Boom5426's picture
WeightCode records and weights (redacted 2026-11-10, see REDACTIONS.json)
4c27541
|
Raw History Blame Contribute Delete
45.7 kB
Claude Code trace
Assistant claude-opus-5-5 2026-11-03T13:30:53.937Z 10,265↓ 54↑ (33,249 cached)
5 tool calls (mcp__wrsi__python)
Assistant claude-opus-5-5 2026-11-03T13:31:22.725Z

Checking the patch on the held-out set: it ties fc1 weights across the colour channels at strength 0.72, which keeps the edit under the size limit.

Assistant claude-opus-5-5 2026-11-03T13:31:22.906Z 771↓ 108↑ (20,964 cached)
2 tool calls (mcp__wrsi__patch_check, mcp__wrsi__declare_final)
Assistant claude-opus-5-5 2026-11-03T13:31:33.991Z 197↓ 494↑ (11,022 cached)

I've checked and declared one patch, tie72.json. On the held-out check set it raised accuracy from 57.6% to 72.4% (S = 590: 807 fixed, 217 broken, out of 4,000 images).

Why the model fails: the deployment images are coloured digits, with the digit appearing in a varying mix of the red, green and blue channels. The first layer's (fc1's) weights differ a lot between channels: the green and blue channel weights correlate only about 0.32–0.35 with the red ones. So the model's answer depends on what colour the digit is, not just its shape. On the random evidence examples it got only 48% right.

The edit: I pulled each fc1 weight towards the average of its three colour channels at the same pixel, so the first layer responds to the digit's shape whatever its colour. This is one tie op over 2,352 numbers, about 0.7% of all parameters.

Why it's only a partial tie: averaging the channels fully would have changed fc1 by 67% of its size, over the 50% limit. I used strength 0.72 instead, which changes it by 48.5%. That costs accuracy: on the random evidence examples the full average scored 85%, but the allowed version scored 77%. Both versions fixed 9 of the 16 known failures.

Trade-off: the patch breaks 217 check-set images the original model got right. The model left in place is the checked patch with the highest S, which is this one. I used 1 of the 8 checks and didn't try other variants.

claude-opus-5-5 Input 11.2k↓ Input Cached 65.2k↓ (85.3%) Output 656↑