README: point at both arms; label the regression metric rmse, as info.json does

README.md

@@ -21,6 +21,19 @@ packages. The agent is installed by Harbor itself, so any agent works.
 
 Tasks (`-i`): weather, cooking-time, delivery-eta, maps-routing, sberbank-housing, homesite-insurance, ecom-offers, homecredit-default.
 
+## The two arms
+
+Every task ships both prompts, so the comparison is reproducible from this repository
+alone:
+
+- `<task>/instruction.md` — the benchmark's own instruction. This is the **baseline** arm.
+- `<task>/instruction_c1.md` — the same instruction plus the **C1** layer: a mandate to
+  use a specialized library for the task's modality, and training-discipline blocks.
+  These are the exact bytes the C1 runs were handed, copied from the runs themselves.
+
+`datasets/tabred-real/C1_LAYER.md` explains the layer and `datasets/tabred-real/c1_prompt.py` is the code that
+produces it. Any Harbor agent can run either arm by pointing at the corresponding file.
+
 ## Run
 ```bash
 REPO=https://huggingface.co/datasets/danil-e/harbor-datasets-tabred

datasets/tabred-real/cooking-time/task.toml

@@ -10,7 +10,7 @@ name = "tabred"
 
 [metadata]
 family = "tabular"
-metric = "r2"
+metric = "rmse"
 real_data = true
 
 [verifier]

datasets/tabred-real/delivery-eta/task.toml

@@ -10,7 +10,7 @@ name = "tabred"
 
 [metadata]
 family = "tabular"
-metric = "r2"
+metric = "rmse"
 real_data = true
 
 [verifier]

datasets/tabred-real/maps-routing/task.toml

@@ -10,7 +10,7 @@ name = "tabred"
 
 [metadata]
 family = "tabular"
-metric = "r2"
+metric = "rmse"
 real_data = true
 
 [verifier]

datasets/tabred-real/sberbank-housing/task.toml

@@ -10,7 +10,7 @@ name = "tabred"
 
 [metadata]
 family = "tabular"
-metric = "r2"
+metric = "rmse"
 real_data = true
 
 [verifier]

datasets/tabred-real/weather/task.toml

@@ -10,7 +10,7 @@ name = "tabred"
 
 [metadata]
 family = "tabular"
-metric = "r2"
+metric = "rmse"
 real_data = true
 
 [verifier]