The question#
Which failures escape a coarse success check but become visible when we check the intended outcome?
This series develops selected records from my engineering work into inspectable case studies. Each article explains the trigger, the mechanism, the repair and the limits of the available evidence. “AI-assisted” describes the working context; it does not establish that AI caused a particular defect.
First publication#
A successful response can describe the wrong operation examines two repairs in Agent Shop. An unsupported method reached an HTML success response. A completion URL served the parent listing while status and metadata checks could pass. The article connects the code changes to historical browser checks and preserves the initial failed assertion.
The accompanying evidence package distinguishes GitHub's reported PR head from the merge revision actually checked out by CI. It includes selected log lines, a claim map, pinned source references and SHA-256 checksums.
My contribution and method#
Agent Shop is my project. I own the requirements and publication, and developed the work with AI coding assistance. The retrospective uses AI to help retrieve records, compare source, inspect test assertions and draft the explanation. That is disclosed assistance, not independent review or a measured comparison of human and AI performance.
The public record contains the source material needed for the article. Private working records are used for provenance and selection; they are not published as a dataset of verified incidents.
What follows#
Further candidates concern writes that resume after account deletion, payment fulfilment whose evidence remains incomplete, and generated handoffs that are structurally valid but misleading. They will join this series when their individual claims have adequate support. They are not results of a completed comparative study.
Moving from these articles to a research paper requires explicit episode selection, unsuccessful and negative cases, comparable before-and-after fixtures, a primary-source literature review and independent assessment where feasible. These publications begin that record without claiming novelty or generalisation from selected examples.
The complementary merged open-source contribution series covers work accepted in other maintainers' projects. This series keeps ownership and evidence scope explicit for my own systems.