Three percent, and why it is not ninety-nine
The most common objection to the network's verification story is a good one: if only about three percent of queries are re-run on a second validator, then ninety-seven percent of answers are unchecked. Why should anyone trust an answer that was probably never audited?
Because the audit is not there to check your answer. It is there to make cheating unprofitable, and those two goals have wildly different prices.
The condition being bought
The validators chapter states the honest-validator condition plainly: detection probability times the cost of being caught must exceed the gain from cheating. Detection probability is bounded below by the audit fraction. The cost of being caught is the validator's stake, which is constrained to a minimum proportional to the largest job value that validator is eligible for. The gain is bounded above by what the protocol pays for a single job.
Put the numbers in. A divergence costs 50 basis points of stake per incident. A Standard-tier validator posts at least 60,000 OGI; a Heavy-tier one at least 200,000. Against that, the gain from returning a cheap wrong answer to one query is a fraction of one job's payment. At a three percent sampling rate the expected cost of cheating still exceeds the expected gain by orders of magnitude — not because three percent is a high chance of being caught, but because the thing at risk is enormously larger than the thing to be won.
Raising the fraction to ninety-nine percent would nearly double the network's compute bill to improve an inequality that is already wide. That compute comes out of the same treasury that pays for the answers themselves.
What the network guarantees
Correctness here is a property of the population over time. A validator that returns bad answers systematically is detected, challenged, and slashed, and it does not remain in the routing table long enough to matter.
That is P3: verification is statistical, and a surface that cannot be checked by sampling is not deployed. The guarantee is economic rather than cryptographic, which is the guarantee that scales to a network anyone can join. Raising the sampling rate would not strengthen it. It would only make it expensive.