GuardWrite: experiment results and full call logs

Results for "Invalidation, Not Information: A Theory and Protocol for Coherent Memory in Multi-Agent LLM Systems" (Suranjan Goswami, Kuldeep Singh, Karan Bharadwaj, Manu Ghulyani; Nasiko Labs). The study asks when a stale entry in shared agent memory still drives an agent's action, and tests GuardWrite, a protocol that invalidates such entries. This record holds every output of its experiments (E1 to E8) across nine language models and more than 120,000 agent runs. It contains agent trajectories, single-call outputs, scores, human validation labels, and one log record per model-call attempt with the raw request and response. Nothing was filtered. Failed calls, refusals, truncated outputs, pilots, smoke tests and excluded runs are all kept. Each excluded run sits in a folder with a REASON.txt. Public inputs (tau-bench, CRMArena-Pro, model weights) are not redistributed; the code supplement fetches them from their sources. Extract all archives into one folder to rebuild the data tree. The 16 blobs archives together form the raw payload store that every call log points to. README.md describes each archive, the folder layout and the log fields. SHA256SUMS lists a checksum for every archive. Files are restricted until the article is accepted.