{"id":"69fdbb94-fcbc-486e-b09b-bbdee9338815","arxiv_id":"2411.13134","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Comparing 16 extraction methods shows that discarding hierarchical relations and adding secondary street data best approximates spatial distance in a confront network of medieval Avignon.","lead":"The authors turn medieval land registries, which locate properties only by their neighbors and directions, into graphs and compare ways to build them. The selected graph partitions 14th century Avignon into neighborhoods that match known urban landmarks and practices.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 0.80 correlation is not an independent validation: the spatial yardstick was built from the same confronts that create the graph edges (§3.4, §6.1), so the main empirical claim is partly self-referential.","rationale":"I agree with the reader's weakest assumption. The paper is transparent about georeferencing uncertainty, which is good, but the evaluator is not independent. The concern is not that the authors are wrong; it is that the empirical support for the central claim does not rule out the alternative explanation that the high correlation comes from constructing the coordinates to match the graph. The invariant-only test is feasible with the published data and would settle whether the .80 is an artifact. I do not see a reason to move the verdict: CONDITIONAL remains the right call, because the method is novel, clearly described, and the historical analysis is suggestive, but the headline correlation needs independent validation before the claim is accepted.","tokens_in":30547,"tokens_out":6479,"duration_ms":73320,"concrete_test":"Recompute Spearman's ρ for EFS_k using only pairs of invariant objects (streets, edifices, gates, walls) whose coordinates come from planimetric/archaeological sources (§3.4) rather than from property confronts. If the restricted ρ is substantially below 0.80 (e.g., <0.5), the headline correlation is inflated by the self-referential yardstick; if it stays near 0.80, the circularity is a minor issue.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is that EFS_k approximates spatial distance with Spearman ρ=0.80 (Table 3). The yardstick for 'spatial distance' is not independent. Section 3.4 states that 2,049 properties were 'arranged manually based on additional information, particularly confronts,' and the paper concedes this localization is 'highly uncertain' and 'relies on the localization of all the invariants, which is often quite hypothetical.' The graph edges are then extracted from exactly these confront relations (§5). Computing Spearman correlation between graph distance and Euclidean distance on coordinates that were manually chosen to be consistent with the confronts therefore measures, at least in part, internal consistency of the georeferencing procedure rather than the graph's ability to recover real spatial layout. The method ranking in §6.6, including the choice of EFS_k over EFW_k, inherits this bias. In addition, k=7 for EFS_k was selected by maximizing distance correlation on the same data (Fig. 15), so the reported 0.80 is a tuned maximum with no uncertainty or hold-out estimate. The community discussion in §7 is historically plausible, but it does not provide an independent check on the distance-correlation claim. These issues make the central claim plausible but not established; a conditional verdict is appropriate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a graph-based approach to representing medieval urban space from land registries, where properties are located only by relative spatial references ('confronts'). Sixteen variants of a confront-network extraction pipeline are compared on a dataset of papal-period Avignon, using two criteria: coverage (number of properties retained in the graph) and reliability (Spearman correlation between graph distance and Euclidean spatial distance). The method EFS_k (extended data, flat relationships, split longest streets) is selected, achieving ρ_d = 0.80, and Louvain community detection on that graph yields 31 communities that the authors interpret as lived neighborhoods. The data and code are made publicly available.","tokens_in":30793,"tokens_out":4674,"duration_ms":45938,"significance":"If the quantitative evaluation were independent, the paper would make a useful methodological contribution for historical spatial analysis when absolute coordinates are unavailable. Its strengths include the open data and code, the systematic comparison of extraction choices, and a careful qualitative historical discussion of the resulting communities. However, the central empirical claim — that the confront network approximates spatial distance with ρ_d = 0.80 — is compromised by the fact that the spatial yardstick used for evaluation was constructed from the same confront relations that generate the graph edges, and by the lack of baseline comparisons and uncertainty quantification. The ranking among extraction variants may still be informative, but the absolute fidelity claim needs re-basing.","major_comments":[{"comment":"The evaluation yardstick for spatial distance is not independent of the graph construction. Section 3.4 states that the 2,049 georeferenced properties were 'arranged manually based on additional information, particularly confronts,' and that this localization is 'highly uncertain' because it relies on the localization of invariants 'which is often quite hypothetical.' The graph edges in Section 5 are extracted from exactly these confront relations. Consequently, the Spearman correlation ρ_d in Table 3 measures, at least in part, the internal consistency of the manual georeferencing procedure rather than the ability of the confront graph to recover an independently known spatial layout. The paper should either provide an independent validation set (e.g., coordinates of invariants derived from archaeological or planimetric sources that were not used in the confront extraction) or clearly reframe the claim as one of internal consistency rather than absolute spatial fidelity.","section":"§3.4, §6.1"},{"comment":"The value k = 7 for EFS_k is selected by maximizing the distance correlation on the same data that are then used to report ρ_d = 0.80 in Table 3. This is a post-selection maximum, and no confidence intervals, significance tests, or hold-out validation are provided. The comparison across methods in Section 6.6 therefore does not account for selection bias. The authors should provide uncertainty quantification (e.g., bootstrap confidence intervals for ρ_d) and a validation procedure that separates parameter selection from evaluation, or at minimum a sensitivity analysis over k showing that the reported advantage of EFS_k is robust.","section":"§6.4, Fig. 15"},{"comment":"No baseline against simpler spatial graph models is reported. Without comparators such as a k-nearest-neighbor graph in Euclidean space, a graph based on shared parish or street membership, or a random-edge null model, the claim that the confront network 'approximates spatial distance' is not calibrated. A baseline would show whether the observed ρ_d = 0.80 is meaningful relative to what can be produced with the same amount of information. The authors should add at least one such baseline and report its distance correlation and coverage in Table 3 or in a separate comparison.","section":"§6.6"},{"comment":"The community discussion in Section 7 is historically plausible and illustrates the qualitative value of the approach, but it is not an independent check on the quantitative distance-correlation claim. Since the communities are derived from the same confront graph that was evaluated circularly, the historical consistency reported in Section 7 does not correct the circularity identified above. The conclusion in Section 8 should not present the community analysis as confirmation of the quantitative fidelity of the graph; it should be framed strictly as a qualitative illustration.","section":"§7, §8"}],"minor_comments":[{"comment":"In the Full graph row, the proportion of properties is reported as 100.00%, but Section 3.2 states the database contains 3,021 properties, and the Full graph contains 2,693 property vertices; 2,693/3,021 ≈ 89.1%. Please reconcile the numbers or clarify the denominator used for the percentage.","section":"Table 3"},{"comment":"The text states that Kendall's τ and Spearman's ρ give 'qualitatively similar results' and that only Spearman results are shown, but no quantitative comparison is presented. A sentence with summary values or a supplementary figure would support this claim.","section":"§6.1"},{"comment":"The sentence 'we empirically determine that a lower threshold of 25 vertices is appropriate' does not specify the criterion used to choose this threshold; please state the metric or heuristic that led to 25.","section":"§5.1.2"},{"comment":"Typo: '162 our of 326 streets' should read '162 out of 326 streets.'","section":"§3.4"},{"comment":"There are minor language issues: 'felt under their lordship' should be 'fell under their lordship,' and the stray 'extsuperscript' in Section 7.2.2 appears to be a LaTeX rendering artifact.","section":"§2, §7.2.2"},{"comment":"The axis labels and legend text in Figures 14 and 15 are very small; increasing the font size would improve readability.","section":"Figures 14 and 15"}],"recommendation":"major_revision","confidential_remarks":"The paper is more applied than methodological, and would fit a venue concerned with digital humanities or applied network analysis. The authors are transparent about the uncertainty in §3.4, which is commendable, but the central quantitative claim needs an independent validation or a clear reframing. The most significant correctable issues are the circular yardstick, the absence of baselines, and the missing uncertainty quantification. If the authors provide an independent validation (even a small one) or substantially temper the absolute claim, the paper could become acceptable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid applied-network paper, not a paradigm shift. The genuinely new pieces are the systematic comparison of sixteen graph-extraction variants on a newly assembled geohistorical database for fourteenth-century Avignon, and the explicit framing of coverage and distance-correlation as conflicting selection criteria. The code and data are public, the historical sourcing is careful, and the authors are unusually candid about the uncertainty in their geolocation. That candor is real and should be credited.\n\nThe method story is clear: hierarchical relations hurt, secondary relations help, and splitting long streets helps. The community segmentation is plausible and historically informed. I would not be surprised if the ranking of variants survives stronger validation.\n\nThe soft spot is the one flagged in the stress-test note, and it is not manufactured. The Euclidean positions used as ground truth were partly produced by arranging properties to be consistent with the very confronts that generate the graph edges; Section 3.4 says this explicitly. So the Spearman rho = 0.80, and the gaps between variants like EFW_k and EFS_k, are partly measures of internal consistency in the georeferencing procedure, not independent evidence that the graph recovers real spatial layout. On top of that, the parameter k is selected on the same data by maximizing the same correlation (Figure 15), no confidence intervals or significance tests are reported, and there is no baseline against something simpler, like connecting properties by shared parish or street co-occurrence. These issues are fixable with hold-out evaluation or sensitivity analysis, but as published the headline claim is best read as “the best variant among those tested, under a yardstick that shares information with the graph.”\n\nA smaller point: the per-community distance correlations in Table 4 range from 0.03 to 0.78, so the graph’s spatial fidelity is uneven. The qualitative neighborhood reading would be stronger if the low-correlation communities were explicitly treated as less reliable.\n\nWho this is for: historians and digital-humanities researchers who want to segment medieval cities without reconstructing parcels, and network scientists interested in extraction choices from relational records. It deserves a serious referee. I would send it out and ask that the validation be made honest about independence: use positions that do not depend on confronts where available, add split-half or leave-one-out uncertainty, and compare against at least one trivial baseline.","headline":"A useful graph-extraction pipeline for incomplete medieval land records, honestly presented but with a partly self-referential validation—worth refereeing, though the 0.80 correlation should not be read as independent confirmation.","tokens_in":31285,"tokens_out":2121,"would_cite":true,"duration_ms":23016,"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 graph built from medieval 'next-door' mentions reproduces city distances well enough to split Avignon into 31 lived neighborhoods without a single exact address.","keywords":["confront networks","spatial graphs","medieval Avignon","land registries","community detection","distance correlation","historical GIS","graph extraction"],"falsifier":"Take the EFS_k extraction, restrict the evaluation to a control set of properties whose medieval positions are independently fixed by surviving archaeology or standing landmarks (churches, gates, walls), and recompute the Spearman correlation; a large drop from 0.80 would show the reported spatial fidelity was an artifact of the confront-based georeferencing loop.","tokens_in":30354,"feed_emoji":"🗺️","tokens_out":5842,"duration_ms":56902,"temperature":0.7,"pith_summary":"Medieval land registers rarely give addresses; they locate a property by naming its neighbors and surroundings—which house faces it, which street bounds it, what lies to its north. This paper argues that these relative mentions, organized into a graph, carry enough spatial information to approximate real distances and to split a city into coherent neighborhoods, without ever reconstructing the parcel plan. Using 14th-century registers from Avignon, the authors extract 16 graph variants and pick the one whose shortest-path distances best track measured spatial distances: a graph that keeps only flat (non-hierarchical) relationships, splits the longest streets into connected segments, and is enriched with secondary-source connections between streets and buildings. That graph reaches a Spearman correlation of 0.80 and, partitioned by Louvain community detection, yields 31 communities that line up with streets, markets, cemeteries, and parish boundaries in a way historians can interpret as lived neighborhoods. If the claim holds, historians can segment and analyze cities whose sources are partial and imprecise without first solving the usually impossible task of locating every parcel.","feed_headline":"Next-door mentions in medieval tax rolls map Avignon at 0.80","feed_subtitle":"A graph of who-borders-whom beats full data at matching real distance and rediscovers 31 lived neighborhoods.","key_machinery":"The load-bearing object is the confront network: an undirected spatial graph whose vertices are properties and urban invariants (streets, gates, churches, walls, river) and whose edges are the relative locations recorded in terriers, normalized to seven relation types. The distance-correlation criterion $\\rho_d$—Spearman's rank correlation between shortest-path (graph) distance and Euclidean (spatial) distance on the georeferenced subset—carries the method comparison; the paper's extraction variants are essentially attempts to make geodesic distance on the graph behave like Euclidean distance. The best variant, EFS_k, combines extended secondary data, flat relationships only, and splitting of the seven longest streets into artificially connected segments, which prevents long linear objects from acting as spatial shortcuts.","core_discovery":"The paper's central claim is that a confront network—a graph whose edges are the spatial relationships tenants themselves declared (x is north of y, x faces y, x is inside borough z)—is a legitimate stand-in for a map of the urban space, for the purpose of spatial segmentation. The authors define the quality of such a graph by two criteria: coverage (how many of the 3,021 recorded properties the graph retains) and reliability (Spearman rank correlation between graph distance and spatial distance over the 2,049 georeferenced properties). Comparing 16 extraction variants, they find that the best graph is not the one that uses all available information but the one that discards hierarchical containment edges, splits the longest streets, and adds relationships from secondary sources; this EFS_k variant reaches 0.80 correlation while keeping 69% of properties. Partitioning that graph with Louvain produces 31 communities with modularity 0.93, which the authors interpret as the lived neighborhoods of papal Avignon, grounded in spatial proximity and shared landmarks rather than administrative parish lines.","pith_inferences":["A transferable benchmark design: other cities with terrier-like registers, where cardinal relations may be rare, could reuse the same 'coverage plus distance correlation' selection protocol, but the optimal trade-off will likely shift with the relation types present in the source.","The 0.80 ceiling is probably not a limit of the method but of the noisy ground truth: the georeferenced properties were manually arranged using the same confronts that make the edges, so part of the measured correlation is circular; an independent control group of properties located via fixed landmarks would reveal the true floor.","The same graph could feed interpolation of missing absolute positions: instead of averaging neighbor coordinates, a graph neural network could use the edge semantics to predict locations, a perspective the authors mention; if that works, the confront network becomes a scaffold for full georeferencing, not just segmentation.","The finding that deleting information improves spatial fidelity is likely general: in any spatial graph, hubs representing broad regions collapse many real meters into two hops, so flat plus split representations are a safer default for distance-based analysis on relational sources."],"forward_implications":["Historians can segment an incompletely documented city without parcel-by-parcel georeferencing, as long as the registers contain explicit relative locations.","Discarding hierarchical containment edges and splitting long streets improves distance fidelity; keep-all-information graphs are spatially misleading, with the full graph reaching only 0.22 correlation versus 0.80 for EFS_k.","Secondary information about how streets connect to each other and to buildings is worth collecting: adding it raised correlation by 0.26 to 0.45 across comparable variants while also improving property coverage.","The 31 Louvain communities of the best graph form spatially coherent units that cross parish boundaries only at uncertain edges, and typically coalesce around a street, market, cemetery, or borough—consistent with neighborhoods as lived spaces, not administrative cells."],"supporting_citations":[{"why":"Supplies the NLP pipeline and relational database of Avignon terriers that all 16 graph extractions draw from.","marker":"[13]"},{"why":"Provides the proposed reconstruction of Avignon's parish territories, used in georeferencing and community interpretation.","marker":"[20]"},{"why":"Provides the medieval street network and street names that anchor georeferencing and the street-splitting step.","marker":"[29]"},{"why":"Locates the boroughs outside the old walls, used as surface invariants for property placement.","marker":"[19]"},{"why":"Locates the cardinalatial liveries, another class of invariant used in property georeferencing.","marker":"[21]"},{"why":"Provides the Louvain algorithm that produces the 31-community partition.","marker":"[6]"},{"why":"Defines modularity, used to assess and compare the candidate partitions.","marker":"[28]"}],"fun_headline_variants":["Medieval neighbor mentions beat full data for mapping Avignon","Best Avignon map ignores some medieval records, adds extras","Neighbor graphs from tax rolls reveal 31 medieval neighborhoods","Tax-roll neighbor links map Avignon with 0.80 accuracy","Ignoring some data yields best medieval urban segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's yardstick is not fully independent: the 2,049 property positions used to test distance accuracy were themselves inferred from the same neighbor relationships that build the graph, so part of the 0.80 match may be the method measuring its own input.","fun_headline_variants_meta":{"raw":{"variants":["Medieval neighbor mentions beat full data for mapping Avignon","Best Avignon map ignores some medieval records, adds extras","Neighbor graphs from tax rolls reveal 31 medieval neighborhoods","Tax-roll neighbor links map Avignon with 0.80 accuracy","Ignoring some data yields best medieval urban segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000167,"raw_usage":{"total_tokens":1323,"prompt_tokens":1079,"completion_tokens":244,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":695,"completion_tokens_details":{"reasoning_tokens":162}},"tokens_in":695,"tokens_out":244,"duration_ms":3264,"temperature":1.0,"reasoning_tokens":162,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:48:04.289359+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the EFS_k extraction, restrict the evaluation to a control set of properties whose medieval positions are independently fixed by surviving archaeology or standing landmarks (churches, gates, walls), and recompute the Spearman correlation; a large drop from 0.80 would show the reported spatial fidelity was an artifact of the confront-based georeferencing loop.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the NLP pipeline and relational database of Avignon terriers that all 16 graph extractions draw from."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the proposed reconstruction of Avignon's parish territories, used in georeferencing and community interpretation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the medieval street network and street names that anchor georeferencing and the street-splitting step."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Locates the boroughs outside the old walls, used as surface invariants for property placement."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Locates the cardinalatial liveries, another class of invariant used in property georeferencing."}],"review_version":1}