RDEP · Documented worked example
Checking the source before acting on AI advice.
When the wrong source became the basis for advice.
AI projects often keep saved instructions and notes to help an agent continue its work. During RDEP’s development, an agent advised changing those project notes—but the text it relied on had come from a different memory location. The project owner checked the actual notes, exposing the mistaken attribution. The sequence below shows what was checked and corrected.
Reconstructed explanation of the mistaken reasoning, not a verbatim AI transcript. The annotations were prepared by us; they are not automatic detection.
The export contains statements that need correction. Those statements must therefore be in the project’s own memory. We should correct the project memory.
- The unsupported step
- The highlighted sentence attributes one source to another. Reading an export does not establish that it represents the project’s own memory.
- The premature recommendation
- The proposed correction depends on that attribution. The source-to-target link needed checking before advice about the project memory was justified.
- What the human supplied
- The owner supplied the text observed in project memory. It was compared separately with the export and a local reference.
- What the comparison supported
- The statements at issue were absent from the owner-supplied text. The attribution was withdrawn; the agent’s access to that project-memory surface remained unverified.
The error
The agent read an exported memory file from a different storage context and treated it as the project’s own saved memory. On that basis, it recommended correcting the project memory without first establishing that the file represented it.
The human check
The project owner inspected the project's memory and supplied the text. This was human intervention; the framework did not automatically detect the mistake.
The comparison
The supplied project-memory text, the export and a local reference document were compared as separate sources. The statements used to justify the proposed corrections were not present in the text supplied by the owner.
The correction
The unsupported attribution to project memory was withdrawn. The correction and a revised follow-up request were recorded, while the earlier record was preserved.
What remained unknown
The owner’s supplied text showed what the owner could read in the project memory. It did not establish that the agent’s memory tool could read the same project-specific area. That access question remained open.
This incident illustrates source checking within the framework. It is not a record of passing all nine Gates. See a separate teaching example alongside the nine Gates ↗
Incident recorded 5 September 2026; public account prepared 7 September 2026; explanation clarified 8 September 2026.
Download the case account ↗Download the worked example (PDF) ↗