NEL1: Claim-Gated Tree-Ring Forecasting under Source Shift and Operational Constraints

Abstract

NEL1 presents a nine-stage technical whitepaper and numerical evidence ledger for one-step tree-ring width forecasting under temporal dependence, source shift, uncertainty, and operational constraints. The study uses two public NOAA/WDS Paleoclimatology collections from Kaindy Lake, Kazakhstan: KAZ002, comprising living-tree series, and KAZ003, comprising dead-tree series. The audited corpus contains 5,693 measurements from 60 independent series and yields 5,393 one-step modelling rows. The evidence chain evaluates data provenance and split dependence, performance beyond predeclared baselines, endogenous memory depth, incremental feature value, source/status proxy identifiability, series-balanced heteroscedastic uncertainty, versioned prequential deployment, selective prediction and abstention, sequential drift recovery, and integrated fail-closed operational closure. The results reveal genuine but qualified predictive structure. Complete-series and source-held-out evaluations contain no forbidden target or future leakage, whereas the row-random diagnostic places approximately 98.4% of test targets or future labels inside training features. Candidate models achieve local improvements over Persistence, and a conservative five-lag endogenous-memory core is retained. However, no candidate passes the universal performance gate. Uncertainty policies remain source-specific and do not transfer universally; selective prediction confirms in only one temporal epoch and fails both source-level gates; and no causal drift-recovery policy survives policy validation. The final supported classification is NO_UNIVERSAL_SAFE_STACK. Neither source satisfies the complete deployment contract, and both therefore enter ABSTAIN_ALL_AUTOMATIC_OUTPUTS. Versioned telemetry, exact replay, artifact verification, ledger integrity, terminal anchoring, tamper rejection, and the public/private release barrier remain operational. The scientific contribution is not an asserted universal forecasting service or a new standalone forecasting algorithm. It is an integrated, falsifiable, leakage-aware and source-aware evidence architecture that separates local predictive skill from deployment admissibility, preserves negative results, and demonstrates safe rejection when local improvements cannot satisfy end-to-end operational gates. This record includes the final technical whitepaper and a consolidated curated public-evidence package containing aggregate tables, figures, manifests, checksums, and public-release audits. Source code, fitted model objects, row-level modelling arrays, exact split assignments, and reconstruction-enabling private materials are not included. The study is not a climate reconstruction, a causal ecological model, a field-deployed forecasting service, or a geographically general model for Kazakhstan.