In spite of widespread acclaims of social audits as low-cost and powerful participatory tools that can bolster awareness and improve public service delivery, a key policy question is what such audits have achieved so far. Using a unique panel data set assembled from official social audit reports, we study the impact of these audits on MGNREGA delivery in Andhra Pradesh, India. Within a dynamic conceptual framework where beneficiaries, auditors and transgressors interact and learn and where beneficiary stakes vary across programme outcomes and irregularity types, we find a positive but insignificant impact of audits on employment generation and a modest decline in the leakage amount per labour related irregularity. These are outcomes with high beneficiary stakes. The latter occur alongside an increase in 'harder to detect' material-related irregularities with lower beneficiary stakes. Although we find evidence suggestive of beneficiary 'learning' from audit participation and of audit effectiveness in detecting irregularities, repeated audits did not deter irregularities. This highlights the need for a time bound process where transgressors are punished and responsibilities for follow-up of social audit findings are laid out and credibly enforced. Our findings suggest a changing anatomy of corruption, where transgressors keep one step ahead of auditors and respond to more intense scrutiny by locating new avenues for rent extraction.