Replication package for "Directing the Machine: A Single-Case Study of Verification When an AI Agent Cannot Reach the Oracle"
收藏资源简介:
Privacy-scrubbed replication package for a single-case study of a non-domain-expert directing an agentic LLM to build a cross-engine game-asset conversion pipeline. Contains the derived interaction event log (4,662 deduplicated events across 10 sessions — 9 development sessions plus one later analysis/writing session excluded from all process measures; the manuscript analyses the 4,522 development-session events; text lengths only, never message text), summary statistics, the forked-session deduplication analysis, the full 46-episode inventory, the coded failure/reversal catalogue (13 episodes, 5-dimension codebook, primary-detector and Auto? assignments) with per-episode verbatim transcript excerpts, the verification-burden measurement, the escalation typology, author-adjudicated labels for all 150 director messages, the coding scheme, a reproducibility appendix (tool and version inventory), a per-component walkthrough of the net-new code, scrubbed copies of the four project working documents and the commit log that the episode evidence cites, and the stdlib-only analysis scripts. No raw transcripts and no game assets are included; scrubbing policy in PRIVACY_NOTES.md. Supplements a manuscript under submission at Information and Software Technology.



