{"id":"d27ae18e-dae8-4b43-84a5-391132cc4987","arxiv_id":"2607.20560","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A three-layer RAG pipeline with exponential temporal decay and an STL freshness gate improves temporal retrieval, but its decay coefficient is tuned post hoc, not derived from the claimed Bayesian grounding.","lead":"Chronofy adds time-stamped decay to retrieval-augmented generation, so a lab result from yesterday counts more than one from six months ago. It is a plausible recipe for time-sensitive RAG, but its main decay parameter is chosen by grid search per dataset, and the freshness gate underperforms the LLM's own confidence signal.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim rests on exponential validity decay (Eq. 2/3) that the paper concedes misses non-monotonic lifecycles and regime shifts; β is grid-searched, not learned or validated against OU rates, so retrieval and STL gate inherit unspecified error.","rationale":"The reader's weakest_assumption identifies the exponential decay factorization in Eq. (2)/(3) as the load-bearing premise, and this is exactly where my stress-test lands. The paper itself acknowledges the limitation in §V-B, making this an internal inconsistency rather than a mere disagreement with outside consensus. The ablation (Table VII) confirms Layer 2 is the dominant contributor, so an error in the decay function propagates directly to the main reported gains.\n\nThe proposed test is concrete and achievable: fitting a Weibull model on GDELT and comparing MRR would falsify or support the exponential assumption. If the Weibull model performs comparably or better, the paper's claim that exponential decay is the appropriate form (and the Proposition 1 OU grounding) is not supported. If, however, exponential remains competitive, the concern would be partially mitigated. The STL layer's underperformance (AUROC 0.479 vs 0.556) and the +2% end-to-end gain further weaken the central claims, but the exponential-decay assumption is the foundational issue.\n\nI agree with the reader's verdict. The paper is well-organized and the architecture is coherent, but the central claim of principled temporal decay modeling is not established beyond a grid-searched exponential on selected benchmarks. The verdict should remain REJECT. No new concern changes this assessment.","tokens_in":8780,"tokens_out":6217,"duration_ms":73082,"concrete_test":"On GDELT, replace the exponential decay exp(−β_j Δt) in Eq. (2) with a per-relation Weibull survival function (shape and scale fitted on the training split, same parameter budget) and re-run the retrieval experiment. If the Weibull-based MRR exceeds the reported 0.293, or if the per-relation optimal shape parameters are far from 1 (which would indicate exponential is a poor fit), then the exponential assumption is not neutral and the reported gains depend on the benchmark's coincidence with Eq. (2).","verdict_should_be":"UNCHANGED","load_bearing_attack":"The architecture's behavior is driven by Eq. (2)/(3): the traversal weight and the STL validity signal factor as q_e · c(tr_j) · exp(−β_j Δt), with a single β_j per fact type. For the central claim to hold, this exponential decay must approximate the true usable lifetime of facts. The paper concedes the opposite in §V-B: 'The exponential form cannot capture non-monotonic lifecycles or sudden regime shifts.' That is exactly the regime of breaking news, legal supersession, and clinical lab values after an intervention. If true validity is instead a step function or has delayed onset, both the retrieval ranking and the Layer-3 gate (which is monotone in βΔt) are systematically miscalibrated for the cases where temporal hallucination is most dangerous.\n\nProposition 1 does not rescue this: β=2κ is optimal only under the OU process (Eq. 8), a specific stationary Gaussian model. No experiment estimates κ_j or validates the OU assumption. Instead, β* is obtained by grid search per benchmark (Tables I, II, III, V), and §V-B admits 'cold-start calibration in novel domains remains open.' The grid search can hide functional-form misspecification by tuning β on a per-benchmark basis.\n\nThe end-to-end evidence does not close this gap: the parsed variant gains only +2.0% (Table VI) over vanilla, and the STL robustness AUROC (0.479) is below LLM confidence (0.556), while recovering only 24.6% of confidently-wrong outputs. Thus the claims of reduced temporal hallucination and principled re-acquisition triggers rely on a validity model whose central assumption is confessed to be inadequate for the phenomena the paper targets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Chronofy, a three-layer temporal RAG architecture: Layer 1 embeds timestamps in a Matryoshka-style temporal subspace, Layer 2 weights graph traversal by exponential decay exp(-β Δt) and an epistemic threshold τ, and Layer 3 applies STL robustness to a knowledge-validity signal, abstaining when the weakest-link validity falls below γ. The authors claim that this reduces temporal hallucination, improves retrieval, and grounds the decay coefficient β as twice the mean-reversion rate of an OU process (Proposition 1), and they provide a weakest-link bound on output confidence (Theorem 1). Experiments are reported on ICEWS14, GDELT, TimE-Lite News, and MIMIC-IV, with end-to-end RAG accuracy and ablations.","tokens_in":9188,"tokens_out":6260,"duration_ms":69963,"significance":"The problem is real and applying STL robustness to knowledge freshness rather than LLM confidence is a worthwhile direction. The open-source release and the explicit treatment of fact age as a first-class representation feature are strengths. If the formal grounding and experiments were solid, this would be a useful contribution. However, the central claims are not established as reported: β is grid-searched rather than learned or estimated from an OU process, Theorem 1 is a definitional consequence of the chosen possibilistic semantics, and the STL gate underperforms LLM confidence while the realistic end-to-end gain is +2.0%. The paper's own limitation section concedes that the exponential decay form cannot capture non-monotonic lifecycles or sudden regime shifts, which are exactly the settings where temporal hallucination is most dangerous.","major_comments":[{"comment":"The architecture's behavior is driven by the memoryless exponential decay w = q·c·exp(-β Δt) and v(s_i)=min_e q_e exp(-β Δt). The paper itself (§V-B) concedes that this form cannot capture non-monotonic lifecycles or sudden regime shifts. No experiment estimates the true validity function or compares against step, Weibull, or non-parametric alternatives on the reported benchmarks; the statement that Weibull gave only 'marginal gains' is not quantified. Because the same exponential form feeds both the retrieval weights and the STL gate, misspecification propagates systematic error into both layers. The central claim is therefore conditional on an assumption the paper admits is violated in relevant settings.","section":"Eq. (2), Eq. (3), §V-B"},{"comment":"β is presented as 'learnable' and grounded as 2κ of an underlying OU process, but no experiment estimates κ or σ, and no diagnostic validates the OU assumption. In Tables I, II, III, and V, β* is obtained by sweeping a grid and picking the best MRR. That is benchmark-specific tuning, not parameter learning, and it makes Proposition 1 a post-hoc labelling rather than a validated grounding. The per-relation heterogeneity of β* in Table III could equally be absorbing functional-form misspecification. The admission in §V-B that 'cold-start calibration in novel domains remains open' is therefore central, not peripheral: the paper provides no procedure to obtain β in a new domain without a labeled validation set.","section":"Proposition 1, Tables I-III, V"},{"comment":"The proof of the weakest-link bound is definitional. After assigning each premise the necessity degree V(e_i,T_q)=q_i exp(-β Δt) and adopting min-aggregation and possibilistic Modus Ponens, the conclusion that output confidence is bounded by min_i V_i follows from the semantics. The theorem does not connect these formal degrees to the LLM's actual output confidence, nor does it justify the assumption that LLM reasoning combines evidence via the min-t-norm. Calling Eq. (6) a 'formal guarantee' overstates what has been shown: it is a consistency property of the chosen representation, not an empirical or independently verifiable bound on hallucination.","section":"Theorem 1, Eq. (6)-(7)"},{"comment":"The evidence for the STL gate is weak. On TimE-Lite News, the STL robustness score has AUROC 0.479 versus 0.556 for LLM confidence, and the gate catches only 24.6% of confidently-wrong outputs. The paper attributes this to variance compression in the closed-context benchmark, but no experiment shows that the gate discriminates in an open-context setting. When the full pipeline is active, Table VIII and the surrounding text state that the gate 'rarely triggers' (mean ρ=0.965). The abstract's claim that the system 'reduces temporal hallucination' is not supported by these results; at best, Layer 3 is a post-hoc filter whose reported discrimination is below the LLM's own confidence signal.","section":"Table IV, §IV-C"},{"comment":"The Layer 1 ablation does not evaluate the mechanism proposed in Section III-B. The paper states that the ablation uses a 'simplified sinusoidal positional encoding that requires no training' rather than the LoRA-trained temporal subspace with CKA contrastive loss. Furthermore, the full model's optimal temporal subspace is t=4, but no experiment studies the trained temporal projection at any dimension. Consequently, the contribution of Layer 1 as designed is not actually measured, and the '+73.9%' full-pipeline gain cannot be attributed to the proposed temporal-subspace component.","section":"§IV-F, §III-B"}],"minor_comments":[{"comment":"The example 'β≈5.0 for vital signs' contradicts Table V, where vital signs have optimal β*=0.001. Please fix the inconsistency.","section":"§III-C vs Table V"},{"comment":"Please define Gold@5 and Accuracy. Also, the parsed variant has Gold@5 0.337, slightly below vanilla's 0.341, so the +2.0% accuracy gain should not be described as confirming retrieval-level improvement.","section":"§IV-E, Table VI"},{"comment":"The notation c(tr_j) is called 'base semantic confidence' but its computation is not specified. The threshold τ in the epistemic filter is swept, not learned; this should be stated alongside the β-sweep.","section":"§III-C"},{"comment":"The figure appears as a text block in the current version. It should be rendered as an actual architecture diagram with readable labels.","section":"Figure 1"}],"recommendation":"reject","confidential_remarks":"The paper has multiple load-bearing issues that cannot be fixed by local editing: the central decay model is admitted to be misspecified in the targeted regimes, β is not learned or grounded empirically, the formal bound is definitional, and the STL gate's reported discrimination is below the LLM confidence baseline. Substantial new experiments and a reworked formal claim would be needed; at present the manuscript is not ready for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Chronofy is a well-written, honest engineering paper. The genuinely new move is applying STL robustness to knowledge freshness rather than LLM confidence—a modest but real re-targeting of Mao et al. The three-layer stack (temporal subspace, decay-weighted graph retrieval, STL gate) is a sensible synthesis of TMRL, TempValid, and STAR-RAG, and the authors deserve credit for open-sourcing the code and for a limitations section that actually names the cracks.\n\nThe retrieval results are real. Decay-weighted PPR beats static retrieval on ICEWS14 (MRR 0.608 vs 0.556) and on GDELT (0.293 vs 0.197), and the per-relation beta heterogeneity in Table III is consistent with the idea that fact-types age differently. The MIMIC-IV sensitivity analysis, with stable diagnoses at beta*=0 and lab values decaying fastest, is a nice sanity check.\n\nThat said, the abstract's central claims are not supported. The one parameter that drives the gains—beta—is grid-searched per benchmark, not learned. \"Learnable decay functions\" are not what is evaluated. Proposition 1 grounds beta as 2-kappa under an OU process, but no experiment estimates kappa or tests the OU assumption. More importantly, Section V-B concedes the exponential form cannot capture non-monotonic lifecycles or sudden regime shifts—exactly the regime where stale evidence is most dangerous. The grid search can hide functional-form misspecification by tuning a single constant per dataset.\n\nThe end-to-end evidence is thin. The parsed variant gains 2.0% over vanilla, the STL robustness AUROC (0.479) is below LLM confidence (0.556), and the gate catches only 24.6% of confidently-wrong outputs. The re-acquisition trigger is described but not evaluated as an intervention. Theorem 1 is a correct formal restatement of possibilistic min-aggregation—it is not an empirical claim about LLM behavior.\n\nNone of this is fatal to the paper as a retrieval-engineering contribution. Decay-weighted retrieval is a useful design pattern, and the STL-as-freshness-gate idea is worth exploring. But as it stands, the headline promises more than the validation delivers. The paper deserves a serious referee, not a desk reject: the architecture is concrete, the code is public, and the failure modes are clearly stated. I would send it to review with a major-revision bar—learn beta or measure kappa, validate the decay form on a case where non-monotonicity matters, or rescope the claims to retrieval precision. A reader looking for a cure to temporal hallucination should not take the abstract at face value.","headline":"A sensible, honest engineering paper whose retrieval gains are real, but the headline claims about temporal hallucination and re-acquisition lean on a grid-searched beta and an exponential-decay assumption the paper itself admits is wrong where it matters most.","tokens_in":9701,"tokens_out":2886,"would_cite":false,"duration_ms":30703,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Chronofy embeds exponential decay into every RAG layer, bounding answers by their stalest evidence.","keywords":["temporal hallucination","retrieval-augmented generation","exponential decay","Signal Temporal Logic","weakest-link bound","Ornstein-Uhlenbeck process","knowledge graph forecasting","information validity"],"falsifier":"Collect timestamped facts with human-annotated usefulness over time and check whether log-usefulness falls linearly with age; if validity flattens, spikes after re-publication, or follows a Weibull shape, Eq. (2)/(3) is violated and both the retrieval weights and the STL gate inherit systematic error.","tokens_in":8624,"feed_emoji":"⏳","tokens_out":7643,"duration_ms":78240,"temperature":0.7,"pith_summary":"The paper argues that retrieval-augmented generation fails when it treats old and new facts as equally valid, letting stale but plausible evidence corrupt answers. It proposes Chronofy, a three-layer architecture that makes a fact's age structurally inescapable: a dedicated temporal subspace in the embedding, exponential decay weights during graph retrieval, and a Signal Temporal Logic check on the freshness of the evidence actually used in each reasoning step. The central formal claims are that the optimal decay rate equals twice the mean-reversion rate of the underlying latent process, and that output confidence can never exceed the validity of the least-fresh evidence in the chain. If those claims hold, 'how old is this fact?' stops being a heuristic and becomes a bound, and the system can demand new data rather than answer from stale context. The experiments report retrieval gains of 9.4% to 48.9% over static retrieval across benchmark knowledge graphs, with lower but real gains in end-to-end QA.","feed_headline":"Old facts bound new answers: decay-weighted RAG wins by 48.9%","feed_subtitle":"A three-layer system puts fact age into retrieval and logic, and asks for fresh data when evidence is too stale to trust.","key_machinery":"The load-bearing object is the exponential decay weight w = q · c · exp(-β Δt) used in retrieval, and its reflection in the reasoning layer as the validity signal v(s_i) = min_{e in facts(s_i)} q_e exp(-β_i (T_q - t_e)), which Signal Temporal Logic aggregates into a robustness score ρ = min_i (v(s_i) - γ). Two identities carry the argument: Proposition 1's β = 2κ, which connects the decay coefficient to the mean-reversion rate of an Ornstein-Uhlenbeck process under squared-error loss, and Theorem 1's weakest-link bound, which follows from possibilistic logic's min-aggregation for conjunctions. Together they make freshness thresholdable, comparable across reasoning steps, and formally tied to","core_discovery":"The paper's central discovery is that temporal validity can be treated as a first-class quantity in retrieval-augmented generation rather than a post-hoc filter. Every fact carries a timestamp, and the system's confidence in a conclusion is clamped by the weakest (most decayed) piece of evidence used to derive it: Theorem 1 gives C_out ≤ min_i q_i exp(-β_i (T_q - t_i)), where q_i is source reliability and β_i the decay rate of the fact's type. The paper grounds β_i in decision theory: if the latent state behind a fact type follows an Ornstein-Uhlenbeck process with mean-reversion rate κ, the value of information decays at rate 2κ, so β_i = 2κ is the optimal exponential coefficient. This sing","pith_inferences":["Editorial extension: the architecture does not actually depend on exponential decay—the STL gate only needs a validity signal—so replacing exp(-β Δt) with a Weibull or hazard-function model, which the paper names as an alternative, would preserve the weakest-link and re-acquisition machinery while fixing non-monotonic lifecycles.","Editorial extension: Theorem 1 is a general property of chaining premises with decaying necessity, so the bound should transfer to multi-hop QA, tool-use agents, and any decision pipeline where evidence age varies across sources.","Testable check on the paper's own identity: fit an Ornstein-Uhlenbeck process to historical fact-validity data for a relation type and compare the fitted κ against the learned β; the proposition predicts β = 2κ, and systematic deviations would localize where the Gaussian/exponential approximation breaks.","Flagged limitation from the paper: the end-to-end ablation uses oracle (gold) timestamps, and the authors report that automatic temporal-focus parsing covers only 37.8% of cases, making extraction the practical bottleneck; the STL robustness AUROC also trailed raw LLM confidence on the closed-context QA subset (0.479 vs 0.556), though it caught 24.6% of confidently-wrong answers."],"forward_implications":["Explicit age discounting beats both static retrieval and naive recency: the paper reports MRR gains of 9.4% on ICEWS14 and 48.9% on GDELT, while recency-only runs degrade performance in every experiment.","The weakest-link bound implies that a single stale fact in a chain caps the reliability of the whole answer—adding fresh evidence cannot compensate for the oldest component.","When the STL robustness ρ falls below the threshold, the system is instructed to re-acquire data instead of emitting a prediction, giving RAG a principled abstention policy.","Per-type decay rates are learnable and heterogeneous: optimal β ranges from 0 for stable diagnoses to 10.0 for fast-moving political events, consistent with the OU-process grounding.","The Temporal Invariance Guarantee means stable facts are preserved indefinitely rather than penalized by age, matching clinical intuition (β* = 0 for diagnoses, β* = 0.05 for lab results)."],"fun_headline_variants":["Decay-weighted RAG cuts hallucination by 48.9%","Most decayed evidence clamps RAG answer confidence","Time-aware RAG: fact age bounds answer trust","Optimal decay for RAG derived from mean-reversion"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that a fact's value decays memorylessly as exp(-β·age) with a single per-type rate β, which the paper itself concedes cannot capture non-monotonic lifecycles or sudden regime shifts and which Proposition 1 ties to stationary Gaussian dynamics.","fun_headline_variants_meta":{"raw":{"variants":["Decay-weighted RAG cuts hallucination by 48.9%","Most decayed evidence clamps RAG answer confidence","Time-aware RAG: fact age bounds answer trust","Optimal decay for RAG derived from mean-reversion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001051,"raw_usage":{"total_tokens":4299,"prompt_tokens":839,"completion_tokens":3460,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":583,"completion_tokens_details":{"reasoning_tokens":3392}},"tokens_in":583,"tokens_out":3460,"duration_ms":32217,"temperature":1.0,"reasoning_tokens":3392,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T20:36:26.795634+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect timestamped facts with human-annotated usefulness over time and check whether log-usefulness falls linearly with age; if validity flattens, spikes after re-publication, or follows a Weibull shape, Eq. (2)/(3) is violated and both the retrieval weights and the STL gate inherit systematic error.","supporting_citations":[],"review_version":1}