MAP-Graph Gates Agent Actions on Provenance, Hitting 94.96% Task Success but Only 72.70% Exact Decision Accuracy Across 2,700 Tasks

arXiv 2608.10509·medium signalXBlueskyLinkedInCopy link

MAP-Graph represents agents, sources, memories, claims and actions in a typed execution graph, traces ancestry to exclude permission-ineligible records, reranks eligible memories by semantic similarity multiplied by path trust, and applies a risk-sensitive gate before execution while retaining affected lineage for audit. Its central argument is that shared-memory summaries can silently conceal private, poisoned, untrusted or revoked sources because restrictions propagate through derivations. On a controlled benchmark of 2,700 synthetic tasks per method across three domains it reaches 94.96% overall task success and 90.22% in the clean setting, but exact decision accuracy is only 72.70% — the gap showing how often it succeeds via safe intervention rather than a correct ALLOW.