VERIFICATION FRICTION DESTROYS APPROVED WORK

Situation

A high-velocity AI agent infrastructure team uses PR approvals and automation assistance to move quickly. A single bot reviewer (ellipsis-dev) accounts for 64.4% of all review activity. Human reviewers depend on bot approvals as a verification signal.

Finding

Two closed-unmerged PRs represent distinct verification failure modes. PR #541: 1,007 lines across 25 files, approved with explicit maintainer signal ('this is looking pretty good / I can resolve conflicts and get it in'), accrued merge conflicts, and was closed unmerged 34 days later when conflict resolution exceeded maintainer bandwidth. PR #513: +19,392 lines authored by Claude Code, submitted with a blank template description, zero human reviews, and no tests — the bot reviewer explicitly refused to process it. Closed unmerged after 10 days. Combined: 20,399 lines of code converted to unshipped inventory without a merge.

Infrastructure

GitHub PRs, ellipsis-dev bot reviewer (64.4% of all review activity), Claude Code (AI code generation), manual verification checklists

Impact

1,007 lines of reviewed and approved code became unshipped inventory after 34 days of contributor engagement. A second 19,392-line AI-generated PR was abandoned with zero human review. The 34-day approval-to-close gap represents captured contributor attention with no outcome — a pattern that discourages future contributions and concentrates knowledge in the core team. The bot-review dominance creates a compliance illusion: 32% of merged PRs carry zero human review despite the team's product being built on the premise of human-in-the-loop AI governance.

Remediation

Implement an approval-to-merge SLO (72 hours from approval signal to merge). When conflict resolution exceeds SLO, assign a core team member to resolve rather than closing. Enforce a pre-merge gate for AI-generated PRs: require human sponsor, filled template, passing CI, and at least one human reviewer. Track unmerged-approved-code as an explicit waste metric.