{"id":"0885835d-17c5-4e05-b7a4-0b0c5ffcb2bf","arxiv_id":"2507.03027","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"The authors build and release an open-source toolkit that turns Dutch population registry data into millions of plain-text personal life histories for LLM-based analysis.","lead":"This paper introduces a procedure, BOLT, that converts large-scale administrative log data into plain-text 'books of life' describing each person's trajectory across work, household, and education. It reports generating over 100 million such books for 4 million people from Dutch registry data and releases the toolkit open source.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Deferred effectiveness is the load-bearing gap: the paper's motivating claim that ~1,000-token, recipe-filtered books of life support LLM analysis is left to a companion paper, and no information-retention check separates summarization loss from LLM capability.","rationale":"The reader's conditional verdict is well supported. The construction pipeline is concrete and reproducible in principle: BOLT is open source, the nine recipe variants are specified, and the throughput numbers in Table 2 demonstrate a plausible path to 100 million books on the described hardware. The paper also does the right thing by stating its limitations early: Section 1 flags limited effectiveness evidence, and Section 5 explicitly leaves the ultimate value open. My stress-test agrees with the reader's weakest assumption but sharpens it: the specific mechanism that could falsify the central claim is information loss from token-limiting and filtering, not LLM competence in general. A failure of the companion prediction task would be ambiguous unless the representation's information retention is measured separately. The concrete test I propose separates those two explanations. Because the concern is about missing evidence rather than a demonstrated internal flaw, it does not change the verdict: the verdict stays CONDITIONAL, with the condition being the companion paper plus an explicit information-retention check.","tokens_in":15076,"tokens_out":8304,"duration_ms":94724,"concrete_test":"Run an information-retention audit inside the CBS Remote Access environment on a random subsample (e.g., 10,000 individuals): generate a full unconstrained book and the exact recipe- and token-limited book used for the PreFer task. Build a fixed set of closed-form factual queries whose correct answers are unambiguously derivable from the raw logs (number of household moves, spells with a new cohabitant, employer changes, months with high sick leave). Ask one fixed LLM, with identical prompts, to answer from the full book, the token-limited book, and the original spells representation. If accuracy on the token-limited books drops materially below the full-book or structured-data baseline, the 1,000-token filtering is a bottleneck for downstream use, and the paper's richness claim needs qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that BOLT turns registry-scale complex log data into plain-text books that combine richness and scale, with LLM analysis as the payoff. For that claim to hold, two things must be true: the text retains the information needed for downstream research questions, and pretrained LLMs can extract patterns from it. The paper provides no evidence for either. Section 3 states that books were limited to approximately 1,000 tokens for computational reasons; Section 4 shows recipes used to meet that limit are aggressively lossy—Book 4 keeps only employment changes, Book 5 only education changes, Book 8 only the five most recent household and address spells, and Book 9 filters household spells to the final five. The paper explicitly defers effectiveness to Satish et al. (2025) and in Section 5 calls the ultimate value something that remains to be explored. This is not an internal inconsistency: the lossless claim in Section 3 refers to a single household-log snippet, and the what/who/how recipe design makes filtering a deliberate user choice. But it means the headline capability is an untested premise. If a downstream LLM performs poorly, the cause could be the lossy 1,000-token representation rather than a general LLM limitation, so the motivating hypothesis is not yet distinguishable from its null. The paper is honest about this and should be read as a proof of concept, but the central richness-and-scale claim is conditional on a check that has not been reported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 'book of life' approach: converting complex log data (e.g., national registry records) into plain-text narratives of individual life trajectories, supported by an open-source toolkit called BOLT. The authors report generating over 100 million approximately 3-page books for 4 million people from Dutch registry data, demonstrating computational feasibility on a standard server. They describe a recipe-based design with 'what', 'who', and 'how' specifications, illustrate the pipeline with nine example book types, and offer a worked example of a two-page book. The motivating goal is to enable LLM-based analysis of life course data at scale, but the paper explicitly defers evidence of downstream effectiveness to a companion paper (Satish et al., 2025).","tokens_in":15313,"tokens_out":4971,"duration_ms":57136,"significance":"If the approach holds up, it offers a reusable, scalable transformation pipeline that lets researchers repurpose administrative and other complex log data for LLM-based analysis, potentially bridging qualitative richness and quantitative scale in life course research. The paper's strengths are its open-source release of BOLT, transparent recipe-driven design, honesty about limitations, and a concrete feasibility demonstration on a large, real-world registry. However, the significance is conditional: the paper currently demonstrates scale and tooling but does not empirically establish that the generated books retain task-relevant information or that LLMs can extract useful patterns from them. Those gaps are load-bearing for the paper's central promise.","major_comments":[{"comment":"Section 3 states that the visual, spells, and textual representations of the household-log snippet are losslessly interchangeable, but the actual books generated under PreFer are deliberately lossy: Section 3 limits books to approximately 1,000 tokens, and Table 2 shows that Books 4, 5, 8, and 9 keep only changes or the five most recent spells. This creates a mismatch between the paper's framing of 'richness' and the delivered artifact. The paper should either provide evidence that task-relevant information survives the filtering and truncation (e.g., a retrieval or prediction probe) or explicitly reframe the claim to say that the current PreFer books are lossy summaries, with the lossless property applying only to the per-event transformation. This is central to the paper's 'richness and scale' claim.","section":"Section 3 and Table 2"},{"comment":"The motivating payoff—that LLMs can analyze these books of life—is entirely deferred to a companion paper. Section 5 says 'the ultimate value—and limitations—of the books of life approach remain to be explored,' and Section 3 motivates the 1,000-token limit by downstream computational constraints. Without any empirical check, the reader cannot separate information loss in the representation from a general limitation of LLMs on this text format. I recommend adding at least a small validation experiment, such as querying an LLM for a known life event (e.g., a move or employment change) from a subset of the generated books and measuring accuracy, or comparing performance on a known outcome using a book-of-life representation versus the full tabular starter pack. Such a check would make the central hypothesis testable rather than asserted.","section":"Section 5 and Satish et al. (2025)"},{"comment":"The headline claim of 'over 100 million approximately 3-page books of life for 4 million people' is not auditable from the information given. Table 2 lists nine book types; generating all nine for every one of 4 million people would yield only 36 million books, not 100 million. The discrepancy needs clarification: what exactly counts as a 'book' (e.g., each recipe variant, each memory span, each output format, or each page), and how many distinct recipe configurations were used? Without a recipe manifest or a precise counting rule, the scale claim is difficult to verify. This is part of the paper's central feasibility argument and should be made transparent.","section":"Section 1 and Table 2"}],"minor_comments":[{"comment":"There is a typo in the sentence 'Some of some of these strengths and weaknesses are described in this paper': the duplicated 'some of' should be removed.","section":"Section 1"},{"comment":"The spelling of 'Waldek' in the section on prior work is inconsistent with 'Wladek' used elsewhere; the same spelling should be used throughout.","section":"Section 2"},{"comment":"In the 'Specifying the how' subsection, 'the norms for how information is written in books of like will evolve' should read 'books of life'.","section":"Section 4"},{"comment":"For the Persoontab row, the temporal resolution is listed as 'Monthly' while the content is described as information at birth; the table should clarify whether this file is updated monthly or contains time-invariant information.","section":"Table 1"},{"comment":"The example book in Figure 4 is described as two pages, while the paper states books are approximately three pages; the example would benefit from a note that two-page and three-page books are both within the intended range.","section":"Section 4, Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is honest and well-scoped as a proof of concept, and the open-source toolkit is a genuine contribution. The main issue is that the central 'richness' and 'LLM payoff' claims rest on deferred or unvalidated assumptions. A revision that adds a modest information-retention or LLM-probe experiment, or sharply narrows the paper's claims, would substantially strengthen it. The 100-million-books counting question should also be resolved before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this if you care about whether registry-scale administrative data can be turned into text that LLMs can analyze. The genuinely new thing here is BOLT: a configurable, open-source pipeline that serializes multi-source, multi-resolution log data into plain-text 'books of life,' with the what/who/how recipe abstraction letting users choose filters, ordering, and parsing dictionaries. That is a real and reusable contribution, and the scale claim (100M+ books, 4M people) is backed by the architecture and throughput numbers, even though the underlying CBS registry data can't be inspected.\n\nThe paper does several things well. It frames the representation problem clearly, situates the work against prior sequence models (Savcisens, Vafa) and text-serialization (Athey, Gardner), and it is admirably honest about being a proof of concept. Section 5 explicitly says the ultimate value remains to be explored, and the authors name the lack of effectiveness evidence upfront. The worked example in Figure 4 and the recipe design are concrete and helpful.\n\nThe soft spot is exactly where the stress-test lands: the motivating claim is that ~1,000-token, recipe-filtered books of life support LLM analysis, but no information-retention check separates summarization loss from LLM capability. Book 8 keeps only the five most recent household/address spells; Book 5 keeps only education changes. These are deliberate user choices, and the lossless claim in Section 3 refers only to a single household-log snippet, so there's no internal inconsistency. The gap is real but external: the companion paper (Satish et al. 2025) is supposed to supply the downstream evidence. Until that lands, the headline 'richness and scale' claim is conditional on an untested premise.\n\nThe citation pattern looks fair, and the paper does not oversell: no free parameters, no invented entities, no circularity burden. If I had to bet, the representation will be more useful for tasks where event order and rare transitions matter than for tasks where detailed spell-length distributions drive the answer, but that's speculation. The lack of public data is a genuine limitation for independent audit, but it is a feature of the registry setting, not a flaw in the authors' claims.\n\nWho is this for? Anyone building on administrative data for life-course or social prediction. It deserves a serious referee; the honest framing and the open-source artifact outweigh the deferred effectiveness. My recommendation: send it to review, and hold the authors to the companion paper for the effectiveness evidence.","headline":"Honest, well-scoped proof of concept for a plain-text life-course representation; the load-bearing effectiveness claim is explicitly deferred, so treat it as infrastructure, not as evidence that LLMs can analyze these books.","tokens_in":15895,"tokens_out":1473,"would_cite":true,"duration_ms":16020,"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":"This paper proposes a hybrid method that converts population-scale log data into short plain-text 'books of life' for each person, so large language models can analyze entire life trajectories.","keywords":["life course research","complex log data","book of life","large language models","population registry","data representation","plain text","BOLT toolkit"],"falsifier":"Run the fertility prediction task using LLMs on the generated books of life and compare against a strong tabular baseline trained on the same underlying logs; if the LLM-on-books approach does not match or exceed the baseline, the claim that the textual representation enables LLM-based analysis at scale loses its practical support. A more direct test is to convert the books back to spell-level records and check for exact round-trip equality, which would verify the asserted losslessness.","tokens_in":14853,"feed_emoji":"📖","tokens_out":6785,"duration_ms":67986,"temperature":0.7,"pith_summary":"For a century, life course researchers have had to choose between rich qualitative accounts of a few lives and quantitative analyses of many lives but little detail per person. This paper argues that two developments—administrative and other 'complex log data,' and large language models that read plain text—now make a hybrid possible. It proposes converting each person's logged life events into a short plain-text 'book of life,' and demonstrates feasibility by generating over 100 million roughly three-page books for 4 million people from Dutch population registry data. The paper's central claim is that these textual books preserve the information in the original logs while giving researchers a representation that LLMs can analyze at scale. The open-source Books of Life Toolkit (BOLT) is the concrete vehicle for making the approach reusable.","feed_headline":"100 million life stories, written from registry data","feed_subtitle":"A toolkit turns Dutch administrative logs into plain-text life narratives that language models can analyze.","key_machinery":"The load-bearing mechanism is the 'book of life' data representation together with the Books of Life Toolkit (BOLT) that generates it. A researcher writes a 'recipe' file making three specifications: the 'what' (which log files and which events become paragraphs in the book), the 'who' (which other people, such as household members, are included, possibly as 'books within books'), and the 'how' (filtering, ordering, and style, including parsing dictionaries that turn coded fields like 'gbageslacht: 1' into 'Sex at birth: Male'). The paper proves the concept by noting that a household spell expressed as a row, a diagram, or a sentence carries the same information, so the textual book is a lossless rendering of the structured logs rather than a lossy summary.","core_discovery":"The central claim is that complex log data—records logged for non-research purposes, spanning multiple life domains and temporal resolutions—can be transformed, losslessly, into a single plain-text narrative of an individual's life, and that this can be done at population scale. The authors demonstrate this by writing over 100 million approximately 1,000-token 'books of life' for 4 million people from five Dutch registry files covering demographics, households, employment, education, and addresses. They show for the household log that the visual, spell-based, and textual representations contain exactly the same information, so one can move between them without loss. The contribution is not a finding about lives themselves but a reusable representation and toolkit intended to let life course research combine qualitative-style richness with quantitative-style scale.","pith_inferences":["My editorial inference: the approach's ultimate value hinges on the companion result that pre-trained LLMs can actually extract predictive or explanatory signal from these roughly 1,000-token narratives; this paper deliberately defers that evidence.","My editorial inference: the same BOLT pipeline could be adapted to other complex log data beyond population registries—such as digital platform logs or administrative health records—wherever the lossless textual rendering remains comprehensible.","A testable extension would be a systematic comparison of LLM predictions from books of life against tabular baseline models on the same prediction task, which would isolate the contribution of the textual representation itself.","I also infer that the governance concerns flagged for centralized registries will become sharper if the approach is applied to sensitive domains like health, since the books make individual trajectories human-readable."],"forward_implications":["Researchers can use BOLT to generate books of life from registry or other complex log data in formats tailored to their own research questions, without needing the engineering effort required for this first implementation.","Because the textual books are lossless renderings of the underlying spells, any pattern an LLM finds in the text corresponds to a pattern in the original log data, not an artifact of aggregation.","The 'books within books' mechanism lets researchers embed linked lives—household members, co-workers, neighbors—into a focal person's narrative, making social context part of the representation.","If the approach scales as demonstrated, life course studies can move beyond the 'rectangle of numbers' and analyze rich, multi-domain trajectories for entire populations.","Longer and richer books are straightforward to produce with more computational resources, so current length limits are a practical constraint rather than a conceptual one."],"supporting_citations":[{"why":"Companion paper that tests the downstream use case of predicting fertility from books of life, supplying the evidence this paper defers.","marker":"Satish et al. (2025)"},{"why":"Describes the fertility prediction data challenge and the Dutch registry sandbox that defines the data and constraints for the proof of concept.","marker":"Sivak et al. (2024)"},{"why":"Documents the system of Dutch register-based social statistical datasets that provides the complex log data used here.","marker":"Bakker et al. (2014)"},{"why":"Origin of the 'book of life' vision that this paper returns to and operationalizes.","marker":"Dunn (1946)"},{"why":"Provides the concept of repurposed, non-research log data that motivates the approach.","marker":"Salganik (2017)"},{"why":"Earlier work writing career histories as plain text, a direct predecessor for the textual representation.","marker":"Athey et al. (2024)"},{"why":"Uses administrative data to construct life sequences, a non-textual predecessor the paper extends to multi-domain text.","marker":"Savcisens et al. (2024)"},{"why":"Represents career trajectories as sequences in a bespoke language, another predecessor the paper's text representation builds on.","marker":"Vafa et al. (2022)"}],"fun_headline_variants":["100M life stories from Dutch administrative logs","BOLT writes 100M life narratives from registry data","Book of life: hybrid method for life course research at scale","Transform log data into 100M text-based life trajectories","From complex logs to 100M plain-text life books"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a roughly 1,000-token plain-text book preserves enough of a person's life for downstream research questions, and that pre-trained LLMs can learn patterns from these narratives; the paper states this as a belief and leaves performance evidence to a companion paper.","fun_headline_variants_meta":{"raw":{"variants":["100M life stories from Dutch administrative logs","BOLT writes 100M life narratives from registry data","Book of life: hybrid method for life course research at scale","Transform log data into 100M text-based life trajectories","From complex logs to 100M plain-text life books"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000479,"raw_usage":{"total_tokens":2361,"prompt_tokens":922,"completion_tokens":1439,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":538,"completion_tokens_details":{"reasoning_tokens":1359}},"tokens_in":538,"tokens_out":1439,"duration_ms":12629,"temperature":1.0,"reasoning_tokens":1359,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:36:21.879582+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the fertility prediction task using LLMs on the generated books of life and compare against a strong tabular baseline trained on the same underlying logs; if the LLM-on-books approach does not match or exceed the baseline, the claim that the textual representation enables LLM-based analysis at scale loses its practical support. A more direct test is to convert the books back to spell-level records and check for exact round-trip equality, which would verify the asserted losslessness.","supporting_citations":[],"review_version":1}