{"id":"b6c4cd3c-3c3a-4ca6-b9b8-0686c41832f3","arxiv_id":"2602.22045","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DLT-Corpus is a 2.98B-token scientific/patent/Twitter corpus for blockchain NLP, plus LedgerBERT (+23% NER vs BERT), a sentiment dataset, and cross-domain innovation-diffusion analyses.","lead":"This paper introduces DLT-Corpus, a 2.98-billion-token text collection spanning scientific papers, US patents, and 22 million Twitter posts about blockchain and distributed ledgers. It also ships LedgerBERT, a domain-adapted language model with modest NER gains, and claims that DLT research tends to appear in scientific literature before reaching patents and social media.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"First-mention diffusion order may be an artifact of divergent source coverage and filtering, not evidence of science-led innovation.","rationale":"The reader's verdict is CONDITIONAL, identifying corpus representativeness and first-mention order as the key weakness. My stress test agrees and sharpens that concern: the specific mechanism by which source construction can bias the diffusion finding is the undocumented temporal coverage of the Twitter/X aggregation and the narrow patent keyword filter. The paper's corpus is likely a useful, large, and carefully documented resource; the LedgerBERT and sentiment datasets also appear plausible. However, the innovation-diffusion analysis—a headline claim repeated in the abstract and conclusions—depends on first-mention dates that are not directly reported and are vulnerable to systematic left-censoring in social media and patents. This does not invalidate the corpus contribution, but it does mean the 'science precedes patents and social media' claim should be treated as provisional pending a sensitivity analysis. Since the reader already set a conditional verdict and my critique reinforces that assessment rather than overturning it, UNCHANGED is the appropriate recommendation. The proposed concrete test—checking earliest mention dates and excluding late-starting tweet datasets—would settle whether the concern lands. If the ordering survives, the paper's central claim is substantially strengthened.","tokens_in":175,"tokens_out":3104,"duration_ms":43211,"concrete_test":"For each technology (stablecoin, DEX, AMM) and each corpus subset, extract the earliest document's publication/post timestamp. Then re-run the diffusion analysis restricted to documents from sources with verified coverage from their reported start year; for social media, determine each Kaggle/academic dataset's actual date range and exclude any dataset that begins after 2015. Also re-run patent retrieval using expanded keywords (e.g., 'stablecoin', 'automated market maker', 'decentralized exchange', 'cryptocurrency') or CPC codes. If the scientific-literature first-mention date remains earlier than social media after excluding late-starting tweet datasets and expanded patent search, the ordering is robust; if it shifts, the 'science first' claim is an artifact.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central finding—that technologies 'originate in scientific literature before reaching patents and social media'—is not established by the reported evidence because the three subsets have unequal temporal coverage and collection filters that can create artificial lags. Social media data are aggregated from Kaggle and academic datasets (§4.1.3) whose earliest coverage is not documented; if those datasets begin later than 2013 or contain early-period gaps, first social-media mentions are biased late. Patents were selected solely by keyword search for 'Distributed Ledger Technology' and 'blockchain' (§4.1.2), potentially missing early patents using alternative terminology (e.g., 'cryptocurrency', 'digital asset', 'smart contract'), biasing patent first-mentions late. Scientific literature was filtered by a BERT model trained on the authors' own DLT NER annotations [36], which may select for post-2008 vocabulary and miss pre-2008 papers describing the same concepts without those terms. Figures 3 and 4 show only normalized yearly proportions, not explicit first-mention dates per technology or per document. Without controlling for source-specific coverage start dates and keyword coverage, the observed ordering is confounded with collection artifacts. The paper's own limitations section acknowledges possible marginal relevance errors but does not address source temporal completeness.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DLT-Corpus, a large multi-source text collection for the Distributed Ledger Technology domain, containing 2.98 billion tokens from 37,440 scientific publications, 49,023 USPTO patents, and 22.03 million Twitter posts. The authors describe the construction and quality assessment of the corpus, train a domain-adapted language model (LedgerBERT), release a crowdsourced sentiment-analysis dataset, and present two utility analyses: (i) correlations between document volumes, market capitalization, and sentiment; and (ii) technology diffusion across scientific literature, patents, and social media. The headline claims are that technologies originate in scientific literature before reaching patents and social media, and that scientific publications lead market expansion by two years (ρ=0.95, p<0.001). The resource is publicly released with code, models, and datasheet-style documentation.","tokens_in":20916,"tokens_out":3274,"duration_ms":34039,"significance":"If the claims hold, DLT-Corpus is a substantial and valuable contribution: it is described as the largest domain-specific DLT text collection, it is publicly released, and it ships reproducible artifacts (code, models, datasets) as well as a FAIR-aligned datasheet. The quality assessment against general-purpose corpora is a useful benchmark. The temporal and cross-source analyses are of interest to innovation-diffusion and bibliometric communities. However, the two central analytical claims — technology-transfer ordering and science-leading-market causality — are currently supported only by correlations and normalized mention proportions, and the evaluation of LedgerBERT is partly in-domain by construction. The resource itself is likely useful even if those analytical claims are weakened; the paper's contribution should be reframed accordingly.","major_comments":[{"comment":"The central claim that technologies 'originate in scientific literature before reaching patents and social media' is not established by the reported evidence. Figures 3 and 4 show normalized yearly proportions, not first-mention dates per technology or per document, and no statistical test for ordering is provided. More importantly, the three subsets have different temporal coverage (scientific: 1978–2025; patents: 1990–2025; social media: 2013–mid 2023; §4.1.3) and different collection filters. The patents subset uses only the search terms 'Distributed Ledger Technology' and 'blockchain' (§4.1.2), which can miss early patents using terms such as 'cryptocurrency', 'digital asset', or 'smart contract', biasing patent first-mentions late. The social media subsets are aggregated from Kaggle and academic datasets whose earliest coverage and sampling gaps are not documented; if those datasets","section":"§6.1, Figs. 3–4"},{"comment":"The abstract states that 'scientific publications lead market expansion by two years (ρ=0.95, p<0.001)', but Table 5 shows that social media also exhibits ρ=0.95 at lag −3, and patents exhibit ρ=0.95 at lag −2 and ρ=0.93 at lag −3. The claim that scientific literature uniquely or particularly leads is not supported by any test for differences among lags or among document types. Moreover, the correlations are computed on annual data from 2013–2024 (n=12 for scientific literature and patents, n=11 for social media) with no correction for autocorrelation or multiple testing; Spearman's ρ on such short, highly autocorrelated series can be large and significant even when no true lead–lag relationship exists. The full text also reports concurrent ρ=0.76 for scientific literature (Table 4), so the 'two-year lead' claim rests on a single lagged correlation without a formal comparison or robustne","section":"Abstract; §6, Tables 4–5"},{"comment":"The primary evaluation of LedgerBERT is partly circular. The scientific-literature subset of DLT-Corpus was filtered by fine-tuning BERT-base-cased on the NER dataset from [36] and predicting domain entities (§4.1.1). LedgerBERT is then evaluated on exactly that same NER dataset from [36] as the 'primary evaluation' (§5.1). This means the model is trained on a corpus that was selected using labels from the very evaluation set, and the reported +23% over BERT-base and +3.5% over SciBERT may be inflated by construction. The authors note that the dataset 'derives from scientific literature, matching our corpus composition', but they do not discuss the selection overlap. An independent held-out evaluation (e.g., NER on patents or social media, or a DLT NER dataset not used in corpus filtering) is needed to support the claim that LedgerBERT improves DLT-specific NER because of domain-adaptive","section":"§4.1.1 and §5.1"},{"comment":"The limitations section is incomplete regarding the two headline analyses. It states that 'marginally relevant DLT papers may remain in the dataset' but does not acknowledge that the science-first diffusion order and the market-lead correlations could be artifacts of (i) the patents subset's narrow keyword selection, (ii) undocumented temporal coverage or gaps in the social-media aggregations, or (iii) the scientific-filtering model's vocabulary biases. Because these threats bear directly on the paper's central claims, they should be addressed explicitly, preferably with supplementary analyses that restrict all three subsets to a common overlapping period and/or use alternative keyword vocabularies.","section":"§9 Limitations"}],"minor_comments":[{"comment":"The text refers to 'Fig. 2a' for stablecoins, 'Fig. 2b' for DEXs, and 'Fig. 2c' for AMMs, but the figures are numbered Fig. 3(a)–(c). Please correct the cross-references.","section":"§6.1 text"},{"comment":"The sentence 'Comparing Fig. 1d with Fig. 5' appears to refer to Fig. 1(d), the temporal evolution panel, but the connection between sentiment and document growth is asserted rather than quantified. Consider reporting a correlation between sentiment fractions and document volumes, or clearly labeling this as a visual comparison.","section":"§6.2"},{"comment":"The description of the sentiment-label construction would benefit from more detail: the 'median minimum votes' filter and the 25th/75th percentile boundaries are mentioned, but the exact thresholds and the number of examples excluded at each step are not reported. This is a reproducibility concern for the sentiment dataset.","section":"§4.2"},{"comment":"The social-media subset is described as 'Twitter/X posts', but several upstream sources are Kaggle datasets (e.g., 'bitcoin-tweets', 'crypto-tweets'). It would be useful to document which sources cover which time ranges, since this directly affects the diffusion analysis.","section":"Appendix A, Table 8"},{"comment":"Minor typographical issues: 'cryptocurrencies price prediction' in the abstract; 'T&Cs at collection time' should be 'terms and conditions'; the footnote on line 30 of §5.2 is incomplete. A careful proofread is recommended.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The resource contribution is real and the public release is commendable, but the analytical claims currently outrun the evidence. I would recommend the authors either substantially weaken the science-first and science-leads-market claims, or add robustness analyses that control for coverage differences and alternative keyword vocabularies. The circularity in the LedgerBERT evaluation should also be addressed. If these changes are made, the paper could be a solid publication; in its current form, the headline claims are not fully supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: DLT-Corpus is worth your time. The 2.98B-token, three-source resource is the real contribution, and the release looks careful and usable. The diffusion and market-lead claims are a side dish, not the main course, and they are seasoned a bit too strongly.\n\nWhat's new: no one has put open-access science, USPTO patents, and Twitter into one public DLT corpus at this scale. The metadata is rich, the legal reasoning is thoughtful (open-access only, pre-2023 Twitter), and they ship code, a sentiment dataset, and LedgerBERT. The keyword-density check against RefinedWeb and C4 is a reasonable way to show domain specificity. That kind of reproducible resource work deserves credit.\n\nThe soft spots. First, the analytics. The abstract's 'scientific publications lead market expansion by two years (ρ=0.95, p<0.001)' does match Table 5 (lag -2, scientific: 0.95, p<0.001) — so the reader's note that it appears only for social media is wrong. But the deeper problem is that these are annual time series with about a dozen observations, heavily autocorrelated, and no adjustment is made. So the 'lead' conclusion is statistically fragile. The 'virtuous cycle' language is storytelling.\n\nSecond, the diffusion analysis. Showing normalized mention proportions is not the same as establishing first-mention order. The corpus subsets have different temporal coverage, and the patent keyword filter is narrow. The stress-test worry about coverage artifacts is legitimate, even if the specific technologies chosen (stablecoins, DEX, AMM) are recent enough that some pre-2013 bias doesn't bite. I'd ask the authors to rephrase 'originate in scientific literature' as 'first appear in our science subset,' and to test sensitivity to query choices and coverage truncation.\n\nThird, the LedgerBERT NER evaluation is partly circular: the same annotation set [36] was used to train the domain filter and to evaluate the model. The +23% over BERT-base is real but inflated by construction. The out-of-domain sentiment result showing no degradation is more telling.\n\nOverall: the resource is solid; the analyses are not the reason to publish. This deserves a serious referee: the corpus should be vetted, the statistics redone, and the claims tamped down. I'd bring it to a reading group, and I'd cite it.","headline":"The corpus is a real resource worth having; the diffusion and market-lead analyses overreach, but that's fixable.","tokens_in":21322,"tokens_out":4660,"would_cite":true,"duration_ms":43735,"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":"A 2.98-billion-token corpus spanning science, patents, and social media shows that DLT concepts first appear in research before markets.","keywords":["distributed ledger technology","blockchain","text corpus","named entity recognition","sentiment analysis","innovation diffusion","patent analysis","domain-adapted language model"],"falsifier":"For any of the three analyzed technologies (stablecoins, DEXs, AMMs), find a patent or social media post dated earlier than the corpus's first scientific mention of that term; such a document would break the claimed science-first ordering.","tokens_in":20514,"feed_emoji":"📚","tokens_out":5011,"duration_ms":46099,"temperature":0.7,"pith_summary":"This paper introduces DLT-Corpus, a publicly released collection of 2.98 billion tokens from 22.12 million documents spanning scientific publications, US patents, and social media posts. It argues that this is the largest domain-specific text resource for distributed ledger technology and demonstrates its value by showing that technologies like stablecoins, decentralized exchanges, and automated market makers first appear in scientific literature, then in patents, then in social media. It also reports that scientific publication volume leads cryptocurrency market expansion by about two years, while social media sentiment stays overwhelmingly bullish even during market downturns. Additionally, a language model trained on this corpus improves named-entity recognition for DLT terms by 23% over a general baseline while preserving general sentiment performance.","feed_headline":"2.98B-token corpus shows DLT ideas start in science","feed_subtitle":"Largest DLT text collection spans 22M documents; papers lead market growth by two years.","key_machinery":"The central object is the corpus itself, which aggregates three timestamped document streams: scientific literature (37,440 open-access publications), US patents (49,023 records), and social media posts (22 million). The analysis relies on the corpus's temporal metadata to compute lagged correlations and first-mention dates, and on its keyword density (8.7 times higher than general web corpora) to argue it provides concentrated learning signal for domain-adapted models. The filtering pipeline, including a BERT-based domain relevance filter trained on an existing DLT NER dataset and manual pruning, defines the corpus's membership and thus drives the first-mention ordering.","core_discovery":"The central claim is that DLT-Corpus, the largest domain-specific text collection for distributed ledger technology, enables both specialized NLP and cross-domain studies of innovation. Using it, the authors find that three economically significant technologies (stablecoins, DEXs, AMMs) consistently appear in scientific literature before they appear in patents or social media, matching a traditional technology-transfer pattern. They also find an asymmetric lagged correlation: scientific publications lead market capitalization by two years (rho=0.95, p<0.001) and the correlation decays when the market leads, while patents show a symmetric correlation and social media follows the market. The c","pith_inferences":["The first-mention ordering (science before patents before social media) is sensitive to how each subset was collected and filtered; a differently filtered corpus could yield a different diffusion ordering, so the technology-transfer conclusion should be tested on alternative query and filtering choices.","The two-year lead of scientific publications over market cap is based on annual aggregates; finer-grained data could reveal shorter lags or lead-lag dynamics within quarters, and the correlation alone does not establish causation.","The persistently bullish social sentiment may reflect the cryptocurrency-focused composition of the social media sources and the sentiment labeling method rather than an objective market mood; a sample of general social media would be a useful check.","The corpus excludes news articles for copyright reasons; a news stream, if added, might show that financial journalism acts as an intermediary between research and market sentiment."],"forward_implications":["Tracking scientific literature can serve as an early-warning system for emerging DLT technologies, since concepts first appear there.","Scientific output appears to lead cryptocurrency market growth by about two years, suggesting research is a leading indicator for DLT market expansion.","Domain-adapted language models trained on DLT-Corpus outperform general models on DLT-specific NER, showing concentrated terminology matters.","The corpus enables integrated analysis across scientific, patent, and social discourse that previously required assembling separate, often inaccessible datasets.","Because social media coverage stops at 2023, the corpus is a fixed snapshot and cannot track post-2023 community discourse."],"fun_headline_variants":["DLT-Corpus: science leads patents and market by 2 years","Largest DLT corpus: research precedes market growth by 2 years","Science first: DLT ideas emerge before patents and social media","2.98B-token DLT corpus: science predicts market trends"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The technology-transfer ordering rests on first-mention dates derived from a corpus whose scientific subset was filtered by a BERT-based relevance model and whose patents were selected by a simple keyword search; different query or threshold choices could change the order.","fun_headline_variants_meta":{"raw":{"variants":["DLT-Corpus: science leads patents and market by 2 years","Largest DLT corpus: research precedes market growth by 2 years","Science first: DLT ideas emerge before patents and social media","2.98B-token DLT corpus: science predicts market trends"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000223,"raw_usage":{"total_tokens":1304,"prompt_tokens":767,"completion_tokens":537,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":511,"completion_tokens_details":{"reasoning_tokens":459}},"tokens_in":511,"tokens_out":537,"duration_ms":5511,"temperature":1.0,"reasoning_tokens":459,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T20:46:59.986391+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"For any of the three analyzed technologies (stablecoins, DEXs, AMMs), find a patent or social media post dated earlier than the corpus's first scientific mention of that term; such a document would break the claimed science-first ordering.","supporting_citations":[],"review_version":1}