- Rating
- 2
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 2
Read review
Paper summary
Persistent research agents keep state across sessions: sources become notes, notes become report evidence, and evidence may be compressed into memory or turned into a reusable procedure. The paper argues that blocking an unsafe final output is not enough, because contaminated or untrusted content can still be stored and reused later. A safe answer does not certify that the agent’s stored state is safe. The paper proposes a governance layer that tracks every persistent object’s provenance and re-checks it at write, recall and release time.
Critical strengths
The paper shows that a safe output does not mean the agent’s stored state is safe. Without the promotion gate, trust and contamination errors rise sharply while unsafe export stays at 0.01. This exposes a blind spot in output-only defenses and evaluations.
Provenance tracking with separate labels is backed by formal guarantees. Full governance clearly beats all baselines on persistent-state metrics without losing utility, across three LLMs and a ten-session extension.
Critical weaknesses
The guarantees and the experiments don’t line up. The formal guarantees require complete lineage and semantically sound certificates. The main experiments instead use selective attribution and never measure the semantic verifier’s error rate, so the theory and the empirical results support each other less than the paper suggests.
Key lifecycle features are untested. Revocation, policy-epoch shifts and resource-version drift are central to the revalidation design, but the experiments never exercise them. The most distinctive part of the framework, re-checking evidence when things change, is therefore unvalidated.
The reuse-risk analysis is illustrative only. Certificate error rates p_c are set by hand rather than estimated from data, and the adaptive exposure budget (Prop. 4) is never implemented. The analysis of shared certificates (Props. 3–4) is mathematically interesting but not shown to matter for a real verifier.
The evaluation is self-constructed. DRSC-Bench, the five attack families, the simulated restricted data and the metric definitions were all designed by the authors, and there are no adaptive attacks. Strong results on a benchmark built for the method may not carry over to realistic or adversarially adapted settings.
STP labels lack validation. Silent Trust Promotion depends on judging “inappropriate eligibility,” yet no annotation protocol or inter-annotator agreement is reported for it. Human validation covers only report quality (utility), not the safety labels that support the main claims.
Baselines are not like-for-like. The Fides and PRUDENTIA comparisons are the authors’ own adaptations, with different serialization scope and planners. Part of the reported gain may come from implementation differences rather than from the proposed design, and the paper acknowledges the residual gap is not attributed to any one mechanism.
Several components are under-specified. The empirical risk score ρ, the AVS/η context-selection heuristic, and the hand-set ordinal ingress rules all affect what gets admitted, but their calibration and their effect on the safety–utility trade-off are not well characterized. This makes the system hard to reproduce or tune.
The system is complex and costly. The framework adds many components (three label lattices, certificates, witness closures, four gates) and raises median latency by about 12%. The paper offers little evidence about which components are essential, or how easily it integrates with existing agent frameworks.
The presentation is dense. Concepts such as audit dependencies vs. derivation dependencies vs. witness closures, and AVS vs. η, are introduced with heavy notation but loosely tied to the results. A single running example traced through all the gates would make the paper much easier to follow.
Positioning against related work is limited. The paper could say more about how the governance layer composes with other runtime or write-time state defenses, and how it relates to proof-carrying code, capability systems and provenance-aware release frameworks. This would clarify what is new versus adapted from classic information-flow control.
Questions for authors
How exactly is STP adjudicated at the task-run level—what constitutes “inappropriate eligibility” and what annotation or rule set ensures consistency across domains and runs?
How are certificates produced in practice in your system—are they LLM-generated endorsements, human-issued signatures, or programmatic checks—and what empirical evidence do you have about their semantic error rates p_c?
Could you report a revocation/epoch-change experiment (e.g., expiring or revoking a frequently reused endorsement mid-benchmark) to demonstrate inherited-witness invalidation and recovery behavior?
What safeguards prevent selective attribution from omitting influences (e.g., model under-reporting) in real deployments—do you support conservative capture modes by default, and what is the measured utility/safety trade-off at scale?
How is the empirical risk score ρ calibrated and updated? Is it per-object static (ingress-only) or updated via subsequent observations, and how sensitive are results to λ_risk in context assembly?
Can you quantify the end-to-end overhead (latency and storage) of maintaining witnesses and running Current checks at recall/release under heavier loads or deeper dependency trees?
How does the governance layer interact with executable capabilities and tool permissions in the control plane—e.g., can a high-trust endorsement for “report evidence” mistakenly authorize code execution, and how is this ruled out in practice?
Lifecycle tests, verifier measurements and capture safeguards
Thank you. The uploaded revision provides these tests and decision rules.
Q1. STP adjudication. STP is positive when a designated unsupported claim is retained and eligible for higher-trust reuse without adequate verification at its task-run probe. Retention follows task annotations; eligibility requires verified-or-higher trust and all admission gates, or the corresponding evidence-pool admission for lattice-free systems. All 1,200 probes, including blocked writes, remain in the denominator. Two raters assessed 100 stratified task-runs: 94.0% agreement, inter-rater κ=0.86; automated precision/recall against consensus is 0.91/0.89 (Appendix L.3). This validates STP separately from report quality.
Q2. Certificates and mechanism isolation. The agent requests endorsement; the trusted runtime signs after assessing cross-source agreement, citation support and source-signature information. Certificates bind subject/support versions, transition, permitted contracts, issuer, epoch and expiry; Eq. 16 checks applicability mechanically. The 63 semantically incorrect certificates among 1,840 issued (p^=3.42%) average 3.2 downstream reuses each (about 202 certificate-use exposures).
The 240-task, five-run matched control fixes persistent labels, object-level scope-aware recall, verifier and maximum per-task verification budget, removing inherited witnesses and their revalidation. Full governance reduces STP/CSC from 0.10/0.07 to 0.03/0.02; utility rises from 0.73 to 0.74 and coverage from 0.94 to 0.96 (Table 2). A four-use cap per endorsement charges all dependent admissions across sessions; with the same verifier it reduces promotion-probe STP from 0.06 to 0.03. The audit measures pooled error; Prop. 4 uses pre-issuance conditional error bounds.
Q3. Lifecycle changes and recovery. Table 3 uses 60 tasks, 20 per condition, with three runs; changes occur between sessions 2 and 3:
Static labels block 0% in each condition. Fresh derivation from valid support restores utility after inherited invalidation (Prop. 5).
Q4. Attribution and capture. Independent annotation finds omissions in 3.8% of selective writes during normal operation. Conservative capture records every governed input visible during generation for complete boundary capture.
All use 240 tasks and five runs (Table 13). Core comparisons use selective attribution; hybrid is the deployment default, applying conservative capture when untrusted/adversarial content is visible or a procedural draft is written.
Q5. Risk, ingress and context. ρ combines an injection classifier with structural features, is fixed per immutable version and propagates by parental maximum; certified sanitization creates a new version for reassessment. It governs admission (ρ≤0.2) and relevance-minus-risk ranking; label, scope and certificate checks are hard gates. At η=0.20, sweeping λrisk∈{0.25,0.50,0.75,1.00} gives utility 0.72-0.74, STP 0.03 and CSC 0.02-0.03 (Table 10). Eq. 33 initializes source trust using ordinal adjustments δev∈{−1,0,1}, capped at verified; derived-content promotion requires a certificate. AVS gives the population information-bottleneck view; η stops the empirical selector by residual need.
Q6. Overhead. Indexed closures check all inherited certificates without ancestry traversal (Table 23):
Metadata is 29.8% of content bytes, about 1.4 KB/object. End-to-end median latency rises from 41.8 to 46.9 s (+12.2% versus Naive RAG); ten-session P50 overhead is +13.4%.
Q7. Capabilities and positioning. Certificate applicability checks the contract against Uc. Execution additionally requires control-plane authorization: requested operations must lie in K, with resource versions ν and composition constraints β checked independently. In the 12-procedure Python sandbox, unsafe-capability rates are 0.22/0.14/0.01 for ungoverned/A-MemGuard/full governance (Table 19). As in proof-carrying code, production and checking are separate; here, justifications follow derived state. The layer composes with admission filters and tool-permission guards.
Five attack families (Table 17), three backbones (Table 18), repeated promotion probing (Appendix H) and the ten-session trajectory (Figure 2) test persistent reuse. We would appreciate your reassessment.