results: reject unmatched random-beta screen

1 files changed, 23 insertions, 0 deletions

diff --git a/results/ep_bias/s2_summary.json b/results/ep_bias/s2_summary.json
new file mode 100644
index 0000000..036f353
--- /dev/null
+++ b/results/ep_bias/s2_summary.json
@@ -0,0 +1,23 @@
+{
+  "stage": "rain_ep_random_beta_s2",
+  "status": "invalid_unmatched_corruption_development",
+  "rows": [
+    {
+      "seed": 1989,
+      "random_beta_validation_accuracy": 0.4815,
+      "positive_beta_steps": 115,
+      "negative_beta_steps": 122,
+      "final_bias_state_difference_rms": 0.00019334602670050772
+    },
+    {
+      "seed": 1990,
+      "random_beta_validation_accuracy": 0.739,
+      "positive_beta_steps": 115,
+      "negative_beta_steps": 122,
+      "final_bias_state_difference_rms": 0.00017428155723407375
+    }
+  ],
+  "reason_not_comparable": "The development injector normalized its fixed field from the first realized clean beta-state difference. A random-sign run initialized from a different beta sign than the fixed-positive runs and realized roughly 2.5 times their bias RMS. The results validate execution and sign accounting but do not compare methods at matched corruption.",
+  "next_protocol": "Normalize hardware readout bias from the free-state activity only, independently of beta sign and task difference, before comparing fixed-positive EP, random-sign beta, constant calibration and innovation.",
+  "test_policy": "development holdout observed; no performance claim"
+}