Companion artefacts for: A held-out damped-persistence baseline and Hobday definition-sensitivity for coastal SST machine learning (controlled ablation of a MARINA-style panel Transformer on 27 Korean stations)

Synthetic hourly sea surface temperature series for 30 Korean coastal monitoring stations, generated so that the released re-analysis code for the companion study executes without the private observational records. Covers a training span of 2018-06-01 to 2024-08-31 (1,644,480 rows) and a validation span of 2024-09-01 to 2025-08-31 (262,800 rows), with columns TIMESTAMP, LOCATION_ID and TEMP, which are exactly the fields the released analysis consumes. Each station's series is generated as a day-of-year climatological mean plus a diurnal cycle plus an AR(1) residual that is split into a shared across-station component and an independent one, so a single parameter controls pairwise correlation. Generation is deterministic given the seed recorded in surrogate_params.json. The record also contains reference outputs from running experiments 005, 006 and 007 on this surrogate, so a user can confirm their own run matches. IMPORTANT: this surrogate does not reproduce the numbers reported in the paper. Model-versus-model results in the companion code record recompute exactly from the archived predictions and need no observations at all (28 of the 36 pre-registered RMSE tests). The remaining eight baseline comparisons and all marine-heatwave indices require the observed validation temperatures; on this surrogate they execute but score synthetic observations against forecasts produced by models trained on the real records, so their values are a smoke test for the code and not a replication of the analysis. That limitation is structural, because the archived forecasts cannot be regenerated for synthetic inputs without retraining. Disclosure footprint: per station a day-of-year climatological mean curve, an AR(1) coefficient, a residual scale, a diurnal amplitude and a diurnal phase; globally one pooled cross-station correlation and one seed. No per-observation value from the source records is retained.