Client-side Verification Wrapper

Documented MET verification math, run entirely in your browser
Theme 01 · Client-side wrapper
Synthetic data
Loading synthetic matched pairs…

Query the matched pairs — faceted filter

Valid time range (ISO, inclusive)
Selection: — rows. Plain-JS columnar scan (gestures at DuckDB-style querying).

Threshold & operator — binarize → 2×2

bounds —
Suggested thresholds

Compiled query — read-only, not executed


      


      

        
Load your own CSV — parsed locally

        

          Drag & drop a matched-pairs CSV here
          
(same schema), or
          
          

        

        
Never uploaded — read in-browser via FileReader.
          Columns: model, lead, region, valid, fcst, obs, fcst_prob, obs_event.

      


      

        
Dataset — provenance

        

      

    


    
    


      

        
2×2 contingency table

        

          

            
              
obs = yesobs = no

            
            
              

                fcst = yes
                —a · hits
                —b · false alarms
              

              

                fcst = no
                —c · misses
                —d · correct neg.
              

            
          

          
n = —

            event when value 

        

      


      

        
Categorical statistics — hover/focus a tile for its exact formula

        

      


      

        
Continuous statistics — from filtered raw pairs

        

        

      


      

        

          
Performance diagram

          
          
x = SR (1−FAR), y = POD. Blue curves = CSI contours
            CSI = 1/(1/SR + 1/POD − 1); gold dashed = bias lines POD/SR = const.

        

        

          
Reliability diagram + sharpness

          
          
Observed relative frequency vs mean forecast probability
            (10 bins). Diagonal = perfect reliability; inset = sharpness (n / bin).

        

      


      

        
MET fidelity — exact formula behind every number

        

        

          Parity with reference MET (point_stat / grid_stat CTS & CNT line types) is the
          headline open question this prototype exists to test — bit-for-bit
          parity is not claimed for the MVP. GSS/HSS use the standard
          sample-climatology reference; empty categories / zero variance render as “—”.
        

      


      

        
Future architecture — named, NOT depended on

        

          

            
Zarr v3

            
Lazy range-request access to chunked gridded model output for
              spatial / grid_stat verification (replaces “load the whole CSV”).

          

          

            
Parquet + DuckDB-WASM

            
Columnar matched-pair store; the faceted filter + “compiled
              query” string above is the placeholder for real SQL pushed to DuckDB-WASM.

          

          

            
Rust + WASM

            
Port the verified stats.js kernels for speed and to
              share one math implementation with MET-adjacent code.

          

          

            
WebGPU / WGSL

            
Per-cell / per-pair recompute on the GPU for interactive what-if
              over very large grids.

          

        

        
None of these are imported or required by this prototype.
          It runs offline with vanilla HTML/CSS/JS only.