RAVEL: Controlling Harmful Propagation in Cross-Agent Semantic Caches

A repository artifact can be correct when produced, stale when retrieved by another agent, and either harmless or outcome-changing in the consumer's dependency context. Cross-agent semantic caches therefore need a decision at every match: reuse the artifact, spend shared capacity to verify it, or recompute it. RAVEL implements this pre-consumption decision from prospective provenance, producer–consumer context mismatch, semantic drift, and remaining verification capacity; hard quarantine intercepts concentrated contradiction risk, while evidence-preserving refresh removes a superseded value from service without deleting its trust history. Fixed-request replay exposes why this control object matters: RAVEL admits 1,127 stale hits versus 772 for Centralized, yet records 126 versus 150 harmful events and 38 versus 45 issue-outcome flips. In a 16-agent repository simulator, RAVEL reaches 81.3$\pm 0.7$% task success at 0.523$\pm 0.008$ relative execution cost, compared with 80.0$\pm 0.8$% and 0.548$\pm 0.009$ for a controller with the same candidate, inputs, actions, quarantine rule, and cap. Across 600 held-out policy-dependent issue-runs, the gains over this logistic control are +1.8 [0.6, 3.0] success points and -0.045 [-0.059, -0.031] cost, with lower verification utilization. Document analysis, adaptive contamination, component ablations, and overlap–budget–churn sweeps support one systems principle: shared memory should govern the consequence of consumption, not freshness alone.