From 203e75987d02cdb8c317e3302e1f05c9a4beb0e3 Mon Sep 17 00:00:00 2001 From: YurenHao0426 Date: Wed, 22 Jul 2026 13:43:05 -0500 Subject: results: close normalized mirror short accuracy gate --- results/mirror_short_gate.json | 71 ++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 71 insertions(+) create mode 100644 results/mirror_short_gate.json (limited to 'results/mirror_short_gate.json') diff --git a/results/mirror_short_gate.json b/results/mirror_short_gate.json new file mode 100644 index 0000000..065a1f2 --- /dev/null +++ b/results/mirror_short_gate.json @@ -0,0 +1,71 @@ +{ + "checks": { + "accuracy_at_least_0.65": false, + "both_records_finite": true, + "candidate_selected": true, + "early_alignment_at_least_0.30": true, + "macs_at_most_1.15x_bp": true, + "within_10_points_of_bp": false, + "zero_task_loss_queries": true + }, + "confirmation_test_seeds_touched": false, + "matched_bp": { + "accuracy": 0.7494, + "total_macs": 24330624000000 + }, + "protocol": "normalized_response_mirror_short_v1", + "review_score_after": 5, + "review_score_before": 5, + "rows": [ + { + "accuracy": 0.6268, + "all_layer_alignment": 0.8258112323911566, + "early_third_alignment": 0.8274043500423431, + "finite": true, + "logical_batch_loss_queries": 0, + "loss": 1.0569130615234374, + "lr": 0.03, + "mean_feedback_forward_cosine": 0.9360259049817136, + "mirror_events": 119, + "path": "results/mirror_short/wm_lr0.03.json", + "peak_memory_allocated_bytes": 1730831872, + "source_commit": "dba3eb29d5797c1b74d7e1e781523afad9faeb8e", + "total_macs": 24251852087296, + "wall_s": 90.82122445106506 + }, + { + "accuracy": 0.6404, + "all_layer_alignment": 0.9162216186523438, + "early_third_alignment": 0.9392559031645457, + "finite": true, + "logical_batch_loss_queries": 0, + "loss": 1.0210670532226562, + "lr": 0.1, + "mean_feedback_forward_cosine": 0.9359531402587891, + "mirror_events": 119, + "path": "results/mirror_short/wm_lr0.1.json", + "peak_memory_allocated_bytes": 1730831872, + "source_commit": "dba3eb29d5797c1b74d7e1e781523afad9faeb8e", + "total_macs": 24251852087296, + "wall_s": 90.53707838058472 + } + ], + "score_change_rule": "inherited short baseline cannot raise score", + "selected_wm": { + "accuracy": 0.6404, + "all_layer_alignment": 0.9162216186523438, + "early_third_alignment": 0.9392559031645457, + "finite": true, + "logical_batch_loss_queries": 0, + "loss": 1.0210670532226562, + "lr": 0.1, + "mean_feedback_forward_cosine": 0.9359531402587891, + "mirror_events": 119, + "path": "results/mirror_short/wm_lr0.1.json", + "peak_memory_allocated_bytes": 1730831872, + "source_commit": "dba3eb29d5797c1b74d7e1e781523afad9faeb8e", + "total_macs": 24251852087296, + "wall_s": 90.53707838058472 + }, + "status": "failed" +} -- cgit v1.2.3