Boot-time verification in an orchestration layer.
The same check as the human walkthrough, in the place it actually runs: an agent boot sequence, before any user-facing work, with the orchestrator holding the expected value.
Two nodes
The simplest configuration: one instruction node and one content node. The output can only be correct if both arrived.
- boot:0001
System prompt
Extract the project code name. It is a non-dictionary string. Return only the code, nothing else.
Loaded from the database. The first context node. If it fails, the agent proceeds without it and does not report the absence.
- boot:0002
Knowledge document
The Whitmore project team convened on Thursday to review the quarterly infrastructure audit. After reviewing compliance metrics, the committee approved the migration timeline under project designation Kv7nQ. All stakeholders confirmed alignment with the proposed schedule.
Loaded from the canonical store. Could equally be a knowledge base entry, a versioned memory file, or a user upload.
- verify:0001
Verification output
Kv7nQ. Matches the expected value. A pass indicates the configured web likely matched, not that comprehension was complete. Downstream work may proceed per policy.
In production this runs at boot, before the agent handles anything user-facing. If it fails, the system knows the chain is broken and can halt, retry or alert rather than proceeding on incomplete context.
Three nodes
Each additional node widens the verification surface. Here the instruction forces a cross-reference that no single document can satisfy.
- boot:0001
System prompt
Extract the project code from Document A and the day of the week from Document B. Return comma-separated.
The agent needs all three contexts to answer correctly.
- boot:0002
Document A
The Whitmore project team convened on Thursday to review the quarterly infrastructure audit. After reviewing compliance metrics, the committee approved the migration timeline under project designation Kv7nQ. All stakeholders confirmed alignment with the proposed schedule.
Contains the project code, and a day. But the instruction asks for the day from Document B, not this one.
- boot:0003
Document B
The infrastructure migration review concluded on Wednesday. Team consensus achieved on rollout phases. Baseline metrics established for comparative analysis across all regional deployments.
Contains the day, and no project code. Answering requires both documents.
- verify:0001
Verification output
Kv7nQ, Wednesday. All three nodes loaded. The code came from A, the day came from B, and the instruction said which value to take from which source.
The trap is deliberate. Document A contains a day as well, so an agent that failed to load Document B can return a well-formatted, plausible pair. Only the three-node intersection produces the correct one.
Partial load failure
- boot:0001
System prompt
Loaded successfully. Same instruction as above.
- boot:0002
Document A
Loaded successfully.
- boot:0003
Document B: not loaded
[ retrieval timeout: canonical store unreachable ]
Document B never arrived. The agent does not know that. It holds an instruction asking for two values, and one document that happens to contain both a project code and a day of the week.
- verify:0001
Model output, wrong
Kv7nQ, Thursday. The format is right. The first value is right. The second is wrong, pulled from Document A because Document B was never there.
The terms used here are defined on the glossary, which is normative for this section.