9  Faithfulness, Confidence, and What Verification Misses

9.1 Chapter Map

• Explain what important properties remain off-screen even when a verifier is strong.
• Show why verified success does not imply faithful reasoning or calibrated uncertainty.

9.2 Main Argument

RLVR can produce correct answers without guaranteeing faithful reasoning, transparent internal computation, or calibrated confidence. That is not a reason to abandon RLVR. It is a reason to specify the limits of verification clearly and resist overclaiming.

This chapter should give the reader a principled way to talk about what verifiers cannot see, even when they are strong at the task they were built for.

9.3 Canonical Examples

• Correct final answers paired with misleading chain-of-thought traces.
• Self-verification loops that sound confident while remaining wrong.
• Agents that satisfy tool-based checks while remaining brittle under minor environment changes.

9.4 Failure Modes

• Treating visible reasoning as if it were complete access to model cognition.
• Conflating confidence expression with calibrated belief.
• Using process traces as evidence of truth without checking whether they are causally relevant.

9.5 What the Verifier Sees

The verifier sees externalized answers, explanations, traces, and confidence expressions that have been explicitly surfaced.

9.6 What the Verifier Misses

It misses hidden reasoning, latent uncertainty, and any mismatch between narrated reasoning and actual causal computation.

9.7 Research Notes

• Which faithfulness claims are empirically defensible in verifier-heavy systems?
• How should confidence be trained or evaluated when correctness is verifiable but uncertainty is not?
• Can richer verifier stacks reduce blind spots without pretending to eliminate them?