Note Nine preregistered runs on one frozen 30B: the extraction hypothesis died, two laws and two artifacts came out. Read this first.

Joakimpalm-Zen/Muse-Glimmer-30B-Surgical-Q4_K-GGUF

Text Generation • 28B • Updated 14 days ago • 290

Note 4.75% of the decoder removed and healed, Q4_K. Passes the bar (97.74% / 0.04810). Research artifact.

Joakimpalm-Zen/Muse-Glimmer-30B-Surgical-Q8_0-GGUF

Text Generation • 28B • Updated 14 days ago • 384 • 1

Note 6.34% removed and healed, Q8_0. Passes the bar (98.34% / 0.03883). Research artifact.

Joakimpalm-Zen/Muse-Glimmer-30B-Surgical-6p34-Q4_K-recovered-GGUF

28B • Updated 14 days ago • 295 • 1

Note The 6.34% model at Q4_K with distillation-recovered scales: inside the bar (97.90% / 0.04949) where the plain Q4_K failed.

Note Preregistration, four-arm scores, harness and HellaSwag records, training logs and code behind the recovered Q4_K.

Joakimpalm-Zen/gemma-4-31B-it-attn48-removed-Q4_0-GGUF

Text Generation • 31B • Updated 14 days ago • 415 • 1

Note One attention block physically removed (Runner 0.4.7+, CPU path). Frees KV cache; llama.cpp refuses it by name.

Joakimpalm-Zen/Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF

27B • Updated 6 days ago • 2.7k • 17

Note ISTA-DASLab's IQ3_S file with every block scale retrained against the BF16 parent. Same bytes, same layout; inside the bar.

Note Run record, training log, fidelity and HellaSwag records, frontier table and code behind the recovered IQ3_S.

Joakimpalm-Zen/Cloudflare-clef-IQ3_S-recovered-GGUF

27B • Updated 11 days ago • 1.61k

Note A 3-bit Clef backbone with two recovery adapters; the house bar (KLD 0.05, top-1 97%) not met, stated on the card.