| # Overlap-arch 1e17-g3-s2 SCREENS (orch 0150 step 3). Standard g3-s2-1e17 config (16k G3-era corpus, | |
| # cosine, gb256/mb64/lr3e-3/3861it/seed1234, GRAIN=3=192exp/top18/moe_ffn58/shared352, CE_FUSION). | |
| # BPB = test_CE_nats / 2.7600 (byte-derived, 3.9819 bytes/tok; the G3-era 16k corpus divisor, NOT the | |
| # original-bpe-16k 2.7568). LOWER better. Standard-arch comparators (same corpus/divisor): temporal 1.2873, moe 1.2708. | |
| # GATE (0150 step4): variant promotes iff temporal_BPB <= 1.2873+0.005=1.2923 AND moe_BPB <= 1.2708+0.010=1.2808. | |
| # UNTESTED/NOT-BUILT variant (per orch 0155): V1 could alternatively route on LN2(x) (same norm as experts, isolates | |
| # earliness from the norm) instead of LN1(x); not implemented - V2 already spans the LN2(x)-reading axis. | |
| variant,cell,method,test_ce_nats,test_bpb,std_bpb,delta,bar,gate | |
| v1_early_router,g3_tmoe_s2_1e17_ovlEarly,temporal,3.562969,1.2909,1.2873,+0.0036,0.005,PASS | |
| v1_early_router,g3_moe_s2_1e17_ovlEarly,moe,3.508395,1.2712,1.2708,+0.0004,0.010,PASS | |
| v2_parallel_ffn,g3_tmoe_s2_1e17_ovlParFFN,temporal,3.625855,1.3137,1.2873,+0.0264,0.005,FAIL | |
| v2_parallel_ffn,g3_moe_s2_1e17_ovlParFFN,moe,3.560735,1.2901,1.2708,+0.0193,0.010,FAIL | |
| # VERDICT: V1 early-router PROMOTES (temporal +0.0036<=0.005 AND moe +0.0004<=0.010) -> 1e18-g3. | |
| # V2 parallel-ffn FAILS both (temporal +0.0264, moe +0.0193). Routing one sublayer earlier is | |
| # nearly free; the full PaLM-parallel restructure costs too much quality at this budget. | |