Sachin Gupta argues that conventional productivity metrics capture code-production speed but not the human attention required to trust the result. Review debt compounds when lightly reviewed code informs future agent changes, architectural decisions receive less scrutiny and higher throughput becomes the new baseline.
His framework combines ten deterministic checks across five signal families: diff size and coupling, test-evidence gaps, ownership spread, AI-authorship indicators, and missing rationale. It avoids an LLM judge so the score remains repeatable and defensible. AI markers only amplify other weaknesses rather than creating a penalty by themselves.
A scan of 524 public pull requests found that volume and structural complexity drove the burden more than authorship. Gupta recommends scoring recent work, setting a justification threshold, posting the result without automatically blocking merges, aggregating the trend by team and using it to guide review capacity and governance.
Watch on YouTube
