Boom5426's picture
WeightCode records and weights (redacted 2026-11-10, see REDACTIONS.json)
4c27541
Raw History Blame Contribute Delete
3.51 kB
{
"schema": "wrsi.patch.v0",
"parent_sha256": "e40c20786ae91c904cfd7a195addb7e03b6601b955d731518da1ff54bc01c251",
"ops": [
{
"op": "add_dense",
"tensor": "classifier.blocks.0.linear1.weight",
"delta": "delta_009.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.0.linear1.bias",
"delta": "delta_010.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.0.linear2.weight",
"delta": "delta_013.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.0.linear2.bias",
"delta": "delta_014.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.1.linear1.weight",
"delta": "delta_020.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.1.linear1.bias",
"delta": "delta_021.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.1.linear2.weight",
"delta": "delta_024.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.1.linear2.bias",
"delta": "delta_025.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.2.linear1.weight",
"delta": "delta_031.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.2.linear1.bias",
"delta": "delta_032.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.2.linear2.weight",
"delta": "delta_035.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.2.linear2.bias",
"delta": "delta_036.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.3.linear1.weight",
"delta": "delta_042.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.3.linear1.bias",
"delta": "delta_043.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.3.linear2.weight",
"delta": "delta_046.npy"
},
{
"op": "add_dense",
"tensor": "classifier.blocks.3.linear2.bias",
"delta": "delta_047.npy"
}
],
"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.",
"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."
}