Nine sequential reports develop and test an exactly invertible forecasting system; a standalone publication and a shared metric charter sit beside the main sequence.
Research programme
Findings so far
Forecast-task pretraining on CMIP6 simulations produced the strongest confirmed deterministic gain. Residual diffusion improved individual members and temporal coherence on validation, but that advantage transferred only partly to the sealed period.
Companion documents, separate from the nine-report sequence.
Publication
The externally readable product of the whole programme: a CMIP6-pretrained deterministic forecaster and a 32-member residual-diffusion ensemble in the relative-SST frame, characterized on 2021–2025 development validation. Self-contained, and deliberately not ranked against the earlier generations.
Charter
The measurement contract: evidence tiers, the persistent metric identifiers, the house bootstrap, and the comparability rules. Read it first if you want to know exactly what any number in the series is allowed to mean.
Nine reports in reading order, each stating its incoming baseline, the single factor it changes, its predeclared gates, and the artifact it hands forward.
Report 1 · System validation
Builds the thing and proves it is built correctly: lossless coefficient packing, exact inverse reconstruction, auditable checkpoint surgery on the official backbone, and fail-closed access to the sealed period. The scope is the pipeline; skill claims come later.
Report 2 · Deterministic ablations
Nine controlled families ask whether direct adaptation on 335 observed windows can forecast at all. It cannot: low error is bought with near-climatological damping, amplitude costs correlation, and residual learning spends itself cancelling persistence.
Report 3 · Self-supervised ladder
Masked coefficient pretraining over four SST products improves reconstruction monotonically and forecast transfer not at all. The durable products are a coefficient-energy explanation of why, and the statistical governance the rest of the series runs under.
Report 4 · CMIP transfer and sealed test
Pretraining on the actual forecast task across eleven CMIP6 models transfers where reconstruction did not, and the resulting model spends the project's one pristine hash-gated read of the sealed period. This is the strongest confirmatory result in the series.
Report 5 · Subsurface transfer
A controlled test of whether ocean heat content adds anything. Inside simulation it helps; through the observation pathway it resolvably hurts. The zeroed-input twin design is what makes that a causal statement rather than a guess.
Report 6 · Probabilistic heads
Gaussian, latent-noise, and diffusion heads surround the frozen center. Diffusion fixes the smudgy individual member that started the programme, but the regional calibration wall survives every tested change, including sixty-four times more data.
Report 7 · Projection-space calibration
Six rungs of explicit calibration establish what post-processing can and cannot buy. Average reliability can be corrected and lifted into sharp fields, but flow-dependent width does not transfer from simulation and cannot be manufactured after the fact.
Report 8 · Temporal diffusion and final test
Re-weighting the objective across wavelet bands is what unlocks temporal coherence — the joint sequence architecture alone does not reach it. The adopted stack is the best validation product of the series, and its one authorized test read shows the advantage only partly transferred.
Report 9 · Reference re-basing and handoff
The control the earlier reference finding was missing: five pipelines differing only in the anomaly reduction, each retrained from scratch. Retraining, not climatology updating, turns out to explain most of the degradation. Ends by freezing the deployment contract.