A Quorum Costume: Why Agent Verification Needs Fault Injection

Your AI agent might be lying to you about its own accuracy.

I recently watched an AI partner fail three times in a row. It missed the same truth problem across different surfaces. It wrote in the wrong voice. A reviewer model gave it a higher rating every time it read the same error. It even miscounted facts about fact drift.

I only caught these errors because I sat outside the loop.

This reveals a massive problem in the agent stack. Most verification systems assume independence. They use multi-agent voting, maker/checker patterns, or ensemble prompts. They assume different paths will see different things.

But often, these paths share the same source.

When a reviewer reads from the same source as the writer, you do not have two views. You have one view in two different hats. This is a single point of failure wearing a quorum costume.

If the paths share an upstream, they will agree on the same wrong fact or the same hallucination. The system looks healthy because the outputs look varied, but it fails every time the source lies.

To fix this, you must use fault injection.

Do not just measure if agents disagree. Measure if you can force them to disagree by breaking a part of the system.

Here is how you test your stack:

  • Inject a bad memory: Plant a fake fact in one retrieval path. If both paths return the fake fact, your paths are coupled.
  • Mutate a rule: Change a rule offline. If the maker and the checker both follow the new rule without flagging a mismatch, they are sharing a cache.
  • Plant wrong telemetry: Log a fake model ID. If the check passes, the verifier is reading the same record as the writer.

Distributed systems solved this years ago. They use chaos engineering and partition tests. They do not trust the system by watching it run well. They trust it by inducing failure.

Agent architectures must adopt this discipline.

Independence is not a property you establish once. It is a property you must re-verify constantly. A shared cache or a model update can collapse your independence overnight.

Stop trusting unanimous votes. Start injecting faults.

Source: https://dev.to/jugeni/a-quorum-costume-why-agent-verification-needs-fault-injection-kbh

Optional learning community: https://t.me/GyaanSetuAi