LLM Strategic Assessment Experiment: Raw Execution Archive and Reconciled Analysis Dataset

Dataset supporting a repeated LLM-based strategic assessment experiment comprising 1,080 scheduled primary runs across three deployment configurations, three prompt strategies, three reasoning-effort settings, four fixed portfolios, and ten repetitions per scheduled cell. The deposit contains two files: Raw execution archive — the authoritative experiment records, including run and attempt provenance, requests, responses, retries, terminal outcomes, bindings, and associated traceability information. Reconciled analysis-ready dataset — the verified 933-row dataset of completed strategic assessments used for the reported statistical analyses. The experiment evaluates configuration sensitivity, repeated-run variability, portfolio separation, ranking stability, evidence-reference consistency, operational failure, and provenance in LLM-based decision support. The underlying strategic-assessment case uses four fixed LIFE Programme project portfolios with real projects and the real European Directive related to the LIFE Programme and a frozen evidence corpus. Raw execution records remain authoritative. The analysis-ready dataset is a reconciled derivative prepared for reproducible analysis. Post-execution scale harmonisation affecting 20 baseline-prompt observations is retained transparently in the provenance and correction records. This record is anonymised for peer review. Creator metadata will be updated after the review process without altering the deposited research files.