{ "clock": "virtual-v1", "deltas_sha256": "f7fdba634b93c9c7bfa3296cb7382a57ed325657279c3ddb2aca560053aa8840", "inputs_sha256": { "experience.npz": "5542c80820e9b94e2e1c6e89f5364254fd27774b1ffdc3ebb8d82a87c27780b8", "forward.py": "2caf44541cb5a365d4475dc05157c7fff20928450e7cecd3c3c9020d16915f5f", "gated_4stage_f499.py": "b7def56c59cfc2fe34fb853fffc203441faadb450ce6aa64c9018dce94659866", "layout.json": "f0d9d3b5725bdaaefc1a5197eb346abc752a1704221d4947c74959e65206e016", "params.npz": "c5e0cd4319c3d5788b61fe0db767c0f423810588cd9106154f4d9e7cff749320" }, "program_file": "gated_4stage_f499.py", "program_sha256": "b7def56c59cfc2fe34fb853fffc203441faadb450ce6aa64c9018dce94659866", "run": { "duration_s": 122.855, "exit_code": 0, "result": { "clock": "virtual-v1", "clock_reads": 0, "hypothesis": "The new condition is a global contrast reduction (per-image, per-channel pixel deviations scaled by c~0.16, c estimated from the experience). Stage 1: in block 0, repurposed least-influential MLP units add a gated rank-20 linear correction. Their gate is the low-contrast-vs-other mean-difference direction of block-0 MLP inputs, and the correction is a label-free map from MLP input to the clean-minus-low-contrast block-0 output difference, fitted on synthetic contrast-reduced copies of the clean experience images. Its gate misses low-contrast images with saturated mean colours, whose normalised tokens look high-energy. Stage 2: in block 1, after stage 1 is applied, a second set of repurposed units uses a gate fitted on exactly the low-contrast tokens stage 1 left closed (vs clean and noisy tokens; threshold at the clean 98th percentile). Its rank-12 map, fitted on those tokens, corrects them. Stages 3 and 4 do the same in blocks 2 and 3 (rank 8) for the tokens all earlier gates left closed. All stages use 0.499 of the Frobenius budget for sharper gates.", "hypothesis_type": "str", "keys": [ "classifier.blocks.0.linear1.weight", "classifier.blocks.0.linear1.bias", "classifier.blocks.0.linear2.weight", "classifier.blocks.0.linear2.bias", "classifier.blocks.1.linear1.weight", "classifier.blocks.1.linear1.bias", "classifier.blocks.1.linear2.weight", "classifier.blocks.1.linear2.bias", "classifier.blocks.2.linear1.weight", "classifier.blocks.2.linear1.bias", "classifier.blocks.2.linear2.weight", "classifier.blocks.2.linear2.bias", "classifier.blocks.3.linear1.weight", "classifier.blocks.3.linear1.bias", "classifier.blocks.3.linear2.weight", "classifier.blocks.3.linear2.bias" ], "predicted_effect": "Same four-stage design as the checked version (new +587, old -5, prev +1), using 0.499 instead of 0.49 of the Frobenius budget per edited tensor, so every gate is slightly sharper (larger gate weight, smaller per-direction weight). Earlier budget increases helped (0.46 -> 0.49 raised the single-stage score from 472 to 494); this step is smaller. On the experience, predictions are nearly identical to the checked four-stage version. Expected: score_new about +580..+610, score_old about -10..0, score_prev within about +-10.", "predicted_effect_type": "str", "status": "ok", "update_s": 122.64758389900089 }, "status": "ok", "stderr": "", "stdout": "estimated contrast factor 0.1637\nstage 0 gate: low-contrast median 5.005 clean q 0.95 3.451 theta 3.451\nstage 0 units 22 lam 0.691 beta 1.588 rel frob linear1 0.499 linear2 0.499\nstage 0 tokens left uncorrected: syn 1594 new 689\nstage 1 gate: low-contrast median 4.237 clean q 0.98 3.621 theta 3.621\nstage 1 units 14 lam 0.5588 beta 2.7254 rel frob linear1 0.499 linear2 0.499\nstage 1 tokens left uncorrected: syn 554 new 211\nstage 2 gate: low-contrast median 4.568 clean q 0.98 3.589 theta 3.589\nstage 2 units 10 lam 0.5072 beta 3.7359 rel frob linear1 0.499 linear2 0.499\nstage 2 tokens left uncorrected: syn 168 new 67\nstage 3 gate: low-contrast median 2.01 clean q 0.98 1.209 theta 2.011\nstage 3 units 10 lam 0.4891 beta 3.5216 rel frob linear1 0.499 linear2 0.499\nstage 3 tokens left uncorrected: syn 91 new 27\n" }, "sandbox": { "bwrap": "/usr/bin/bwrap", "driver_sha256": "3c154552687c8b84c00eb85ada4bea68838fa0f1526949c8650092c42f9feabe", "file_size_limit_bytes": 268435456, "max_output_chars": 30000, "memory_bytes": 8589934592, "python": "/home/boom/BoomLi/WeightRSI-nsess/.venv-sandbox/bin/python", "threads": 4, "tmpfs_bytes": 1073741824 }, "shown": { "duration_s": 122.855, "program_stderr": "", "program_stdout": "estimated contrast factor 0.1637\nstage 0 gate: low-contrast median 5.005 clean q 0.95 3.451 theta 3.451\nstage 0 units 22 lam 0.691 beta 1.588 rel frob linear1 0.499 linear2 0.499\nstage 0 tokens left uncorrected: syn 1594 new 689\nstage 1 gate: low-contrast median 4.237 clean q 0.98 3.621 theta 3.621\nstage 1 units 14 lam 0.5588 beta 2.7254 rel frob linear1 0.499 linear2 0.499\nstage 1 tokens left uncorrected: syn 554 new 211\nstage 2 gate: low-contrast median 4.568 clean q 0.98 3.589 theta 3.589\nstage 2 units 10 lam 0.5072 beta 3.7359 rel frob linear1 0.499 linear2 0.499\nstage 2 tokens left uncorrected: syn 168 new 67\nstage 3 gate: low-contrast median 2.01 clean q 0.98 1.209 theta 2.011\nstage 3 units 10 lam 0.4891 beta 3.5216 rel frob linear1 0.499 linear2 0.499\nstage 3 tokens left uncorrected: syn 91 new 27\n" }, "tensors": [ "classifier.blocks.0.linear1.weight", "classifier.blocks.0.linear1.bias", "classifier.blocks.0.linear2.weight", "classifier.blocks.0.linear2.bias", "classifier.blocks.1.linear1.weight", "classifier.blocks.1.linear1.bias", "classifier.blocks.1.linear2.weight", "classifier.blocks.1.linear2.bias", "classifier.blocks.2.linear1.weight", "classifier.blocks.2.linear1.bias", "classifier.blocks.2.linear2.weight", "classifier.blocks.2.linear2.bias", "classifier.blocks.3.linear1.weight", "classifier.blocks.3.linear1.bias", "classifier.blocks.3.linear2.weight", "classifier.blocks.3.linear2.bias" ], "wall_s": 300.0 }