When AI Agents Commit: Cognitive Serializability Across Data, Evidence, Policy, and Authority
cs.AI · cs.DC
Quick verdict
TCT makes AI-agent commits verifiable against volatile inputs; theory is careful, but evaluation is only self-designed conformance tests plus ablations.
Summary
The paper addresses the gap between stochastic, long-latency AI-agent reasoning and deterministic database commit: an agent may derive a mutation from database reads, retrieved evidence, policy, and delegated authority, all of which can change before the effect becomes durable. It proposes TCT, a typed dependency contract enforced by a trusted admission-and-commit gate. The strict profile TCT-S requires every value exposed to derivation to remain unchanged at a common logical event extending through durability, yielding a serializability-style guarantee (Theorem 1) via guard-first strict 2PL and an acyclic lock-point order. The weaker TCT-C recertifies the concrete effect against a simultaneously protected current dependency vector and current policy using a registered joint predicate. Four mechanisms (co-transactional guards, envelope-bound witnessed grants, atomic-commit participants, hard-real-time certificates) preserve external validity to the durability event; co-committed receipts gate belief activation and idempotent grant finalization. A PostgreSQL 16 prototype passes all 28 self-designed falsification histories (100 runs each) without admitting an anomaly, adds 3.22 ms mean commit overhead, and reaches 1,842 txns/sec (TCT-C) at 32 threads; at 20% drift TCT-C recertifies 92.8% of proposals versus TCT-S's 80.1% commit rate. Limitations: only self-ablation baselines (no head-to-head with Mnemosyne, CommitGuard, or Atomix), one workload and machine, no variance reporting, and strong closed-input/trusted-mediation assumptions the authors concede do not fail closed under dependency omission.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Under Assumptions 1-7 (trusted capture, registry, guard-first strict 2PL, complete participation, observation rule, external validity, durability), every TCT-S commit admits a serial order and a logical event at which all derivation inputs were simultaneously valid through durability (Theorem 1).
- Effect-Compatible Cognitive Admission (TCT-C) lets drifted proposals commit when a registered joint predicate accepts the complete, simultaneously protected current dependency vector and current policy, without serializing the original stochastic derivation.
- Database serializability and ATP-style constraint gates alone cannot establish a common valid point for a mutation and its derivation inputs (Proposition 1), and no protocol can be both zero-error sound and progress-guaranteeing over gate-indistinguishable history pairs (Lemma 1).
- A PostgreSQL 16 prototype enforces both profiles, producing the specified admission outcome on all 28 falsification histories (100 runs each) while adding 3.22 ms mean commit overhead and sustaining 1,842 txns/sec at 32 threads.
Strengths
Precise, honest separation of correctness profiles
TCT-S (Definition 5, derivation-faithful serializability) vs TCT-C (Definition 6, effect recertification only) is kept rigorously distinct; Proposition 1 (constraint-relative non-implication) and Lemma 1 (observational-equivalence boundary) are explicit negative results that pre-empt overclaiming about what constraint gates can guarantee.
Exceptional implementation disclosure
Algorithm 1 is fully specified and Appendix A (Listings 1-6) gives complete PostgreSQL DDL for guard materialization, permanent identity bindings, envelope/discharge tables, grant lifecycle (RESERVED/CONSUMED/RELEASED with release-proof tables), receipts, and outbox — rare depth for a systems-concept paper and a strong reimplementation substrate.
Falsification-suite methodology includes implementation-order conformance tests
Candid interpretation boundary
Theorem 1 rests on sound classical substrate
Weaknesses
No head-to-head evaluation against closest related systems
Closed-input capture assumption may be unrealizable for LLM agents
Thin statistical reporting
Liveness under failure is asserted, not measured
Novelty is composition and typing rather than new theory
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper targets the epistemic-to-ACID gap (Section 2.1): LLM agents derive durable mutations from volatile inputs — database reads, retrieved evidence, policy, delegated authority — none of which the DBMS can fence. TCT is a typed dependency contract enforced by a trusted admission-and-commit gate, with two profiles. TCT-S (Definition 5) demands a database serial order plus a logical event s at which every mediated derivation value is unchanged, with validity fences held until the runtime durability event d. TCT-C (Definition 6) only requires the concrete effect to pass a registered joint recertifier Phi over a simultaneously protected current dependency vector and current policy. Central claims: an acyclic lock-point order implies strict cognitive serializability (Theorem 1, under Assumptions 1-7); ATP constraint sets that omit dependencies or commit-spanning fences imply neither profile (Proposition 1); effect-receipt correspondence and no premature belief activation (Theorems 2-3); and a Lemma-1 impossibility showing no protocol can be both zero-error sound and progress-guaranteeing over gate-indistinguishable history pairs. The scope is explicitly conditional: Section 10 enumerates the TCB (mediators, canonicalizer, registries, resolvers, gate, DBMS, issuer validity semantics) and concedes an omitted dependency is not guaranteed to fail closed.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is Cognitive Serializability?
It is this paper's correctness notion for AI-agent mutations: a committed agent effect must admit a serial database order plus a logical event at which every value the agent used during reasoning (database state, evidence, policy, observed authority) was simultaneously unchanged, with that stability lasting until the durable commit actually lands. The strict profile TCT-S enforces this; the weaker TCT-C only re-validates the final effect against current state and policy.
How much overhead does TCT add to agent commits?
In the authors' PostgreSQL 16 prototype, the guard-first commit protocol added a mean of 3.22 ms per commit (median 2.85 ms, 99th percentile 8.41 ms), which they note is under 0.5% of typical multi-second agent reasoning latency. Peak throughput measured was 1,842 transactions/sec for the compatible profile at 32 worker threads on a 16-core Xeon Gold server.
What problem do AI agents have with database transactions?
An agent may reason for seconds before proposing a write, using inputs (database rows, retrieved documents, policies, permissions) that live outside the DBMS and can change meanwhile. A serializable commit can then be internally valid but justified by stale evidence, revoked approvals, or old policy — a time-of-check-to-time-of-use gap the paper formalizes as four anomaly classes, including premature belief updates like an internal 'order placed' state after an aborted commit.
Is the TCT prototype code available?
The paper fully specifies the commit algorithm and includes the complete PostgreSQL metadata schema in its appendix, but no public code or artifact release is stated, so reproducing the results currently requires reimplementing the gate from the paper and reconstructing the benchmark workloads.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Text extraction truncated several sections (parts of Sections 3, 5, 8, 10.2, and the conclusion), so some protocol details, Proposition 2, and the full benchmark description were unavailable.
- Figures 1-3 could not be inspected (text-only extraction); throughput and drift curve shapes are taken from surrounding prose only.
- Several key references (Mnemosyne, CommitGuard, Atomix, ATCC, CoAgent, S-Bus) are 2026 arXiv preprints, not peer-reviewed; novelty claims could not be cross-verified offline.
- Appendix listings were heavily truncated near the end; schema completeness beyond Listings 1-6 could not be fully verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.