summaryrefslogtreecommitdiff
path: root/results/kp_dynamic_projection_confirmation_gate.json
diff options
context:
space:
mode:
Diffstat (limited to 'results/kp_dynamic_projection_confirmation_gate.json')
-rw-r--r--results/kp_dynamic_projection_confirmation_gate.json60
1 files changed, 60 insertions, 0 deletions
diff --git a/results/kp_dynamic_projection_confirmation_gate.json b/results/kp_dynamic_projection_confirmation_gate.json
new file mode 100644
index 0000000..875bdef
--- /dev/null
+++ b/results/kp_dynamic_projection_confirmation_gate.json
@@ -0,0 +1,60 @@
+{
+ "checks": {
+ "all_mechanism_query_tracking_and_cost_invariants": true,
+ "all_records_trajectories_and_metrics_finite": true,
+ "at_least_four_seeds_within_2_points_of_clean_kp": true,
+ "every_dynamic_test_accuracy_at_least_0p87": true,
+ "mean_dynamic_test_accuracy_at_least_0p89": true,
+ "mean_early_alignment_at_least_0p90": true,
+ "mean_paired_deficit_to_clean_kp_at_most_0p015": true,
+ "one_sided_95pct_upper_deficit_at_most_0p025": true
+ },
+ "metrics": {
+ "accuracy_by_seed": {
+ "clean_kp": [
+ 0.9158,
+ 0.9155,
+ 0.9087,
+ 0.915,
+ 0.9144
+ ],
+ "dynamic": [
+ 0.9152,
+ 0.9158,
+ 0.9165,
+ 0.9151,
+ 0.9166
+ ]
+ },
+ "bp_50k_mac_reference": 1216531200000000,
+ "dynamic_early_alignment_by_seed": [
+ 0.9996865193049113,
+ 0.9996783634026846,
+ 0.9997114737828573,
+ 0.9996656179428101,
+ 0.9996906618277231
+ ],
+ "invariant_failures": [],
+ "mean_accuracy": {
+ "clean_kp": 0.91388,
+ "dynamic": 0.91584
+ },
+ "mean_dynamic_early_alignment": 0.9996865272521973,
+ "mean_paired_clean_kp_deficit": -0.001960000000000006,
+ "paired_clean_kp_deficit": [
+ 0.0005999999999999339,
+ -0.00029999999999996696,
+ -0.007800000000000029,
+ -9.999999999998899e-05,
+ -0.0021999999999999797
+ ],
+ "paired_deficit_one_sided_95pct_upper_bound": 0.001305694255308511,
+ "source_commit": "0008f2cd96c06bea98f43560867d93967c597711"
+ },
+ "protocol": "kp_dynamic_neutral_projection_confirmation_v1",
+ "review_score_after": 7,
+ "review_score_before": 6,
+ "score_change_rule": "only a complete untouched paired five-seed test panel can establish the strict accept bar; depth scaling remains oral work",
+ "status": "passed",
+ "test_evaluations": 10
+}