Delete RESEARCH_VALIDATION.md

RESEARCH_VALIDATION.md

@@ -1,79 +0,0 @@
-# Research validation notes
-
-This release is a **fully synthetic benchmark**. The validation goal is internal structural, physical and statistical consistency, not proof that the generated distribution is empirically identical to real Kolkata monitoring data.
-
-## Why the generator changed
-
-The earlier 100M draft had two important benchmark weaknesses:
-
-1. PM2.5 correlations were almost distance-invariant across synthetic sites, so geographic coordinates did not carry enough predictive structure.
-2. Missingness was all-or-nothing across all variables, unlike monitoring systems where pollutant channels often fail independently.
-
-The current generator fixes both issues.
-
-## Spatial dependence
-
-The generator uses 24 latent spatial knots with persistent time-varying AR processes. Each site mixes its four nearest knots with Gaussian distance weights. Pollutant families use different latent fields for fine particles, coarse particles, combustion gases and ozone.
-
-On the final materialized data, a 120-site × 6,000-hour PM2.5 diagnostic gives:
-
-- distance vs raw pairwise correlation: **r = -0.7125**
-- mean raw correlation below 5 km: **0.9546**
-- mean raw correlation at 30+ km: **0.9133**
-- mean cross-sectionally residualized correlation below 5 km: **0.2610**
-- mean residualized correlation at 30+ km: **-0.0570**
-
-The point is not that these exact coefficients are universal physical truths. The point is that spatial proximity now has measurable statistical meaning rather than being decorative metadata.
-
-## Temporal fidelity
-
-The table is a complete hourly grid, including timestamps at which some channels are unavailable. Final timestamp metrics are:
-
-- order consistency = 1.0
-- timestamp uniqueness = 1.0
-- hourly regularity consistency = 1.0
-- grid completeness = 1.0
-
-Station-0 autocorrelations show persistent hourly, diurnal and weekly structure rather than independent row sampling. The validator also records first-difference diagnostics so sudden-step behavior can be inspected directly.
-
-## Missingness
-
-Mean missingness in a 100-site diagnostic, confirmed by the full-table scan, differs by variable. Full-table rates are stored in `validation_report.json`. Typical rates are approximately:
-
-- temperature: 1.9%
-- relative humidity: 2.1%
-- wind: 2.4%
-- rainfall: 2.5%
-- PM2.5: 3.2%
-- PM10: 3.8%
-- NO2: 5.5%
-- CO: 6.5%
-- SO2: 7.5%
-- O3: 6.0%
-
-Outages occur in contiguous runs, with a whole-site telemetry component plus variable-specific failures.
-
-## Cross-variable logic
-
-The final station-0 diagnostics include:
-
-- PM2.5 vs PM10: strong positive correlation, but not identity
-- PM2.5 vs wind: negative
-- PM2.5 vs rainfall: negative
-- NO2 vs O3: negative under the synthetic titration/photochemical mechanism
-
-PM10 is generated as PM2.5 plus an independently varying coarse component. NO2, CO and SO2 share a combustion-related spatial field but have separate station-level innovations so they are related without being mechanically identical.
-
-## Literature alignment
-
-Three recent research directions directly informed the validation design:
-
-- **Seq2Synth (arXiv:2607.15606)**: static marginals are insufficient for sequential data; timestamp validity, cross-sectional dynamics, first differences, autocorrelation and trajectory-level structure should be checked directly.
-- **DynLMC (arXiv:2604.05064)**: realistic multivariate synthetic data should not rely only on fixed correlations; time-varying and lagged cross-channel dependence matters.
-- **AirQualityBench (arXiv:2605.05854)**: air-quality benchmarks benefit from explicit missing-data masks, physical-scale evaluation, graph structure and evidence that spatial dependence decreases with distance.
-
-The generator incorporates these ideas through dynamic spatial fields, lagged rainfall scavenging, pollutant-specific observation masks, a released k-NN graph and temporal/spatial diagnostics.
-
-## Remaining scientific limitation
-
-No synthetic generator can be declared empirically equivalent to Kolkata observations solely from internal checks. A stronger future validation would compare this benchmark against a clean, legally redistributable real reference set at matched hourly/monthly grain using distributional, spectral, cross-correlation and forecast-transfer metrics. The current release therefore uses the label **literature-informed synthetic benchmark**, not historical reconstruction or digital twin.