Automated fact-checkers usually treat a claim as simply true or false. Many claims, though, are true only during a certain period, and quantitative claims in experimental physics change over time in a particular way. Some numbers get refined: a measured neutron-skin thickness narrows its error bars but stays correct. Some numbers get overtaken: a record for how much beam power a target survives is broken by a later result. Some numbers start as plans and later become achieved results: a proposed 600 kW beam is not a demonstrated 600 kW beam. A verifier that ignores these patterns will either call a correct 2012 measurement "refuted" because a later paper reports a different central value, or accept a design target as if it were an operating record. We propose a verifier in which each claim carries a time anchor (the date the claim refers to). The system first classifies how that kind of claim ages, and then uses that classification during evidence retrieval, before any verdict is reached. Its output is a verdict tied to a time interval, not a single true/false label.
The test domain is nuclear-structure and accelerator-target claims, chosen because the file documents real cases of each aging pattern. The neutron skin of lead-208 is a measured quantity that later electroweak and hadronic measurements constrain further. Target power-handling figures such as the 20 kW/cm² rotating-target capacity are capability records. Facility proposals such as IsoDAR state design parameters. The method itself is general: we test it on the public open-web benchmarks AVeriTeC and MultiFC to see whether it improves the evidence retrieval step, which is the known weak point of scalable fact-checking pipelines.
Core Claim
Temporal information should decide which evidence is retrieved, and how strongly time is weighted should depend on the kind of claim. The main hypothesis is that adding claim-type-dependent temporal weighting to retrieval will improve evidence recall and verdict accuracy. Adding dates to the verifier's input after retrieval, which is what current systems do when they use time at all, should not produce the same gain. A second hypothesis applies to scientific numbers. A verdict that compares stated uncertainties will correctly mark older, wider measurements as "consistent but superseded in precision." A plain textual-entailment check (asking whether the evidence text logically supports the claim) would mark them as refuted.
Relation to Prior Work
Three papers motivate the gap. Implicit Temporal Reasoning for Evidence-Based Fact-Checking groups the time gap between claim and evidence, and the dates mentioned inside the text, into learned "time buckets." On MultiFC this raises a BiLSTM verifier from 32.39% to 47.43% macro-F1. However, the time features are used only in the final classifier, and the authors name retrieval-stage temporal reasoning as future work. That missing step is what we build. HerO at AVeriTeC is a strong open-weight pipeline. It combines HyDE-FC (an LLM writes a hypothetical fact-check article that serves as the search query) with BM25 keyword search and dense re-ranking. It reaches a 0.57 overall shared-task score, but its retrieved-answer quality is only 0.2801, and it uses no time information during retrieval. AVeriTeC supplies the temporally split benchmark. It also reports that web search returned useful evidence in only 9 of 20 audited cases, which shows that retrieval is the bottleneck.
Our core weighting method comes from HALO. HALO borrows radioactive half-life to down-weight outdated facts in temporal knowledge graphs, which store facts as (subject, relation, object, time) tuples. It learns half-lives of roughly 80 days on ICEWS14 and roughly 160 days on ICEWS18, and raises RE-GCN MRR from 37.67% to 41.38% on ICEWS14. HALO also limits our design in two ways. It works only on structured tuples, and it uses a fixed pruning threshold that, as its authors note, can wrongly delete facts that are old but still true. We replace the hard prune with a soft re-ranking weight. We also add a claim type, "refinement," for which age alone should not reduce validity.
CiteGuard and Scientific Fact-Checking: A Survey show that scientific claims can be verified automatically. Both treat claims as static, however, and neither handles measurement uncertainty.
The physics papers provide the domain, not the method. The PREX measurement reports a lead-208 neutron skin (the thickness of the outer layer of excess neutrons) of $R_n - R_p = 0.33^{+0.16}_{-0.18}$ fm. Its precision was limited by how much beam its diamond-cooled lead target could take. Later work on cesium and iodine neutron radii comes from coherent neutrino-nucleus scattering (Average CsI Neutron Density Distribution from COHERENT Data) and storage-ring proton scattering (Neutron radius determination of 133Cs). Together these form a chain of measurements that refine one another. The power-handling figures form a second, separate chain: JUNA reports 20 kW/cm² on rotating targets, SPES reports 10 kW on a 2000°C multi-disk target, and the IsoDAR proposal states a 600 kW design. These chains matter because they mix all three aging patterns within a small, checkable body of literature.
Research Program
The system has three parts.
The first part is a claim-type tagger. It labels each claim as stable, refinement, record/capability, or plan-vs-achieved. A stable claim does not age. A refinement claim is a measurement whose later values tighten rather than overturn it. A record claim states a best-so-far figure that can be surpassed. A plan-vs-achieved claim depends on whether something was designed or actually demonstrated. The tagger also extracts the claim's time anchor. It is a small fine-tuned classifier, so it runs at negligible cost compared with LLM calls.
The second part is time-weighted retrieval. Candidates come from the same hybrid BM25 + dense + HyDE-FC pool that HerO uses. Each candidate document $e$ for claim $c$ is then re-scored:
$$s(c,e) = r(c,e)\cdot \mathbb{1}[t_e \le t_c]\cdot \exp\!\left(-\ln 2\,\frac{t_c - t_e}{\tau_{k}}\right)$$
Here $r$ is the original relevance score, $t_c$ is the claim's time anchor, and $t_e$ is the evidence date. The indicator term discards any evidence dated after the claim, which removes the post-claim leakage that AVeriTeC warns about. $\tau_k$ is a half-life learned separately for each claim type $k$. For refinement claims, $\tau_k$ is fixed at infinity, and the ordering instead favors the most precise earlier measurement. Evidence dates come from document metadata, and from extracted in-text dates when metadata is missing. Missing or truncated dates are a known problem in MultiFC snippets, so we model them as a separate "unknown date" case rather than guessing a value.
The third part is an interval verdict. The verifier outputs the period over which the claim held: as of the claim's own date, and as of today. For numerical claims, a deterministic step parses the stated values and uncertainties and checks whether they agree within those uncertainties. This avoids asking an LLM to judge numerical agreement.
Baselines
The reference point that ignores time is HerO's retrieval and verifier, run unchanged on the same candidate pool. On AVeriTeC it reports an overall score of 0.57 and an answer-quality score of 0.2801 (HerO). We rerun it with the same Llama-3.1 checkpoints and the same two-H100 budget, which HerO reports as 500 claims in 6.6 hours. Beating this baseline means a higher answer-quality score and a higher overall score on the 2021 test split, with throughput within 10% of HerO's.
The strongest existing temporal method is Allein et al.'s two-level time-bucket verifier (Implicit Temporal Reasoning). It reports 47.43% macro-F1 on MultiFC. The best MultiFC system overall, which adds metadata, reports 49.2% macro-F1 (MultiFC). We reimplement the bucket features on top of the same backbone and evidence set as our system, so that the only difference is where time enters the pipeline. Beating it means exceeding 49.2% macro-F1 on MultiFC. Matching it while also improving retrieval recall would count only as a partial result.
The ablation that isolates the key ingredient keeps our full pipeline but sets a single half-life shared by all claim types. A HALO-style fixed hard threshold is a second ablation variant. Without the claim-type distinction, our method reduces to "HALO on text." If this ablation matches the full system, the claim-type idea has failed.
Why not just…?
Why not just restrict search to documents published before the claim date? A hard date cutoff removes leakage, but it treats a five-year-old background document the same as last week's update. The design already includes this cutoff as the indicator term, and we run it alone as an ablation. If the cutoff alone captures most of the gain, the contribution shrinks to leakage control. The experiments report this case honestly.
Why not just put the publication dates into the LLM prompt and let it reason? This is the most serious threat, because modern verifiers can read dates. It cannot help, though, when the right-dated document was never retrieved, and AVeriTeC's 9-in-20 audit suggests that happens often. We test this directly with a HerO-plus-dates-in-prompt arm. We also measure retrieval recall@k against gold evidence for each arm. If dates in the prompt match our verdict accuracy and recall, we drop the retrieval-stage claim and keep only the interval-verdict and uncertainty contributions.
Why not just use Allein-style buckets, which already gave a 15-point jump? The buckets act on whatever evidence was already retrieved, and they cannot tell "older and more precise" apart from "outdated." The refinement claims in our physics set are designed to expose exactly this failure. Using the same backbone makes the comparison fair.
Proposed Experiments
We run four experiments.
First, on AVeriTeC's temporal split, we compare the HerO reference, HerO with dates in the prompt, the bucket verifier, the shared-half-life ablation, and the full system. Measures are answer-quality score, overall score, gold-evidence recall@10, and the rate of retrieved post-claim documents.
Second, on MultiFC, we run the same comparison using macro- and micro-F1. Separate results are reported for snippets that have dates and snippets that do not.
Third, we build a small expert-annotated claim set from the nuclear and target literature in this file and from papers it cites. Each claim receives a time anchor, an aging type, and an interval label. Annotators are two physicists, and we report inter-annotator agreement next to AVeriTeC's κ = 0.619. No result on Lacuna was found for any temporal verifier on scientific numerical claims, so this set establishes the first numbers. Its decisive measure is the rate at which a refinement claim is wrongly marked refuted. One example is the PREX 0.33 fm value judged against later cesium radii, which concern a different nucleus and should be recognized as not contradicting it. Another example is a plan claim, such as IsoDAR's 600 kW, being wrongly marked as achieved.
Fourth, we report learned half-lives for each claim type. If the half-life learned for record claims is no shorter than the one learned for stable claims, the claim-type premise is falsified.
References
- Implicit Temporal Reasoning for Evidence-Based Fact-Checking
- HerO at AVeriTeC: The Herd of Open Large Language Models for Verifying Real-World Claims
- AVeriTeC: A Dataset for Real-world Claim Verification with Evidence from the Web
- HALO: Half Life-Based Outdated Fact Filtering in Temporal Knowledge Graphs
- CiteGuard: Retrieval-Augmented Citation Verification for LLM-Powered Peer Review
- Scientific Fact-Checking: A Survey of Resources and Approaches
- Measurement of the Neutron Radius of<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mmultiscripts><mml:mi>Pb</mml:mi><mml:mprescripts/><mml:none/><mml:mn>208</mml:mn></mml:mmultiscripts></mml:math>through Parity Violation in Electron Scattering
- Average CsI Neutron Density Distribution from COHERENT Data
- Neutron radius determination of 133Cs and its impact on the interpretation of CEνNS-CsI measurement
- Progress of Jinping Underground Nuclear Astrophysics Experiment (JUNA) project
- Future Perspectives of the Legnaro National Laboratories: The SPES project
- Proposal for an Electron Antineutrino Disappearance Search Using High-Rate<mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mmultiscripts><mml:mi>Li</mml:mi><mml:mprescripts/><mml:none/><mml:mn>8</mml:mn></mml:mmultiscripts></mml:math>Production and Decay
- MultiFC: A Real-World Multi-Domain Dataset for Evidence-Based Fact Checking of Claims