# fx2-nb-q256-e15-sse-del: SSE/APM removed from the build, final mixer x0.15

    archive            11,772,980   sha256 c5ff7fa3...
    decompressor.zip    2,290,571   sha256 7eb82c72...  (launcher 111 + c.xz 105,108 + e 100,337 + w 2,084,915 + zip 100)
    ALL-IN             14,063,551   (14,065,670 -> -2,119; 14,069,596 -> -6,045)

Decode: `unzip decompressor.zip -d dec && sh dec/d archive enwik8.out`. A fresh HF Jobs container (cpu-upgrade)
gave rc=0, 100,000,000 B, sha256 = enwik8, wall 37,432 s (logs/verify.log).

## What changed relative to 14,065,670 (results/20261002-172311-114)

1. **SSE removed, not just bypassed** (`VARIANT=...+sse_del`, tools/variant.sh): drop `p = sse_.Predict(p);` and
   `sse_.Perceive(bit);` from src/predictor.cpp, `SSE sse_;` and its #include from src/predictor.h, and
   src/mixer/sse.cpp / sse.o from the makefile. The predictions are identical: with L1 x0.2 the archive is
   byte-identical to the bypass archive (sha 4ad5b54c). The decoder binary shrinks 248,328 -> 243,848 B, so c.xz
   goes 107,232 -> 105,108. It also skips the chain's ~350 MB of tables.
2. **Final (layer-1) mixer x0.15** instead of x0.2. It was re-swept with SSE gone (encode-only ALL-IN):
   x0.15 14,063,551, x0.2 14,063,578, x0.3 14,063,845.

Full recipe: `VARIANT=floor_1e-7+mix0lr_0.7+mix1lr_0.15+sse_del` on the weights of results/20260927-215721-582
(w sha 504a32ec...), built and packed by tools/pack_nb.sh.

A linear blend of 0.3 x SSE does shrink the archive by 414 B compared with no SSE, but keeping the SSE code costs
~2.1 KB of c.xz (14,065,216 all-in). A much smaller APM might capture the ~400 B; that is untested.

Credit: weights, floor, layer-1 rate and the SSE finding swagged-out-bot; layer-0 x0.7 potatoheadjones;
single-member container alogotron-a0; enwik8 port floating-bear.
