From data availability to evidence sufficiency. PIH MS as an auditable framework for diagnosing hydrogeological knowledge gaps in Mato Grosso do Sul, Brazil

Carlos Busón , Sandra Gabas

Groundwater registries can be extensive yet fail to demonstrate the evidence chain required to answer a hydrogeological question. We present PIH MS, an experimental framework that audits a bounded corpus of acquired records and makes knowledge gaps visible without predicting groundwater occurrence. The framework organises 12 evidence layers, nine knowledge dimensions, 39 declared requirement instances, five questions, and six families of scientific units. The corpus contains 3,877 provisional well identifiers. The audit recovered 1,591 stable hydraulic-test identities, 1,970 stable chemical-sample identities, 3,213 exact well-to-test-date links for hydraulic parameters, and 2,557 consistent chemical result-to-sample links. These relationships did not form a complete evidence chain. Captured-interval identity, verified units, complete analytical quality assurance, stable same-variable series, and a declared monitoring-network design remained unavailable. No complete case was demonstrated for Q01 to Q04 or Q05a, and Q05b could not be evaluated. This result diagnoses corpus fitness under explicit evidence contracts. It neither indicates methodological failure nor implies an absence of hydrogeological knowledge in the state. Controlled synthetic inputs reached UNKNOWN and every implemented stage from P1 to P4, confirming that P4 is attainable. Experimental context gates influenced intermediate stages and mapped patterns but have not yet undergone external validation. Full-chain blockers were assessed separately. Twelve question-requirement combinations, corresponding to eight unique fields, acted as universal blockers. Enumeration of 1,472 subsets yielded minimum failure distances of one, three, three, two, and four requirements for Q01 to Q05, respectively. Five grid scales and four origins were compared at 14,284 fixed support points. Overall agreement among origins ranged from 64.24% to 66.81%, whereas agreement at locations classified under at least one origin ranged from 29.82% to 34.86%. The primary grids comprised 45,145 cell-question pairs, of which 32,405 were UNKNOWN, 2,801 P1, 5,448 P2, 4,491 P3, and none P4. P5 was not implemented. A single execution regenerated 21 priority outputs as byte-identical files. The proposed contribution is an auditable architecture that specifies evidence before mapping, preserves unknown states, exposes concurrent blockers, evaluates logical and spatial sensitivity, and retains a reverse path to source records.