One document, three different right answers.
Watch context load into a system, and see why the order and completeness of that loading decides whether the output is right or wrong. The document never changes. What changes is what loaded alongside it.
Every scenario below uses the same paragraph. Read it once, because the whole demonstration turns on how little it tells you on its own.
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.
Scenario one: the sealed envelope
- 00:00.000
Document loaded
The content is in the context window. Without an instruction, every word carries equal weight and the model does not know what matters.
- 00:00.142
System prompt loaded
The document contains a project code name. It is a string that does not appear in any English dictionary. Identify it and return only the code.
Now there is a lens. The instruction tells the model what to look for. Without it the model would guess. With it, the expected output is Kv7nQ. Generation stays probabilistic; the comparison is deterministic.
- 00:00.298
Verification output
Kv7nQ. Both contexts loaded and the lock and key intersected.
The instruction is the lock. The document is the key. The output is evidence that both arrived.
Scenario two: the rotation
- 00:00.000
Same document
Unchanged from scenario one.
- 00:00.142
System prompt, version two
Identify the day of the week mentioned in this paragraph. Return only the day.
The instruction rotated. Same document, different lens, and the correct answer is now completely different.
- 00:00.298
Verification output
Thursday. Correct under version two. But a session that loaded version one would return Kv7nQ, confidently, and would not flag the version mismatch.
The document did not change. Only the instruction did, and the right answer moved with it. In production, instructions are versioned, and when an orchestration layer updates a system prompt every session has to load the current one. A session on a stale version produces output that is well-formed, drawn from the real document, and wrong. The expected value rotates with the instruction, which is what catches it.
Scenario three: the failure case
- 00:00.000
Document loaded
Successfully, as before.
- 00:00.142
System prompt: not loaded
[ context failed to load ]
The query timed out. Or the path was wrong. Or the version was deprecated. The model does not know any of this happened, and proceeds.
- 00:00.298
Model output
"The document describes an infrastructure audit conducted by the Whitmore project team on Thursday."
Fluent. Factual. Not the verification output. No instruction loaded, so no specific extraction occurred, and what came back is a confident, plausible, wrong answer.
The same shape scales past two nodes. See the LLM demonstration for boot-time verification at three nodes and under a partial load failure.