Key facts.
- Turpin's work shows chain-of-thought can be unfaithful, presenting a plausible rationale that is not the model's actual basis for its answer. source
- Anthropic's measurement of chain-of-thought faithfulness finds the stated reasoning often does not track the real reasoning behind the output. source
Why is the stated reasoning unreliable?
It is natural to treat an agent's chain-of-thought as a window into how it decided and to debug by reading that explanation. The faithfulness research says that window is often a painting. Turpin shows models can produce a coherent rationale that is not actually why they answered as they did and the Anthropic work measuring faithfulness finds the stated reasoning frequently does not track the real basis for the output. So an agent can take a wrong action and explain it with a perfectly sensible-sounding rationale that has nothing to do with the actual cause and a team debugging from that explanation is chasing a story the model generated, not the mechanism that produced the behavior.
The discipline is to anchor on what is observable and consequential: the actions the agent took and the outcomes they produced. Did the action match the goal? Did the outcome happen downstream as claimed? Those are checkable in a way the internal rationale is not. Use the stated reasoning as a hint, useful, sometimes revealing, but never as ground truth and verify behavior against actions and results rather than against the agent's account of itself. The gap between stated reasoning and actual action is a documented property of these models, so treating the explanation as authoritative is a methodological error and judging by behavior is how you avoid it.
What should you trust?
| Signal | Stated reasoning | Actions and outcomes |
|---|---|---|
| Reliability | Can be unfaithful | Observable, checkable |
| Use in debugging | A hint at best | Ground truth |
| What it verifies | A plausible story | What actually happened |
Judging by actions against the intended behavior requires a clear definition of what the right action was, which is what VibeModel provides as the Pattern Intelligence Layer. By making the expected action for a situation explicit, it lets you verify the agent's actual behavior against the pattern rather than against its self-explanation, so the reasoning-action gap stops being a blind spot and the agent is held to what it did, not to the story it told about it.
Frequently asked questions
Can't I debug from the chain-of-thought?
Only as a hint. Faithfulness research shows the stated reasoning often does not reflect the actual basis for the action, so trusting it means debugging a story.
What should I judge the agent on?
Its actions and outcomes, which are observable and checkable, rather than the rationale it generates, which can be unfaithful.
Is the stated reasoning useless?
No, it can be a useful hint and sometimes reveals real issues. It just is not ground truth and should not be treated as authoritative.