Pith. sign in

REVIEW 3 major objections 5 minor 25 references

Agentic replicas should agree on operational belief, not bitwise token state, via Epistemic State Replication that separates evidence from belief lineage.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-14 16:31 UTC pith:ZRTJSJWV

load-bearing objection Useful systems abstraction for agent control planes; the coherence theorem hinges on an unmeasured contractive graft the authors themselves flag. the 3 major comments →

arxiv 2607.09748 v1 pith:ZRTJSJWV submitted 2026-07-03 cs.AI cs.DCcs.LOcs.MA

Replicating Belief, Not Bits: Epistemic State Replication for Agentic Systems

Jun He , Deying Yu This is my paper
classification cs.AI cs.DCcs.LOcs.MA
keywords epistemic state replicationagentic systemssemantic linearizabilitybelief lineageverifiable semantic rollbackbounded eventual coherencestate machine replicationLLM agents
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Classical state-machine replication demands that correct replicas end up with identical bit-for-bit states. When the replicas are generative agents, that demand is both impractical and harmful: different reasoning paths can still produce the same safe operational decision, and forcing token-level equality stalls the system or erases useful context. This paper introduces Epistemic State Replication (ESR), which treats each node’s knowledge as a pair consisting of an immutable shared evidence log and a stochastic belief lineage. Safety is governed by Semantic Linearizability (operations must reflect the latest committed operational meaning within a verifier-bounded distance) and Bounded Eventual Coherence (expected semantic divergence stays bounded under fair delivery and a contractive graft). Faulty premises are removed by Verifiable Semantic Rollbacks that prune the causal descendants of a bad belief while repairing only partially dependent ones, so the agent does not suffer context amnesia. A preliminary cloud-control-plane simulation illustrates that the approach is feasible and can sharply reduce secondary cognitive faults after recovery.

Core claim

In agentic distributed systems the correct replication target is operational belief, not physical bits. By modeling each replica’s state as K = (L, B)—an append-only evidence log L plus a DAG belief lineage B—and by enforcing Semantic Linearizability and Bounded Eventual Coherence on the projected meaning of B, replicas can diverge in tokens and reasoning traces yet remain inside a verifier-bounded compatibility class of admissible actions. Verifiable Semantic Rollbacks then surgically remove poisoned premises and their causal descendants without erasing independent knowledge, preventing the agent from re-committing the same error after physical recovery.

What carries the argument

Epistemic node state K = (L, B) together with the grafting operator ⊕ on epistemic deltas; Semantic Linearizability and the Bounded Eventual Coherence theorem that bounds expected semantic distance under a contractive graft.

Load-bearing premise

The coherence guarantee rests on the grafting step being a contraction: merging the same shared insight must pull two belief lineages closer together rather than farther apart—an assumption that holds only under a disciplined ontology and audited verifier, not for arbitrary language-model reasoning.

What would settle it

Run heterogeneous agent replicas under identical evidence streams with and without ESR’s graft and rollback; if expected semantic distance fails to stay inside the predicted neighborhood, or if secondary re-faults after rollback remain comparable to plain physical rollback, the central claim fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Infrastructure control planes can keep stochastic LLM agents without forcing identical prompts or token buffers.
  • Poisoned or hallucinated premises can be surgically excised from active memory while independent knowledge is preserved.
  • Replication traffic shrinks from full reasoning traces to lightweight epistemic deltas, raising cognitive-transaction throughput.
  • Safety-critical action sets can stay agreed upon even when non-critical stylistic or retrieval choices diverge.
  • Classical consensus remains only for the evidence log; belief consistency is lifted to a verifier-bounded semantic projection.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the contractive-graft condition can be enforced by a small, audited ontology rather than by the raw LLM, ESR becomes a practical middleware layer between existing consensus stores and multi-model agent fleets.
  • The same justification DAG could serve as an audit trail for regulators who need to know which evidence authorized an autonomous infrastructure change.
  • Partial-dependency repair suggests a general pattern for any memory system that must roll back poisoned context without total amnesia—applicable beyond cloud control planes.
  • Heterogeneous model clusters become first-class citizens once deltas are model-agnostic; the bottleneck shifts from model agreement to verifier quality.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Epistemic State Replication (ESR) for agentic distributed systems whose replicas use stochastic generative models. Classical SMR’s bitwise equality is replaced by agreement on operational belief: each node state is K=(L,B), with an immutable, consensus-ordered evidence log L and a stochastic belief lineage B represented as a DAG of justified beliefs. The authors define Semantic Linearizability (classical linearizability on a verifier-bounded semantic projection E(S)) and Bounded Eventual Coherence (a recurrence bound on expected semantic distance under fair delivery, monotonic evidence, bounded verifier noise, and a contractive graft operator). They specify epistemic deltas, a grafting/contradiction protocol, and Verifiable Semantic Rollbacks with partial-dependency repair, and report a single-trace cloud-control-plane simulation claiming higher CTPS and ~97.7% fewer secondary cognitive faults versus bitwise rollback.

Significance. If the framework holds under realistic verifier and grafting conditions, it would give the field a usable consistency vocabulary for multi-agent LLM controllers that classical SMR and CRDTs do not provide: separate physical evidence from cognitive state, keep safety-critical actions within a semantic neighborhood, and align physical recovery with belief repair so that context amnesia and re-execution loops are reduced. The K=(L,B) split, the CT lifecycle with distinct pre-execution and post-commit validation, the rollback safety invariants (Theorems 2–3), and the explicit acknowledgment that contraction is not free LLM behavior are genuine conceptual contributions. The manuscript does not ship machine-checked proofs or reproducible multi-seed experiments; its value is primarily definitional and architectural, with simulation used only as a feasibility sketch.

major comments (3)
  1. §3.3 Theorem 1 (Bounded Eventual Coherence) is load-bearing for the claim that heterogeneous stochastic replicas stay in a verifier-bounded semantic neighborhood. The recurrence E[dt] ≤ γ^t d0 + (1−γ^t)/(1−γ) δerr rests entirely on the contractive graft assumption (Eq. 12): dM(Eb(BA⊕Δe), Eb(BB⊕Δe)) ≤ γ · dM(Eb(BA), Eb(BB)) with γ∈(0,1). The paper correctly states this is strong and not a property of arbitrary LLM reasoning, yet §6.2 grafts across Gemini 1.5 Pro, Claude 3.5 Sonnet, and Llama 3 70B via natural-language deltas and LLM/cache contradiction checks without measuring γ, without showing that ⊕ reduces rather than amplifies dM, and without an ablation under the stated ontology/verifier constraints. Either provide empirical contraction measurements (or a restricted grafting regime where γ is enforced by construction) or demote Theorem 1 from a system guarantee to a conditional char
  2. §6 and Table 4: the evaluation is a single unreproducible simulation trace (no archived seeds, prompt templates, model/API settings, token/Δe size distributions, or repeated trials). The reported 4.8× CTPS gain (12.4 vs 2.6) and 97.7% secondary-fault reduction (6.4 → 0.15) are therefore illustrative only, as the authors themselves note. These numbers currently underwrite the abstract’s “show feasibility… and illustrate reductions” claim and the reader’s strongest claim. For a journal contribution they must either be removed from the abstract/results framing or replaced by a multi-seed evaluation with the metadata listed in §6.4, confidence intervals, and an explicit baseline that isolates semantic rollback from other protocol effects.
  3. §3.1–3.2: Semantic Linearizability is defined via dM(E(S),·)≤ε and three alternative action-compatibility conditions (policy-compatible selected actions, safety-critical set Acrit, or bounded symmetric difference β). The manuscript never fixes which condition is used in the prototype, how ε/β/Acrit are calibrated, or how the Semantic Quorum Validate procedure in Table 2 implements them for heterogeneous models. Without an operational instantiation, “verifier-bounded semantic compatibility” remains a free parameter rather than a checkable safety property, which weakens the claim that ESR governs execution safety rather than only describing it.
minor comments (5)
  1. §2.3 CT lifecycle: the distinction between pre-validation (execution-safety gate) and the Finalized linearization point is important; a short diagram of the state machine would make the linearization argument easier to follow.
  2. Tables 1–3: the pseudocode is useful but uses heavily spaced identifiers (e.g., E x t r a c t D e l t a); normalize to standard monospaced identifiers for readability.
  3. §4.2 contradiction resolution: Kmax and the priority ordering by (c,τ) are introduced without guidance on choosing Kmax or handling equal-confidence conflicts; a short paragraph on defaults would help implementers.
  4. Related work §7.1: CRDT and provenance citations are appropriate; a brief comparison to Byzantine SMR / BFT under non-deterministic application logic would situate ESR more clearly for the distributed-systems audience.
  5. Eq. (1)–(2) and Definition 1: notation for ≈ε and dM is clear; ensure E vs Eb is used consistently in later sections (occasional bare E for belief projection).

Circularity Check

0 steps flagged

No significant circularity: formal claims are conditional definitions/theorems; simulation is labeled illustrative, not a forced prediction.

full rationale

ESR’s load-bearing formal content is definitional and conditional, not a fit-or-rename of its own inputs. Semantic Linearizability is a specification (classical linearizability on a verifier-bounded projection E), not a derived empirical law. Theorem 1 (Bounded Eventual Coherence) is a standard contraction recurrence: under fair delivery, monotonic evidence, E[ηt|Ft]≤δerr, and a contractive graft with factor γ∈(0,1), one obtains E[dt]≤γ^t d0+(1−γ^t)/(1−γ)δerr and lim sup E[dt]≤δerr/(1−γ). The contractive-graft step is an explicit assumption the authors flag as strong and not automatic for arbitrary LLMs; the bound is not obtained by defining γ from the target neighborhood. Rollback Theorems 2–3 follow from the pruning/repair construction and stated invariants. The single same-author citation (the post-deterministic manifesto) is motivational framing, not a uniqueness theorem or load-bearing external fact that forces the design. Simulation CTPS and secondary-fault figures are presented as unreproducible single-trace illustrations under the modeled workload, not as predictions forced by fitted free parameters. No self-definitional identity, fitted-input-as-prediction, or ansatz-smuggled-via-citation reduction is present in the derivation chain.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 6 invented entities

ESR’s safety/liveness story rests less on fitted constants than on strong domain assumptions about verifiers, delivery, and especially a contractive graft operator over belief lineages. Free parameters are mostly thresholds and protocol knobs (ε, γ, δerr, Kmax, β, quorum). Invented entities are the paper’s core abstractions; none have independent external measurement outside this framework. The central empirical claims further depend on an unreproduced simulation regime rather than on parameter-free external benchmarks.

free parameters (5)
  • semantic compatibility radius ε
    Defines ϵ-semantic compatibility and the target neighborhood for coherence; deployment-specific and not derived from first principles.
  • graft contraction factor γ
    Assumed γ∈(0,1) in Theorem 1; controls the steady-state bound δerr/(1-γ) but is not measured for real LLM grafting.
  • verifier/noise bound δerr
    Upper bound on expected semantic disturbance from validators, sampling, retrieval, and repair; enters the limsup coherence bound directly.
  • contradiction recursion bound Kmax
    Hand-set depth limit before operator/deterministic arbitration; affects liveness and quarantine behavior.
  • action disagreement bound β / critical action set Acrit
    Deployment-specific relaxations of action-set equality used to make semantic compatibility operational.
axioms (6)
  • domain assumption Fair eventual delivery of epistemic deltas to all correct replicas
    Assumption 1 for Bounded Eventual Coherence (§3.3); standard distributed-systems liveness premise.
  • domain assumption Evidence log L is append-only, consensus-ordered, and eventually available to correct replicas
    Separates deterministic physical truth from stochastic belief; relies on classical consensus TCB (§2.4, §3.3).
  • ad hoc to paper Grafting operator ⊕ is a contraction on belief projections for shared deltas
    Load-bearing for Theorem 1; paper admits it is not generic LLM behavior and requires disciplined ontology/verifier regime (§3.3).
  • domain assumption Semantic verifier errors and sampling noise have bounded conditional expectation E[ηt|Ft] ≤ δerr
    Needed for expected neighborhood convergence rather than point convergence (§3.3).
  • domain assumption Trusted Computing Base includes consensus on L, semantic quorum, policy boundary P, and verifiable justification links J
    Trust boundary stated in §2.4.3; safety claims are relative to these components.
  • ad hoc to paper Repair transactions recommit replacements with independently checkable justification sets excluding the faulty premise
    Used for Theorems 2–3 on rollback safety (§5.5); assumes CT_repair behaves as specified.
invented entities (6)
  • Epistemic node state K=(L,B) no independent evidence
    purpose: Split deterministic evidence from stochastic belief lineage as the unit of agent replica state
    Core modeling move of the paper; not an independently measured physical object.
  • Belief lineage DAG with justification sets J no independent evidence
    purpose: Track causal ancestors of beliefs to enable grafting and semantic rollback without context amnesia
    Operational structure introduced for ESR; related to provenance but used online for replication/recovery.
  • Semantic Linearizability no independent evidence
    purpose: Consistency condition for reads/writes over semantic projections rather than bitwise state
    Defined as classical linearizability applied to E(S) with ϵ-compatibility (§3.2).
  • Bounded Eventual Coherence no independent evidence
    purpose: Bound expected semantic divergence among correct replicas under stated assumptions
    Theorem 1 neighborhood bound; depends on contractive graft and δerr.
  • Epistemic delta Δe and graft operator ⊕ no independent evidence
    purpose: Propagate essential beliefs with justifications instead of raw token traces
    Protocol primitives central to ESR efficiency and coherence claims (§4).
  • Verifiable Semantic Rollback / partial-dependency repair no independent evidence
    purpose: Prune faulty premises and repair partially dependent beliefs aligned with physical recovery
    Recovery operator claimed to prevent secondary cognitive faults (§5).

pith-pipeline@v1.1.0-grok45 · 22308 in / 3862 out tokens · 39594 ms · 2026-07-14T16:31:07.735012+00:00 · methodology

0 comments
read the original abstract

In distributed systems, the classical State Machine Replication (SMR) model assumes that correct replicas execute deterministic transitions to yield identical bitwise states. However, the rise of agentic distributed systems -- where autonomous, stochastic, and model-driven agents orchestrate infrastructure -- presents scenarios where deterministic, bitwise replication is insufficient. Replicas operating with generative models may exhibit divergent reasoning paths, summaries, and token boundaries, yet reach semantically equivalent and correct operational decisions. Forcing bitwise agreement across these stochastic participants degrades execution flexibility, induces context amnesia, and limits performance. We argue that in such settings replicas should agree on belief, not bits. We propose Epistemic State Replication (ESR), a belief-replication layer for agentic distributed systems that shifts the replication boundary from data visibility to knowledge visibility. We formalize the epistemic node state as a pair K = (L, B) separating the deterministic, immutable evidence log (L) from the stochastic, evolving belief lineage (B). To govern execution safety, we define Semantic Linearizability, which requires operations to reflect the latest committed operational meaning within a verifier-bounded semantic compatibility metric, and Bounded Eventual Coherence, which bounds expected semantic divergence under fair delivery, monotonic evidence, bounded verifier disturbance, and a contractive graft operator. We outline protocols for propagating derived insights using structured epistemic deltas, and formalize Verifiable Semantic Rollbacks to prune faulty premises from belief lineages without inducing context amnesia. We prototype ESR and report preliminary simulation results that show feasibility under the stated assumptions and illustrate reductions in secondary cognitive faults.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

25 extracted references · 13 linked inside Pith

  1. [1]

    Schneider

    Fred B. Schneider. Implementing fault-tolerant ser- vices using the state machine approach: A tutorial. ACM Computing Surveys (CSUR), 22(4):299–319, 1990. 14

    1990

  2. [2]

    In search of an understandable consensus algorithm

    Diego Ongaro and John Ousterhout. In search of an understandable consensus algorithm. InProceedings of the 2014 USENIX Conference on Usenix Annual Technical Conference (ATC), pages 305–320, 2014

    2014

  3. [3]

    Bernstein, Vassos Hadzilacos, and Nathan Goodman.Concurrency Control and Recovery in Database Systems

    Philip A. Bernstein, Vassos Hadzilacos, and Nathan Goodman.Concurrency Control and Recovery in Database Systems. Addison-Wesley, 1987

    1987

  4. [4]

    The post-deterministic manifesto: A new foundation for trustworthy autonomous infrastructure

    Jun He and Deying Yu. The post-deterministic manifesto: A new foundation for trustworthy autonomous infrastructure. arXiv preprint arXiv:2606.01722, 2026

    Pith/arXiv arXiv 2026

  5. [5]

    Herlihy and Jeannette M

    Maurice P. Herlihy and Jeannette M. Wing. Lin- earizability: A correctness condition for concur- rent objects.ACM Transactions on Programming Languages and Systems (TOPLAS), 12(3):463–492, 1990

    1990

  6. [6]

    Dynamo: Amazon’s highly available key-value store

    Giuseppe DeCandia, Deniz Hastorun, Madan Jam- pani, Gunavardhan Kakulapati, Avinash Lakshman, Alex Pilchin, Swaminathan Sivasubramanian, Peter Vosshall, and Werner Vogels. Dynamo: Amazon’s highly available key-value store. InProceedings of the Twenty-First ACM SIGOPS Symposium on Operating Systems Principles, pages 205–220, 2007

    2007

  7. [7]

    Conflict-free replicated data types

    Marc Shapiro, Nuno Preguiça, Carlos Baquero, and Marek Preguiça. Conflict-free replicated data types. Symposium on Self-Stabilizing Systems, pages 386– 400, 2011

    2011

  8. [8]

    Provenance in databases: Why, how, and where next?ACM SIGMOD Record, 38(4):37–47, 2009

    James Cheney, Tom Chothia, and Joan Feigenbaum. Provenance in databases: Why, how, and where next?ACM SIGMOD Record, 38(4):37–47, 2009

    2009

  9. [9]

    Holland, Uri Braun, and Margo Seltzer

    Kiran-KumarMuniswamy-Reddy,DavidA. Holland, Uri Braun, and Margo Seltzer. Provenance-aware storage systems. InProceedings of the USENIX Annual Technical Conference (ATC), pages 43–56, 2006

    2006

  10. [10]

    Re- act: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2022

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. Re- act: Synergizing reasoning and acting in language models.arXiv preprint arXiv:2210.03629, 2022

    Pith/arXiv arXiv 2022

  11. [11]

    Reflexion: Language agents with ver- bal reinforcement learning.arXiv preprint arXiv:2303.11366, 2023

    Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath,KarthikNarasimhan,andShunyu Yao. Reflexion: Language agents with ver- bal reinforcement learning.arXiv preprint arXiv:2303.11366, 2023

    Pith/arXiv arXiv 2023

  12. [12]

    O’Brien, Carrie J

    Joon Sung Park, Joseph C. O’Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, and Michael S. Bernstein. Generative agents: Inter- active simulacra of human behavior.arXiv preprint arXiv:2304.03442, 2023

    Pith/arXiv arXiv 2023

  13. [13]

    Voyager: An open-ended em- bodied agent with large language models.arXiv preprint arXiv:2305.16291, 2023

    Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Man- dlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended em- bodied agent with large language models.arXiv preprint arXiv:2305.16291, 2023

    Pith/arXiv arXiv 2023

  14. [14]

    Patil, Ion Stoica, and Joseph E

    Charles Packer, Sarah Wooders, Kevin Lin, Vivian Fang, Shishir G. Patil, Ion Stoica, and Joseph E. Gonzalez. Memgpt: Towards llms as operating sys- tems.arXiv preprint arXiv:2310.08560, 2023

    Pith/arXiv arXiv 2023

  15. [15]

    Langchain.https://github.com/ langchain-ai/langchain, 2022

    Harrison Chase. Langchain.https://github.com/ langchain-ai/langchain, 2022

    2022

  16. [16]

    Retrieval-augmented generation for knowledge-intensive nlp tasks.arXiv preprint arXiv:2005.11401, 2020

    Patrick Lewis, Ethan Perez, Aleksandra Pik- tus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-augmented generation for knowledge-intensive nlp tasks.arXiv preprint arXiv:2005.11401, 2020

    Pith/arXiv arXiv 2005

  17. [17]

    Self-rag: Learning to re- trieve, generate, and critique through self-reflection

    Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. Self-rag: Learning to re- trieve, generate, and critique through self-reflection. arXiv preprint arXiv:2310.11511, 2023

    Pith/arXiv arXiv 2023

  18. [18]

    Semantic caching for language model inference.arXiv preprint arXiv:2311.02535, 2023

    Zichun Guo et al. Semantic caching for language model inference.arXiv preprint arXiv:2311.02535, 2023

    Pith/arXiv arXiv 2023

  19. [19]

    Long context rag per- formance of large language models.arXiv preprint arXiv:2411.03538, 2024

    Quinn Leng, Jacob Portes, Sam Havens, Matei Za- haria, and Michael Carbin. Long context rag per- formance of large language models.arXiv preprint arXiv:2411.03538, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Yu Philip

    ShaoxiongJi,Shirui Pan,ErikCambria,PekkaMart- tinen, and S. Yu Philip. A survey on knowledge graphs: Representation, acquisition, and applica- tions.IEEE Transactions on Neural Networks and Learning Systems, 33(2):494–514, 2021

    2021

  21. [21]

    Poisoning language models during instruc- tion tuning

    Alexander Wan, Eric Wallace, Sheng Shen, and Dan Klein. Poisoning language models during instruc- tion tuning. InProceedings of the International Conference on Machine Learning (ICML), pages 35413–35426, 2023

    2023

  22. [22]

    Poisoning retrieval corpora by injecting adversarial passages.arXiv preprint arXiv:2310.19156, 2023

    Zexuan Zhong, Ziqing Huang, Alexander Wettig, and Danqi Chen. Poisoning retrieval corpora by injecting adversarial passages.arXiv preprint arXiv:2310.19156, 2023

    Pith/arXiv arXiv 2023

  23. [23]

    Multiagent sys- tems: A survey from a machine learning perspective

    Peter Stone and Manuela Veloso. Multiagent sys- tems: A survey from a machine learning perspective. Autonomous Robots, 8(3):345–383, 2000. 15

    2000

  24. [24]

    Autogen: Enabling next-gen llm applications via multi-agent conversa- tion.arXiv preprint arXiv:2308.08155, 2023

    Qingyun Wu,Gagan Bansal,Jieyu Zhang,Yiran Wu, Beibin Li, Erkang Zhu, Chi Li, Shubham Toshniwal, Ning Zhang, Michel John, et al. Autogen: Enabling next-gen llm applications via multi-agent conversa- tion.arXiv preprint arXiv:2308.08155, 2023

    Pith/arXiv arXiv 2023

  25. [25]

    Tenenbaum, and Igor Mordatch

    Yilun Du, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. Improving factual- ity and reasoning in language models through mul- tiagent debate.arXiv preprint arXiv:2305.14325, 2023. 16

    Pith/arXiv arXiv 2023