125 MB

Ctrl+K

1 contributor

Campaign B v2 (InfoNCE mix) both seeds PASS; retro-erank corrected after overwrite catch (HF copies were the true v1): spread v1 180 > v2 139 (baseline 239) — target granularity governs spread, not loss family

1ff50cb verified about 17 hours ago

campaign_a
frame-A campaign: A1 front-fusion PASSED (0.87-0.91 both variants/seeds, bit-exact off), A2 consumption REFUTED (0.009/0.010 vs 0.15 line, controls 0.000/0.0003), A3 gated; all arms incl. controls shipped about 19 hours ago
campaign_b
Campaign B v2 (InfoNCE mix) both seeds PASS; retro-erank corrected after overwrite catch (HF copies were the true v1): spread v1 180 > v2 139 (baseline 239) — target granularity governs spread, not loss family about 17 hours ago
entry_g_2b.npz

1.05 MB
xet

frame prototype v0.1: byte-state arm beats index twin (digits 0.275 vs 0.045, twin has only 200 keys for 999 states); RSA says the codebook kept C ancestry; entry channel real vs failing random control about 22 hours ago
entry_ledger_2b.json

511 Bytes
frame prototype v0.1: byte-state arm beats index twin (digits 0.275 vs 0.045, twin has only 200 keys for 999 states); RSA says the codebook kept C ancestry; entry channel real vs failing random control about 22 hours ago
frame_anchors_v01.npz

4.39 MB
xet

frame prototype v0.1: byte-state arm beats index twin (digits 0.275 vs 0.045, twin has only 200 keys for 999 states); RSA says the codebook kept C ancestry; entry channel real vs failing random control about 22 hours ago
frame_ledger_v01.json

2.26 kB
frame prototype v0.1: byte-state arm beats index twin (digits 0.275 vs 0.045, twin has only 200 keys for 999 states); RSA says the codebook kept C ancestry; entry channel real vs failing random control about 22 hours ago
phase0_meta.json

174 Bytes
frame prototype v0.1: byte-state arm beats index twin (digits 0.275 vs 0.045, twin has only 200 keys for 999 states); RSA says the codebook kept C ancestry; entry channel real vs failing random control about 22 hours ago
proto_entry_2b.py

9.47 kB
frame prototype v0.1: byte-state arm beats index twin (digits 0.275 vs 0.045, twin has only 200 keys for 999 states); RSA says the codebook kept C ancestry; entry channel real vs failing random control about 22 hours ago
proto_phase0.py

8.53 kB
frame prototype v0.1: byte-state arm beats index twin (digits 0.275 vs 0.045, twin has only 200 keys for 999 states); RSA says the codebook kept C ancestry; entry channel real vs failing random control about 22 hours ago
proto_train_frame.py

13.2 kB
frame prototype v0.1: byte-state arm beats index twin (digits 0.275 vs 0.045, twin has only 200 keys for 999 states); RSA says the codebook kept C ancestry; entry channel real vs failing random control about 22 hours ago
proto_v02_entry.py

7.56 kB
v0.2: sweep retracts the fusion-gap reading (storage-position asymmetry: bert front-loads, t5 back-loads; student concat 0.954 > teacher); sequence entry: all 999 states, control still fails, embedding table byte-arbitrary about 21 hours ago
proto_v02_extract.py

3.22 kB
v0.2: sweep retracts the fusion-gap reading (storage-position asymmetry: bert front-loads, t5 back-loads; student concat 0.954 > teacher); sequence entry: all 999 states, control still fails, embedding table byte-arbitrary about 21 hours ago
proto_v02_gauge.py

4.01 kB
v0.2: sweep retracts the fusion-gap reading (storage-position asymmetry: bert front-loads, t5 back-loads; student concat 0.954 > teacher); sequence entry: all 999 states, control still fails, embedding table byte-arbitrary about 21 hours ago
t5_walk_of_C.json

108 kB
frame prototype v0.1: byte-state arm beats index twin (digits 0.275 vs 0.045, twin has only 200 keys for 999 states); RSA says the codebook kept C ancestry; entry channel real vs failing random control about 22 hours ago
v02_entry_ledger.json

545 Bytes
v0.2: sweep retracts the fusion-gap reading (storage-position asymmetry: bert front-loads, t5 back-loads; student concat 0.954 > teacher); sequence entry: all 999 states, control still fails, embedding table byte-arbitrary about 21 hours ago
v02_sweep_ledger.json

2.92 kB
v0.2: sweep retracts the fusion-gap reading (storage-position asymmetry: bert front-loads, t5 back-loads; student concat 0.954 > teacher); sequence entry: all 999 states, control still fails, embedding table byte-arbitrary about 21 hours ago