External Publication
Visit Post

Shannon Prime Lattice

Hugging Face Forums [Unofficial] June 14, 2026
Source

The paper series

A staggered set of short, independently citable, receipts-first papers — each carries its own one-command reproduction.

  • 01 — Two-ring memory — query-directed recall + byte-addressable KV offload (the needle-off-NVMe result above).
  • 02 — The reducing loader — output-preserving transcode + zero-copy load (the ~50%-smaller, bit-faithful result).
  • 03 — Frobenius calibration-free quantization (staged).
  • 04 — The Oracle & the Teacher (written) — oracle-grounded backend verification: KL 2.7e-10 port, teacher-forced decode — plus the case study where a hand-written oracle measured gemma-4’s true PPL at 4.68 and convicted the GGUF ecosystem (192–506) while exonerating llama.cpp’s forward.
  • 05 — The Probe Suite (written) — bisection, isolation and benchmark hygiene as one set — from the 12.65× phantom and the 0/256 K-quant bug to ecosystem-scale forensics and simulate-before-build (artifact matched the simulator to four decimals).
  • 06 — Computing on the Zip File (complete, citable) — the dp4a bandwidth ladder (f32 1× → int8 ~3.8× → Q4 ~7.06×), the OK_Q4B block-scaled kernel, the sovereign quantization pipeline, and the gated headline: 26.1 tok/s at PPL 5.12 on an RTX 2060 12GB.
  • 07 — The Auditable Latent Crossbar (staged draft) — a frozen 12B steered by direct KV-cache transplant, no tokens: 15/15 incorporation + 15/15 selectivity , self-transplant null 7/7 bit-identical, gold-instrument coherence (X-R1).
  • 08 — O(1) KV: a context-decoupled cache (staged draft) — a learned 512×32 LSH router decouples the KV cache O(1) from context: +0.47% PPL @8× (oracle −0.08%), O(1) VRAM (8k↔16k flat ~50 MiB) , NIAH needle survives 10/50/90% (frozen-router control misses) (X-R2).
  • 09 — KAIROS: a resident 12B daemon (staged draft — release gated on the in-flight soak) — disciplined silence + O(1) bit-exact rewind : 24-tick crucible perfect, rewind byte-identical across 48 layers, metal 0.0073 vs prefix-grow 0.924 s/action (KAIROS-01/02). The ≥24 h endurance soak is running now — no verdict from a mid-run log; the paper ships on the receipt.
  • 10 — Receipts or it didn’t happen (staged draft) — bit-exact-or-bounded as the contribution: the four documented self-corrections that prove the gates discriminate.

Discussion in the ATmosphere

Loading comments...