deterministic reconstruction

Traditional CRDTs struggle to ensure state consistency in Byzantine environments due to their reliance on update filtering mechanisms. This work proposes a deterministic state reconstruction approach that decouples update propagation from state derivation: all updates are accepted, but only valid ones contribute to the final state. By structurally rejecting or transforming malicious updates, the system guarantees convergence of replicas under arbitrary update injection and supports a layered security model encompassing authentication, authorization, and confidentiality. Built upon delta-state CRDTs, the resulting Melda system is tailored for JSON documents and formally verified to preserve state consistency despite message reordering, loss, or forgery. Theoretical analysis confirms that identical update sets yield indistinguishable states, achieving coordination-free strong eventual consistency with Byzantine fault tolerance.

Traditional distributed systems struggle to support modern autonomous infrastructures that integrate stochastic models and autonomous agents. This work proposes the Post-Deterministic Distributed System (PDDS) model, introducing for the first time its five architectural pillars. Its core innovation is a "cognitive state replication" mechanism that extends consistency from data visibility to knowledge visibility, alongside a novel fault classification framework. By leveraging protocol-driven development, verifiable agent infrastructure, and semantic quorum guarantees, PDDS enables coordination among semantically equivalent yet executionally divergent agents. This approach achieves verifiable semantic rollback and cross-agent reasoning consistency, establishing a theoretical foundation for trustworthy autonomous systems.

Data regulations such as the GDPR mandate compliant data deletion; however, the semantics of “deletion” remain ambiguous—especially in relational databases with functional dependencies and foreign keys—where residual data may still enable inference of deleted records, resulting in factual leakage. Method: We formally define Dependence-Aware Secure Erasure (DASE), a novel notion ensuring that post-erasure data reveal no additional knowledge about erased tuples. Our approach integrates dependency graph modeling, propagation-aware erasure semantics, and a dynamic cost-optimization algorithm supporting batch deletion and configurable retention policies. Contribution/Results: DASE enables tunable trade-offs between security guarantees and system overhead. Experiments on real-world and synthetic datasets demonstrate that our method significantly reduces inference-based leakage risk while maintaining low latency and high scalability.

This work addresses the challenge that local update efficiency in graph-indexed deterministic state systems often depends on the global system scale. To resolve this, the authors propose the Bounded Local Generator with Constraints (BLGC) framework, which enforces a finite interaction range and a bounded state space. Under explicit locality and boundedness constraints, the framework establishes—for the first time—a rigorous proof that the computational complexity of single-step updates is constant, i.e., O(1), thereby achieving structural decoupling between local computation and the overall system dimensionality. The approach integrates graph-indexed modeling, Hilbert space embedding via ℓ²(V;ℝᵈ), and operator norm analysis. Crucially, as the number of nodes M → ∞, the per-step computational workload remains invariant, substantially reducing the evolution cost for large-scale dynamic systems.

This study addresses the problem of reliably observing causal order in shared-memory concurrent systems (COP), formalizing its observability limits and proving that strong consistency—defined as both completeness and reliability—is generally unattainable. The key insight is that the placement of monitoring instrumentation, rather than the choice of timestamp mechanism, fundamentally determines observability guarantees. To this end, the work proposes three non-blocking monitor implementations: FAInc (a centralized atomic counter), Striped (a decentralized counter), and Collect (an iterative register snapshot). Theoretically, all three provide equivalent COP guarantees. Experimental evaluation on a 64-core NUMA architecture demonstrates that Striped achieves throughput comparable to Collect while maintaining linearizability and substantially alleviating the cache contention bottleneck inherent in FAInc.

This work proposes the first randomized algorithm that, using only hardware-supported bounded primitives such as compare-and-swap (CAS) and registers, implements m linearizable load-link/store-conditional (LL/SC) objects with constant expected step complexity under a weakly adaptive adversary. The construction achieves a space complexity of O(nτ + m) and satisfies quiescent history independence (QHI), offering both space efficiency and practicality. When m = O(1), the scheme matches the known theoretical lower bound; when m = Ω(n), it enables the deployment of the QHI dynamic hashing algorithm from STOC 2025 on existing hardware without increasing asymptotic step or space complexity, thereby bridging a critical gap in efficient software implementations.

This work addresses the challenge of accurately sampling valid trajectories from language model agents under policy constraints in state-dependent environments. The authors propose Stateful CARS, a method that freezes sound state-continuation patterns, prunes trajectories containing invalid continuations, and employs an exact residual Doob transform to enable unbiased sampling from the target conditional distribution. The approach introduces a verifiable bisimulation condition for future validity, providing theoretical guarantees including pattern soundness, adaptive exactness, i.i.d. output, and almost-sure termination, while supporting cross-historical reuse of invalidity certificates. Experiments on enumerable workflows demonstrate that the method achieves conditional distribution matching accuracy up to 10⁻¹⁶ (with a valid probability of 6×10⁻⁸), substantially outperforming state-aware local decoding baselines.

This work addresses the persistent degradation of agent reasoning and tool use caused by erroneous memories—such as contamination, staleness, or misattribution—which existing approaches struggle to correct without discarding valid knowledge. The paper formalizes, for the first time, the post-failure memory recovery problem and introduces a dependency-guided rollback repair mechanism. By constructing a typed memory-action dependency graph, the method tracks downstream effects at runtime, selectively deactivates unreliable memories, and replays only those computations relevant to the final answer. Evaluated on a controlled benchmark of 150 cases, the approach achieves an 85.3% recovery rate—surpassing the best baseline (77.3%)—while fully eliminating error sources and preserving all benign memories. In 50 stress-test scenarios, it attains a 68.0% recovery rate and significantly outperforms baselines, achieving the highest statement invalidation F1 score of 0.669.