ablations/alignment_cells.csv

cell,run_name,paradigm,test_bpb,baseline_bpb,delta_bpb,a3_overlap_pct,expert_union,num_experts,eff_experts,note
L0,g3_tmoe_s0_1e16,temporal-standard-aux,1.4753,1.4753,0.0000,29.57,158.12,192,183.87,L0 alignment reference (shipped selection)
A0-ctl,g3_moe_s0_1e16_sigmoid,fullMoE-sigmoid-aux,1.4475,1.4585,-0.0110,20.23,192.00,192,190.07,attribution control (score fn); baseline=softmax MoE 1.4585
A1,g3_moe_s0_1e16_auxfree,fullMoE-auxfree,1.4495,1.4475,+0.0020,20.68,191.56,192,176.27,"vs A0-ctl (bias ctl, neutral); vs softmax 1.4585 -0.009; H1 not supported (A3 flat 20.2->20.7)"
A2,g3_tmoe_s0_1e16_auxfree,temporal-auxfree,1.4765,1.4753,+0.0012,34.59,147.37,192,126.15,A3 +5.0pt vs L0 29.57 at BPB parity; diversity DOWN (eff 184->126 union 158->147) but no frozen-pool collapse
A3,g3_tmoe_s0_1e16_karen,temporal-auxfree-momentum,1.4722,1.4753,-0.0031,36.05,153.40,192,129.26,Karen demand-momentum b=1.0 g=0.125; A3 +1.5 vs A2 +6.5 vs L0; BPB below baseline (best rung); partial diversity recovery vs A2 (eff 126->129 union 147->153)
A2-s2,g3_tmoe_s2_1e17_auxfree,temporal-auxfree,1.3005,1.2873,+0.0132,37.96,150.92,192,128.23,attribution cell (s2@1e17 on h100): SUBSTRATE guilty — bare auxfree already eff 128 / max-res 81.5% / pin 0.2 (~= Karen) and BPB +0.013 vs plain; momentum = BPB repair not cause
A3q,g3_tmoe_s0_1e16_karen2t,temporal-auxfree-doublemom,1.4715,1.4753,-0.0038,35.82,149.39,192,126.27,double-momentum alpha=1.0 gamma_q=0.015625; GATE FAIL (eff 126.3<170 AND maxres 91.2%>40%); PASS BPB/A3/pinned=0; diversity NOT recovered vs single-mom (eff 129->126 maxres 98->91)
A3p,g3_tmoe_s0_1e16_mom,temporal-plain-momentum,1.4717,1.4753,-0.0036,28.37,162.80,192,181.82,momentum beta=2.0 g=0.125 on PLAIN substrate (h100 0046): A3 gate FAIL (28.4% ~ L0) — momentum A3 gain is a substrate artifact; diversity intact (max-res 16.3% pin 0); small genuine BPB gain
A3s,g3_tmoe_s0_1e16_karen_soft,temporal-auxfree-gentlemom,1.4703,1.4753,-0.0050,35.92,152.83,192,126.45,gentle single momentum b=0.5 g=0.5; GATE FAIL (eff 126.5<170 AND maxres 69.1%>40%); PASS BPB(best yet)/A3/pinned=0; maxres best of family (98->91->69) but eff stuck ~126 -> concentration structural not momentum-tunable
A3p-seed2,g3_tmoe_s0_1e16_mom_seed2,temporal-plain-momentum,1.4806,1.4753,+0.0053,27.93,156.90,192,183.76,SEED=2 of A3p (h100 0049): RETRACTS the A3p BPB win — seed pair 1.4717/1.4806 |d|=0.0089 (~18x nominal noise) 2-seed mean 1.4762 = TIE vs L0; momentum on plain substrate is a robust no-op (A3-inert diversity-safe BPB-neutral both seeds)
LG1,g3_tmoe_s0_1e16_momr,temporal-plain-logratio-momentum,1.4714,1.4753,-0.0039,28.63,164.60,192,182.20,"logratio momentum b=0.25 (h100 0060): A3-INERT on healthy substrate (28.6~L0) but diversity-safe (union UP 160.8->164.6, 0 pinned); BPB single-seed within family noise; gates A3 FAIL -> OVERALL FAIL"
LG2,g3_tmoe_s0_1e16_bursty,temporal-plain-bursty-loss,1.5206,1.4753,+0.0453,79.19,42.30,192,33.20,bursty window loss l=0.02 W=32 (h100 0061): catastrophic Goodhart — eff collapse 183.9->33.2 max-res 100% pinned 5/layer BPB +0.045; loss never minimized (rising-plateau anticipatory fingerprint); 4/5 gates FAIL
LG1s2,g3_tmoe_s0_1e16_momr_seed2,temporal-logratio-mom,1.4782,1.4753,+0.0029,28.75,159.88,192,183.65,logratio mom b=0.25 NO auxfree (seed2 of LG1); DIVERSITY RESTORED eff 183.65>=170 maxres 19.2%<40% pinned0 (PASSES Karen-failed gates) but A3 28.75 baseline-level (NO alignment gain); seed pair 1.4714/1.4782 mean 1.4748=PARITY w/L0; mirror-image of Karen
A3s-s2,g3_tmoe_s0_1e16_karen_soft_seed2,temporal-auxfree-gentlemom,1.4775,1.4753,+0.0022,35.10,154.10,192,128.78,seed2 of karen_soft; A3 35.10 (reproduces ~35.5 gain) eff 128.78 maxres 74.44 pinned0; seed pair 1.4703/1.4775 mean 1.4739=PARITY w/L0 (single-seed 1.4703 win was noise); maxres 69->74 (best-of-family but seed-dependent)
A1s2,g3_moe_s0_1e16_auxfree_seed2,fullMoE-auxfree,1.4558,1.4475,+0.0083,21.86,191.97,192,176.48,seed2 of A1 (full MoE auxfree NO temporal); A1 pair 1.4495/1.4558 mean 1.4526; vs sigmoid-ctl 1.4475 +0.0051 (within seed spread 0.0063) vs softmax 1.4585 -0.0059 better; eff 176.48/union 192 reproduce; substitution safe vs softmax baseline