refiner combined scorecard: baseline(v1)+refiner(v3 ReZero/byte-NLL)

refiner scorecard (GLR-1 port)

refiner results (2 pts, refiner_b4:s1)

refiner results (1 pts, refiner_b4:s0)

refiner results (1 pts, refiner_b8:s0)

refiner: ReZero gates (a_attn/a_ffn init 0) -> identity at init, stable deep weight-shared iteration (fixes eval-BPB divergence)

vr grid results (text, 3 pts, fixed:4:s0)

refiner results (2 pts, baseline:s1)

refiner results (1 pts, baseline:s0)

ablate: MODE=refiner (baseline vs Trident+refiner + test-time budget sweep)

GLR-1 refiner: PatchRefiner (weight-shared iterative patch attention) + deep-supervision/PonderNet losses; end-to-end wired, causal, tested

refiner scorecard (GLR-1 port)

refiner results (2 pts, refiner_b2:s0)

refiner results (1 pts, baseline:s0)

operator results (4 pts, learned_N8192:s0)

operator results (3 pts, learned_N4096:s0)

operator results (2 pts, learned_N1024:s0)

operator results (1 pts, baseline:s0)

vr grid results (text, 2 pts, fixed:1:s1)

ablate: MODE=operators capacity sweep + KG-05/KG-08 scorecard

Section 10: sparse executable operator memory (rank-1 pool, top-k router, KG-05 controls)

vr grid results (text, 1 pts, fixed:1:s0)

0.5.6: one-patch read-latency finding + recall probe fix

recall fix: answer one patch after query (1-patch read latency); verified read exposure

KG-04 runner: dense recall packing (write-capacity stress) + factor_emb/local_attn env

test: factor symmetry-break

audit 0.5.6: mechanism self-audit + KG-04 fixes

KG-04 fixes: orthogonal factor_emb (de-redundancy) + optional local attn

vr grid results (text, 6 pts, fixed:2:s2)

vr grid results (text, 5 pts, fixed:2:s1)

vr grid results (text, 4 pts, fixed:2:s0)

vr grid results (text, 3 pts, fixed:1:s2)

vr grid results (text, 2 pts, fixed:1:s1)

vr grid results (text, 1 pts, fixed:1:s0)

mixed-R kernel: exact vectorized patch-expansion chunk scan for R>1/adaptive (2.4x CPU, far more on GPU)

KG-04: cost-aware rate objective + honest cost accounting (mean_factors/active_R)

checkpoint ckpt_r150_trident/latest @ step 500

baseline ckpt_r150_xf/final @ step 1500

baseline ckpt_r150_xf/latest @ step 1500

baseline ckpt_r150_xf/latest @ step 1000

head-to-head scorecard (headtohead_fair_longctx.json)

baseline checkpoints_baseline_longctx/final @ step 1500

baseline checkpoints_baseline_longctx/latest @ step 1500

baseline checkpoints_baseline_longctx/latest @ step 1000

baseline ckpt_r150_xf/latest @ step 500

cpu inference benchmark (cpu_bench_optimized.json)

Upload scripts/cpu_bench.py with huggingface_hub

Upload src/trident/decoder.py with huggingface_hub

Upload src/trident/byte_lane.py with huggingface_hub

Upload src/trident/layers.py with huggingface_hub