1. Introduction and classification
A completed computation is not necessarily a validated forecast. These entries distinguish models fitted or evaluated against archived environmental estimates from tests on constructed numerical fields, replays of retained output arrays and 3D visual teaching models. Archived reanalysis or satellite–gauge gridded products are not identical to independent station truth. A retrospective time split is not automatically a prospective test if overlapping years were previously inspected. Numerical conservation of a supplied aggregate is not atmospheric water-budget closure.
The record was reconstructed from accessible public research source files, associated dated experiment reports and the author's connected research archive. Different manuscript editions may describe related runs; repeated filenames were not counted as separate discoveries. This is a maintained research inventory as of 31 October 2026, not a claim that every historical notebook has been located or reproduced.
2. Archived environmental-data experiments
These studies were executed against documented historical climate products. They differ in outcome and scale, so cross-study scores must not be ranked as if they shared the same target or units.
Monthly Jamaican rainfall: hard, soft and unconstrained downscaling
Rainfall · Spatial allocation | retrospective (Recorded experiment using archived environmental product; not necessarily prospective or independently observed.)Inspect a sampled field
The carriage represents an observation with a particular footprint.
- Material
- CHIRPS v2 monthly rainfall, January 2000–December 2025; 360 Jamaican fine cells within 29 coarse parent regions. Retrospective scoring: 2022–2025.
- Method
- Matched neural downscalers; an exact parent-aggregate equality, a soft penalty and an unconstrained network. The test supplies the observed target-month coarse rainfall, so it is conditional downscaling and not a prospective precipitation forecast.
- Recorded result
- Three-seed source-reported RMSE: unconstrained 15.36 mm/month; exactly constrained 14.23 mm/month (7.4% lower). A seasonal spatial-allocation baseline scored 9.73 mm/month and outperformed both neural variants.
- Limitations
- The original training has not been independently rerun in this website review. Results are conditional on a coarse field derived from the target month and are not independent-station or operational forecast skill.
Linked working paper or evidence ↗Source 2 ↗Source 3 ↗
Source record: Jamaica Conservation Benchmark, revised 8 October 2026. Research archive documents are not redistributed through this page.
What happens when the coarse rainfall driver is biased?
Downscaling · Sensitivity | retrospective (Recorded experiment using archived environmental product; not necessarily prospective or independently observed.)Account for one shared transfer
Two chambers begin with a combined amount of one constructed unit.
- Material
- Same retained 2022–2025 monthly Jamaican CHIRPS fields and frozen downscaling models; imposed coarse rainfall perturbations and historical-driver substitutions.
- Method
- Vary input coarse rainfall by ±10% and ±20%, then replace target-month observed coarse values with past-available seasonal or previous-year alternatives. The model is not retrained for each intervention.
- Recorded result
- With +20% driver perturbation, hard-model RMSE reached 34.87 mm/month; the compulsory coarse-aggregation-error lower bound was 31.67 mm/month. Substituting a seasonal driver gave 84.49 mm/month RMSE; previous-year same-month coarse rainfall gave 109.84.
- Limitations
- The perturbations are controlled sensitivity exercises, not a distribution of operational forecast driver errors. Low conservation residual does not imply useful rainfall estimation.
Linked working paper or evidence ↗Source 2 ↗
Source record: Coarse-driver error in conserving Jamaican rainfall downscalers, revised 8 October 2026. Research archive documents are not redistributed through this page.
When recent rainfall is unavailable: input-latency sensitivity
Rainfall · Information age | retrospective (Recorded experiment using archived environmental product; not necessarily prospective or independently observed.)Inspect a time archive
A time slice selects a stored state; its coordinate is not a physical journey into the past.
- Material
- CHIRPS daily Jamaica-wide and Kingston-area series; 1,461 test dates in 2022–2025; assumed extra input delays of 0, 7, 21 and 45 days.
- Method
- Compare a 1,377-parameter two-hidden-layer neural regressor, ridge regression, persistence and seasonal climatology, keeping target dates aligned. The delay is imposed on archived predictors.
- Recorded result
- Jamaica-wide zero-delay RMSE: neural 6.871 mm/day versus seasonal 7.598. With an assumed 45-day delay: neural 8.015 versus seasonal 7.614, reversing the comparison. Ridge beat the neural model for the Kingston cell even at zero delay.
- Limitations
- Observed product release vintages were not reconstructed. These are idealised information-availability scenarios, not an operational latency audit.
Linked working paper or evidence ↗Source 2 ↗
Source record: When recent rainfall is unavailable, research draft, 6 October 2026. Research archive documents are not redistributed through this page.
Can isotonic recalibration improve Caribbean wet-day probabilities?
Probability · Calibration | retrospective (Recorded experiment using archived environmental product; not necessarily prospective or independently observed.)Compare a reference and an estimate
The left specimen is fixed at 0.60 constructed unit.
- Material
- Seven selected Caribbean CHIRPS grid cells; wet day defined as ≥1 mm/day; 2022–2025 test period; isotonic calibration fitted on 2020–2021.
- Method
- Compare pooled neural probability estimates, local logistic models, site-specific isotonic adjustments and monthly-frequency references using Brier score and secondary verification.
- Recorded result
- Equal-site-weighted Brier scores: raw logistic 0.18454, raw neural 0.18675, recalibrated neural 0.18846. Isotonic recalibration worsened the neural Brier score at all seven assessed cells; raw logistic beat the raw neural at five.
- Limitations
- Seven selected cells are not a representative sample of Caribbean countries; retrospective CHIRPS-derived events are not independent station truth.
Linked working paper or evidence ↗Source 2 ↗
Source record: Recalibration does not always help, working draft, 6 October 2026. Research archive documents are not redistributed through this page.
Regional pooling with one or three years of local rainfall labels
Data scarcity · Geographic transfer | retrospective (Recorded experiment using archived environmental product; not necessarily prospective or independently observed.)Coordinate local numerical fields
Each station represents a numerical subdomain with an explicit physical state.
- Material
- Kingston, western Jamaica and eastern Jamaica held out on 2022–2025 CHIRPS daily rainfall, with one or three local training years; additional source cells in the Caribbean.
- Method
- Pool six regional source cells into a supervised neural model, compare against identical local neural architecture, local ridge and seasonal references. Test restricted local label supply with recent rainfall-history covariates still available.
- Recorded result
- Pooling beat the local neural predictor in all six target/window configurations, but the local ridge model remained better than pooling at Kingston. West Jamaica, one local year: pooled neural RMSE 11.275 versus local neural 13.119 mm/day.
- Limitations
- More data produced more gradient updates; compute was not matched. The experiment mixes within-Jamaica and cross-country sources and is not zero-shot geographic transfer.
Linked working paper or evidence ↗Source 2 ↗
Source record: Regional pooling under restricted local labels, research draft, 6 October 2026. Research archive documents are not redistributed through this page.
Timestral foundation checkpoints on Jamaican rainfall
Foundation models · Rainfall | retrospective (Recorded experiment using archived environmental product; not necessarily prospective or independently observed.)Compare candidate trajectories
Each cartridge represents a candidate trajectory on the same comparison window.
- Material
- Actual Timestral-5M and Timestral-28M pretrained inference, 312 held-out weekly dates in each of two CHIRPS-derived Jamaican rainfall series, 2020–2025.
- Method
- Zero-shot model forecasts compared with persistence, Ridge and climatology at one-, three- and seven-day accumulation horizons. Evaluate idealised observation age and an imposed 45-day delay.
- Recorded result
- At zero assumed delay, island next-day RMSE was 7.34 mm for Timestral-5M, 7.21 for Timestral-28M, 6.62 for persistence and 6.75 for Ridge. Of 15 observed one-day extremes, each Timestral model detected three while persistence detected nine. Under a 45-day input delay, climatology outperformed both pretrained systems across site–horizon RMSE comparisons.
- Limitations
- Timestral-209M was not executed; no spatial nowcasting, gauge-truth evaluation or historically accurate product-vintage replay was established. Public model ranking is not regional skill.
Linked working paper or evidence ↗Source 2 ↗
Source record: Jamaica Timestral Research Preprint v0.2, 16 October 2026. Research archive documents are not redistributed through this page.
TimesFM, Chronos and a reduced physics-guided Jamaican temperature forecast
Foundation models · Temperature | retrospective (Recorded experiment using archived environmental product; not necessarily prospective or independently observed.)Compare candidate trajectories
Each cartridge represents a candidate trajectory on the same comparison window.
- Material
- Daily ERA5-derived two-metre temperature at three Jamaican-labelled grid cells, 192 weekly seven-day forecast origins over 2023–September 2026 (38 in 2026).
- Method
- Compare AR(14), frozen TimesFM generations, Chronos-2 and reduced graph/PINN corrections under common retrospective origins; satellite-derived predictors and six air-cell histories are selectively added.
- Recorded result
- Full-period source-reported physics-containing ensemble MAE 0.4966 °C, 3.89% below AR(14) and only 0.35% below TimesFM 2.5. For 38 origins in 2026 MAE was 0.5017 °C, 1.41% worse than TimesFM 2.5. Matched physical-loss advantage in the 2026 subset was 2.20%, not a full-model win.
- Limitations
- Exploratory comparisons reused previously inspected historical periods. Targets were ERA5-derived grid cells, not independent weather stations. Extra satellite features and added cells lacked consistent gain.
Linked working paper or evidence ↗Source 2 ↗
Source record: Satellite Predictors, Reduced Physics and Foundation Models for Daily Temperature Forecasting in Jamaica, 14 October 2026. Research archive documents are not redistributed through this page.
Archived WeatherNext 2 / ECMWF IFS forecast comparison and post-hoc correction
Numerical weather prediction · Calibration | retrospective (Recorded experiment using archived environmental product; not necessarily prospective or independently observed.)Compare candidate trajectories
Each cartridge represents a candidate trajectory on the same comparison window.
- Material
- 16 common archived forecast origins in 2026; separate eight-week calibration/evaluation halves from that cohort.
- Method
- Compare actual archived WeatherNext 2 ensemble-mean and ECMWF IFS forecast products with matched local baselines; fit a simple site-wise mean-bias offset on the first eight dates and apply it on the remaining eight.
- Recorded result
- Large raw site bias, especially at the Mandeville-labelled grid cell, contributed to forecast error. Chronological bias adjustment reduced some WeatherNext error, but did not beat TimesFM on the eight reported evaluation weeks.
- Limitations
- A post-hoc diagnostic after viewing cohort-level outcomes is exploratory, not a confirmatory holdout. The atmospheric archive has only 16 origins; product comparison does not demonstrate operational superiority.
Linked working paper or evidence ↗Source 2 ↗Source 3 ↗
Source record: Jamaica TimesFM/PINN research manuscript; source calibration manifest in public/papers/evidence/weather_calibration.json. Research archive documents are not redistributed through this page.
3. Reproduction and source verification
Saved-checkpoint replay and rainfall budget verification
Reproducibility · Numerical integrity | verification (Recount or internal numerical consistency of saved artifacts; not a fresh model fit.)Compare a reference and an estimate
The left specimen is fixed at 0.60 constructed unit.
- Material
- 36 saved neural model checkpoints and retained prediction arrays from the October Jamaican conservation revision package.
- Method
- The archived follow-up validation script replayed checkpoint predictions, rescored 24 groups and inspected aggregate conservation residuals. These tests are internal to the stored experiment.
- Recorded result
- Recorded replay agreement had zero maximum prediction difference. Follow-up maximum exact-constraint residual was approximately 0.0001905 mm/month, below the script threshold; 21 of 36 fits reached their optimisation cap.
- Limitations
- A checkpoint replay checks serialization and calculations, not independent training reproduction, true moisture conservation or geographic generalisation.
Linked working paper or evidence ↗Source 2 ↗
Source record: EMPIRICAL READINESS — dLEGOS Jamaica Conservation Results and 15 January Plan, 10 October 2026. Research archive documents are not redistributed through this page.
Recalculation of archived temperature metrics and source-quality audit
Reproducibility · Data quality | verification (Recount or internal numerical consistency of saved artifacts; not a fresh model fit.)Compare a reference and an estimate
The left specimen is fixed at 0.60 constructed unit.
- Material
- 3,520 MAE/RMSE rows from archived prediction arrays, 192 forecast origins, NASA POWER radiation and NOAA OISST missing-data records.
- Method
- Recompute published arithmetic independently from saved predictions; preserve source freeze hashes, detect duplicated solar radiation grid histories, enforce past-only forward fill, and document data lags.
- Recorded result
- The source verification manifest records PASS for independent recount of all 3,520 primary rows. NASA POWER series had 61 missing tail days in several requested locations, and multiple requested locations shared identical radiation histories.
- Limitations
- Arithmetic recount does not mean retraining a foundation model, independent temperature truth or reissuing real operational weather forecasts.
Linked working paper or evidence ↗Source 2 ↗Source 3 ↗
Source record: public/papers/evidence/verification.json and source_audit.json, 14–16 October 2026. Research archive documents are not redistributed through this page.
4. Constructed numerical experiments
The numerical models in this section use constructed fields, with equations and parameters selected to test code behaviour. They are not atmospheric forecast experiments.
Constructed graph heat-exchange learning and exact ledger
Numerical method · Graph coupling | constructed (Executed model tests with mathematically constructed data, not observed climate.)Coordinate local numerical fields
Each station represents a numerical subdomain with an explicit physical state.
- Material
- 12 constructed thermal nodes, 11 shared edges, 192 constructed training fields and 96 unseen constructed fields on the same graph.
- Method
- Train a 3–8–1 positive-conductance MLP for 1,200 steps against a constructed graph diffusion operator; check gradients by finite differences, stability bound, conservation, convex energy and rollouts.
- Recorded result
- Source report: learned tendency RMSE 0.003464 K/s compared with 0.634466 K/s for a fixed-conductance control, with maximum recorded energy drift 1.36 × 10⁻¹² J. These units and capacities are constructed.
- Limitations
- Targets and learner use the same diffusion class on the same graph; no new geography, meteorology, forcing or incomplete observation is tested.
Linked working paper or evidence ↗Source 2 ↗Source 3 ↗
Source record: public/papers/development/results/report.json, 16 October 2026. Research archive documents are not redistributed through this page.
Calculated energy-balance values
Finite-volume advection refinement, projections and CRPS checks
Numerical method · Verification | constructed (Executed model tests with mathematically constructed data, not observed climate.)Compare a reference and an estimate
The left specimen is fixed at 0.60 constructed unit.
- Material
- Constructed one-dimensional transport at 32, 64, 128 and 256 cells; constrained projection and a two-member ensemble score example.
- Method
- Compare numerical cell averages to exact advection, measure L1 convergence and mass drift; test equality projection with a non-negativity counterexample; compute empirical versus fair CRPS.
- Recorded result
- Advection L1 error decreases from 0.04282 at 32 cells to 0.00551 at 256 cells, approximately first-order convergence, with zero recorded mass drift. Equality projection exactly hits its sum but is not guaranteed nonnegative; two-member empirical/fair CRPS example yields 0.5/0.0.
- Limitations
- This validates a toy numerical method and accounting identities, not general atmospheric transport skill.
Linked working paper or evidence ↗Source 2 ↗Source 3 ↗
Source record: public/papers/development/results/report.json, 16 October 2026. Research archive documents are not redistributed through this page.
CLEAR-PINN synthetic source-closure and adverse controls
PINN · Identifiability | constructed (Executed model tests with mathematically constructed data, not observed climate.)Account for one shared transfer
Two chambers begin with a combined amount of one constructed unit.
- Material
- One-dimensional manufactured transport/source problem, matched noise controls and several learned-source penalties; no real precipitation measurements.
- Method
- Compare zero-source transport, source-learning neural variants, tuned regularization, adaptive weighting, coefficient-space dead zone and information-matched averaged noisy replicas.
- Recorded result
- The information-matched pooled baseline produced sourced-case RMSE 0.004612 against 0.006545 for the dead-zone learner, a worse result for the proposed adaptation. Adaptive variants also deteriorated in a zero-source control; all neural runs reached iteration cap.
- Limitations
- The novel closure tolerance and adaptive rules did not establish an incremental advantage in the synthetic study; full atmospheric architecture, nowcasting skill and efficiency remain untested.
Linked working paper or evidence ↗Source 2 ↗
Source record: CLEAR-PINN v0.1 research draft, 20 October 2026. Research archive documents are not redistributed through this page.
5. Interactive model demonstrations
World-model numerical budget and 3D visualisation stress checks
Interactive simulation · Constructed physics | illustrative (Software animation, visual or numerical demonstration; no operational validation.)Inspect a miniature scenario table
The table is a constructed scenario with model-space dimensions.
- Material
- Five environmental demonstration presets and a procedurally illustrated coastal community in the public world-model; figures, population and hazards are constructed.
- Method
- Test finite transforms, simple water storage/budget closure, overflow and drying boundaries, time stepping and animation state across scenarios; render a coherent visual explanation of linked local phenomena.
- Recorded result
- Repository review records passing numerical and scene checks, including budget closure, finite transforms and scenario transitions. These are interface/numerical checks, not observations or forecasts.
- Limitations
- Synthetic terrain, residents and weather effects cannot establish local impact predictions or a trained three-dimensional atmospheric model.
Linked working paper or evidence ↗Source 2 ↗
Source record: reviews/2026-10-23-world-model-design-audit.md and tests/world-model.mjs. Research archive documents are not redistributed through this page.
6. What the evidence supports and what remains open
The strongest recurring result is methodological rather than promotional: a model can obey a numerical constraint, receive additional predictors or use a larger pretrained network without improving predictive skill over a well-matched local baseline. The seasonal allocator was more accurate than both tested neural rainfall downscalers. Weather-data latency reversed a favourable neural-versus-climatology comparison. Calibration and adaptive residual weighting produced negative results in some configurations. All are scientifically useful because they narrow the conditions under which added model complexity warrants further investigation.
The following claims are not supported by the inspected experiment records: a validated three-dimensional Caribbean weather model, formal proof of the learned atmospheric constraints, operational extreme rainfall nowcasts, independent station verification for every target, or successful physical reconstruction of Hurricane Beryl. Those remain research objectives. For their formulations and falsification criteria, consult the theory and ideas manuscript.
Research gaps identified in the source register include incomplete source-release vintage audits, insufficient independent stations, unmatched computation between pooled and local learners, limited cross-event and cross-basin holdouts, the need for strict no-future-information issue-time protocols, and the absence of a verified full-architecture atmospheric model. These limitations should guide the next experiment rather than be omitted from paper abstracts.
7. Reproducibility and source provenance
The climate measurements derive from archived research packages, including CHIRPS, ERA5-derived products, NASA POWER and NOAA OISST. The web series contains temperature-model comparisons, observation-support methods, executable numerical development and the source bibliography. Original private code packages, model checkpoints and some manuscript drafts remain in the research archive and are not automatically public. The table below specifies what the website can substantiate without distributing those private materials.
| Source family | Verification and limitation | Accessible record |
|---|---|---|
| Jamaican monthly CHIRPS | Source-reported results, internal checkpoint replay; stronger seasonal comparator retained. | Working research article |
| Jamaican temperature | 3,520 archived arithmetic rows independently recounted; not fresh foundation-model training. | Verification manifest |
| Manufactured graph and advection | Executed constructed tests, gradients and energy identities; not atmospheric skill. | Run report and source |
| Original source manuscripts | Several studies available in owner-controlled Drive with audit and source-version requirements. | Methods and boundaries |
Selected methodological literature: Funk et al. (2015), CHIRPS; Raissi et al. (2019), PINNs; Gneiting & Raftery (2007), proper scoring; Harder et al. (2023), hard-constrained downscaling.
8. Conclusion
This programme now has several interpretable empirical and numerical components. Its unresolved central hypothesis is whether conservation-aware local atmospheric models, supplemented by reliable data and explicit uncertainty, can improve Caribbean storm reconstruction and prediction under realistic observational constraints. The next decisive results must come from independent data, event holdouts and comparisons in which simpler methods receive a fair computational budget.
Disclosure: AI-assisted coding, writing and editorial review were used in this research programme. The author remains responsible for confirming the recorded source results, rights and venue disclosures before formal publication. These pages are research working manuscripts, not peer-reviewed journal articles.
Continue to the theory and hypothesis register → · Return to the eight linked papers