Atalaya: A Calibrated Multi-Evidence Affinity Graph over an Open-Data Catalog, with a Shuffled-Alignment Negative Control for False Discoveries

Open-data portals publish thousands of datasets as a flat searchable list, but the relational knowledge that makes the data useful, which datasets can be joined, which describe the same thing, and which actually correlate when aligned by a common key, stays implicit and is infeasible to recover by hand. Existing dataset-discovery tools each rank relatedness on a single signal (value containment for joinability, embedding cosine for semantic search, statistical association for correlation mining), and each signal is individually misleading: two datasets can be embedding-similar yet share no join key, joinable on year yet describe unrelated phenomena, or correlate spuriously through a shared driver. We present Atalaya, a system that harvests a national open-data catalog (1,017 datasets from Chile's Data Observatory; a 303-dataset, 24.8 GB government subset mirrored and profiled in full), mines five orthogonal kinds of cross-dataset relation into an explorable knowledge graph, and ranks pairwise relatedness with a calibrated multi-evidence affinity that fuses the evidence with standard empirical-CDF calibration and reliability weighting. The affinity fuses three orthogonal evidences into one interpretable [0,1] score after passing each through its empirical null distribution, so the score means 'stronger than chance' rather than 'large raw number', and it down-weights any evidence that contradicts the others. Two validation measurements accompany the graph. First, the correlation miner is tested with a shuffled-alignment negative control: on the same alignments with the keys shuffled, 343 candidate correlations yield 0 survivors under a permutation null with Benjamini-Hochberg control (empirical false-discovery rate 0.00), which gives empirical support to the 24 correlations reported on real data. Second, the semantic encoder is compared with a classical lexical baseline, over which its measured margin is modest (+0.014 top-5 neighbour-theme coherence over TF-IDF). Every edge carries its evidence in the graph database, and the whole affinity is pure numpy, so the browser recomputes it live as an analyst reweights the evidences. The graph reports where datasets can be related and how strongly relative to chance; it does not assert causation. Source code and derived artifacts (MIT): https://github.com/fsantibanezleal/CAOS_ATALAYA . Live instance: https://atalaya.fasl-work.com .