Add masked-reconstruction evaluation results

README.md

@@ -40,12 +40,33 @@ weights remain frozen when training the language-model translator.
 | `config.json` | architecture and input contract used to construct the encoder |
 | `normalization.json` | input mean, standard deviation, and residual scale |
 | `training_args.json` | optional pretraining recipe and checkpoint-selection record |
+| `eval_results.json` | machine-readable masked-reconstruction evaluation |
 
 `encoder.safetensors`, `config.json`, and `normalization.json` are sufficient
 for inference. `training_args.json` is provenance for training reproduction;
 the optimizer and mixed-precision scaler are not distributed. EGMS
 measurements are not included in this model repository.
 
+## Evaluation
+
+The released checkpoint was evaluated on 1,000 held-out tiles containing
+2,047,451 point histories. The same central 88-step interval, 30% of the
+294-step input, was masked for every point in a tile.
+
+| metric | value |
+|---|---:|
+| normalized MSE | 0.0702 |
+| MSE | 2.279 mm² |
+| RMSE | 1.510 mm |
+| MAE | 1.007 mm |
+| pooled EGMS residual | 1.433 mm |
+| per-point RMSE P10 / P50 / P90 | 0.54 / 1.03 / 2.43 mm |
+
+The masked-reconstruction RMSE is close to the pooled residual level of the
+source observations. This result indicates that the frozen representation
+retains the dominant temporal structure under this masking protocol. It is not
+a forecasting result.
+
 ## Use
 
 Install the code and download the checkpoint:

eval_results.json

@@ -0,0 +1,30 @@
+{
+  "schema_version": "egms-qa-encoder-evaluation-1.0",
+  "evaluation": "masked_reconstruction",
+  "dataset": "risenyard/egms-qa-dataset",
+  "split": "test",
+  "sample": {
+    "tiles": 1000,
+    "point_histories": 2047451,
+    "input_steps": 294
+  },
+  "masking": {
+    "strategy": "central_synchronized_block",
+    "masked_steps": 88,
+    "masked_fraction": 0.3
+  },
+  "aggregation": "point_weighted_over_masked_observations",
+  "metrics": {
+    "normalized_mse": 0.0702,
+    "mse_mm2": 2.279,
+    "rmse_mm": 1.51,
+    "mae_mm": 1.007,
+    "pooled_egms_residual_mm": 1.433,
+    "per_point_rmse_mm": {
+      "p10": 0.54,
+      "p50": 1.03,
+      "p90": 2.43
+    }
+  },
+  "scope": "Masked reconstruction of held-out observations; not forecasting."
+}