- Rating
- 2
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 3
Read review
Paper summary
The paper studies when LLM memory produces the same behavior under reordered and retried event delivery. COVER interprets immutable events using their exact causal ancestors, validates extracted records, preserves unresolved source material, and uses deterministic merging and readout. The theoretical analysis separates schedule invariance from semantic correctness.
Critical strengths
The most useful contribution is the separation between interpretation and storage: a convergent merge cannot repair an event interpretation that already depended on an arbitrary concurrent arrival. The causal-ancestor interface makes the required assumptions explicit, including fixed compilation and trustworthy dependency information. This is a meaningful systems-design contribution even where the underlying algebra is classical.
Critical weaknesses
-
The empirical need for ancestor-restricted interpretation is less directly demonstrated than the general need for deterministic handling of delivery. Exact-ID deduplication already removes all correctness flips in the retry experiment, and Appendix H’s cached canonical control matches COVER’s record maps and compilation lookups. These are appropriate controls, but leave less direct evidence about how often interpreting an event against concurrent, non-ancestor records causes a consequential failure in realistic traces.
-
The learned representation does not yet have a consistently beneficial operating configuration. In the separate official-code comparison in Section 5.3, typed COVER averages 26.12 versus 44.20 for raw sources; the two-call cascade also loses to Raw-2 for Llama in Appendix K. The paper acknowledges this, so I view it as a limitation on practical significance rather than a contradiction of the convergence guarantees. Total compilation and readout cost is also not matched by equalizing reader-call limits.
Questions for authors
Which experiment compares ancestor-conditioned compilation with current-store-conditioned compilation while keeping the model, events, merge rule, and readout fixed? If there is no such comparison, which empirical result supports the specific benefit of excluding non-ancestor context?
Where do the parent links come from in each evaluation? Which results assume externally supplied causal metadata, and which test dependencies inferred from the observations?
Are replicas reusing the same compiled records or independently invoking the LLM? In the latter case, what enforces the fixed-compiler assumption required by the convergence result?
Do the diagnostics identify where the accuracy loss in Section 5.3 occurs: compilation, retention, retrieval, or interpretation by the reader?
What are the total token and model-call costs of the facts/raw cascade relative to Raw-2, including amortized compilation? Equal reader-call limits do not establish matched end-to-end cost.