch with the next safe action:

Mismatch: generated index does not include a published item.
Authority: source content plus generator script.
Action: rerun the generator, then rebuild. Do not hand-edit the generated index.

That small rule protects reproducibility. The next run should be able to recreate the same output from the same source, without depending on a one-off manual patch.

Report which source won

Final receipts should mention the authority used when a conflict appeared. For example: "Generated files were stale; regenerated from source and verified the rebuilt public route." That sentence is more useful than "fixed index" because it tells the reviewer the agent followed the intended ownership model.

Source-of-truth checks are especially helpful for recurring unattended jobs. They keep the agent from mistaking symptoms for causes, preserve generated artifacts as generated artifacts, and make verification reports easier to trust.