Shannon Prime Lattice
Hugging Face Forums [Unofficial]
June 17, 2026
The paper series Update
A staggered set of short, independently citable, receipts-first papers — each carries its own one-command reproduction.
- 12 — The Memo Curator (written, citable — X-C2) — autonomous discrete recall above the crossbar: the loop drives the closed substrate on its own, inert when off (PPL 4.6665 bit-identical), addresses memories with a 256-bit LSH / integer-Hamming key (reduction-order-immune; sign-binarize collapses at r=32, ship r=256), and promotes the matched recall / discards the corrupted one (+0.000% / +40106% safety valve).
- 13 — Speculate and Undo (written, citable — X-222) — O(1) bit-exact rewind of latent memory: replay into the resident cache is load-bearing, the rewind resets the prefix byte-identical (12B + E2B; and across a sliding-window-ring wrap) — the §4-trap guarantee made mechanical, so the curator can speculate and undo for free.
- 14 — A parameter-free neocortex (written, citable — X-R3VSA) — VSA/HRR Ring-3 consolidation from discrete NTT, zero training : retrieve-and-verify (the P2.b top-5 verdict honored), with the consolidation loss a step function (hit lossless / miss caught by the 2% gate) and the idle GC demoting 349.8 MB resident KV → a 16.3 KB index.
- 15 — The organism breathes (written, citable — X-ORG) — real audio to episodic memory: real speech → the EAR on physical Intel GNA 2.0 (KAIROS-04) → the 12B pivots 7/8 → the audio-conditioned KV state becomes a curator-indexable, replayable Ring-2 episode whose signature separates cleanly from text memories (self 211/256, margin +79).
See SERIES.md for the manifest and release cadence, LEDGER.md for the master claims ledger (every number traced to a command), and METHODOLOGY.md for the gate vocabulary and the “no number without a command” discipline.
The system: a four-tier memory hierarchy plus a latent crossbar
The original “two-ring” framing has grown into a four-tier hierarchy with an inter-model lane on top. Architecture ground truth lives in the lattice repo (papers/RFC-XBAR-auditable-latent-crossbar.md); this is the public map, each component tagged with its status.
┌────────────────────────── VRAM (owned arena) ───────────────────────────┐
│ │
│ Exec (generator, e.g. Memo (small curator, │
│ gemma-4-12B OK_Q4B) frozen-small) │
│ causal forward, generates non-causal pass over the episode │
│ │ ▲ │ ▲ │
│ ▼ write │ attend ▼ propose │ read │
│ ┌─ Ring 1 ─┐ ┌── Ring 2 (hippocampus) ┐ ┌─ Ring 2′ (shadow) ─┐ │
│ │ working │ │ verbatim Spinor KV, │◄─│ Memo's proposals │ │
│ │ KV │ │ recent + bounded │ │ promote-on-accept │ │
│ └──────────┘ └────────────────────────┘ └─────────┬──────────┘ │
│ ▲ recall from BOTH │ promote (gated) │
│ │ ┌── Ring 3 (neocortex) ───┐◄──┘ │
│ └────────────────│ adapter pseudo-tokens, │ G-R3-LOSS bounded │
│ │ consolidated long-term │ (irreversible) │
│ └─────────────────────────┘ │
│ modality lanes (one CRT prime per modality): │
│ audio adapter, video, ... │
└─────────────────────────────────────────────────────────────────────────┘
Ring 2′ promotions: coherence/PPL delta → accept or REWIND (transient, reversible).
Ring 3 promotions: G-R3-LOSS bounded BEFORE source eviction (permanent, irreversible).
The four tiers
| Tier | Substrate | Representation | Lifetime | Biological analogue | Status |
|---|---|---|---|---|---|
| Ring 1 | RAM/VRAM working window | verbatim KV, full attention | the live turn | sensory / working memory | [PROVEN] — the stock model path; everything else is bit-exact-when-off relative to it |
| Ring 2 | byte-addressable storage (Optane validated), raw episodic store | verbatim Spinor KV blocks | recent episode (bounded) | hippocampus — recent, detailed, lossless | [PROVEN] — needle off physical NVMe, poison-gated, 7.57 µs/read (01-R3/R4); bounded on purpose: the composed 32k recall at a 64× selection budget MISSed (01-R9) |
| Ring 2′ (shadow) | transient staging copy | proposals awaiting the gate | one consolidation pass | (no analogue — it is the audit mechanism) | [WIRED] — the C1-lite curator: clone → propose → gate → atomic promote / rewind, exercised on real recall, every promotion receipted |
| Ring 3 | consolidated long-term store | adapter-compressed pseudo-tokens (n→k gist) | long-term | neocortex — old, dense, semantic | [DESIGN] — under the irreversible-aware G-R3-LOSS gate: consolidation loss is quantified and bounded before the raw source is evicted; un-compressible episodes stay verbatim in Ring 2 (a valid, logged outcome) |
| Latent crossbar probe: a 12B steered by direct KV-cache transplant, no tokens | 15/15 incorporation, 15/15 selectivity (2×2 double dissociation), max single-token rank pull 3.69 orders | gated + citable (X-R1). Coherence held under the gold instrument (steered-text PPL 1.70–4.10 vs gold 4.68); self-transplant null bit-identical 7/7; raw KV splice is a deliberately blunt instrument — the learned-adapter phase exists to refine it |
|---|---|---|
| O(1) KV decoupled from context on a 12B, needle retained | learned 512×32 LSH router at +0.47% PPL @8× (oracle −0.08%, frozen +4.17%); O(1) VRAM: 8k↔16k flat within ~50 MiB ; NIAH needle survives 10/50/90% depth | gated + citable (X-R2). The KV term is O(1); the absolute footprint in this backend-direct harness still carries the resident model. Frozen-router negative control MISSES — isolates the learned projection as the cause |
| Resident 12B daemon: disciplined silence + O(1) bit-exact rewind | 24-tick crucible perfect (21/21 NO_OP, 3/3 ACTION, 0 false / 0 missed / 0 drift; 0.6B control collapses); rewind byte-identical across all 48 layers (diffs=0); metal 0.0073 vs prefix-grow 0.924 s/action (127×) | gated + citable (KAIROS-01 / KAIROS-02). Scripted 24-event tape, not live sensors. The ≥24 h endurance soak is IN-FLIGHT (no verdict from a mid-run log) — paper 09’s release waits on its receipt |
| The crossbar writes : a stored episode replayed into a 12B’s cache, bit-exact + load-bearing, without breaking perplexity | replay intact == baseline (diffs=0); replay zeroed diverges 12/12 ; replaying a foreign episode deflects PPL +1.38% (< the 2% gate) — proven on the 12B and the smaller E2B | gated + citable (X-R3). The deflection is over n=42 positions, a single chunk — deterministic (replay, not sampling), but a larger-N run is the named hardening lever before any headline |
| The model gains a real-world audio sense on physical silicon | real speech → 12B pivots 7/8 (the 1 miss is a conservative ACTION→NO_OP); the front-end runs on a physical Intel GNA 2.0 accelerator at 0.877 token-recovery == software-emulation == FP32 (a naive int16 sheared it to 0.667; calibrated int16 PTQ fully recovers) | gated + citable (KAIROS-04). Scripted event set, one model / one card / one GNA part; the audio “EAR” is a separate-but-related sibling of the latent-memory crossbar, sharing the same KV-inject seam |
| The Memo curator drives the crossbar autonomously | inert when off (PPL 4.6665 bit-identical); a 256-bit LSH / integer-Hamming address (reduction-order-immune); promote matched recall +0.000% / discard corrupted +40106% | gated + citable (X-C2). The float→discrete course-correction (sign-binarize collapses at r=32, recovers ≥128, ship r=256). 2-episode registry vs synthetic noise; deflection single-chunk; Ring-2 verbatim recall only |
| O(1) bit-exact rewind of latent memory | replay into the resident cache load-bearing (zeroed reads back all-zero); rewind resets the prefix byte-identical (layer-diffs=0) on → 12B <->E2B ← , and across a sliding-window-ring wrap | gated + citable (X-222). The O(1) is in the byte-count (proven by byte comparison); the latency slope is KAIROS-02/03; persistent-ABI scorer port is a follow-on |
| Parameter-free Ring-3 consolidation (VSA/HRR, zero training) | superposition recall@1=1.0 to N=32; consolidation loss a step function (hit +0.000% / miss +8.04% caught by the 2% gate); idle GC demotes 349.8 MB resident KV → a 16.3 KB index | gated + citable (X-R3VSA). Retrieve-and-verify (P2.b top-5 honored). VSA retrieve is host-numpy (the Z_q/NTT engine port is the named follow-on); Path B (trained adapter) budget-gated, untouched |
| The organism breathes : real audio → episodic memory | audio-conditioned KV serialized as a canonical Ring-2 episode (11,206,656 B); signature separates (self 211/256, margin +79); round-trip clean (RT_EXIT=0) | gated + citable, step 1 (X-ORG). The +1989% deflection is foreign-by-design (cross-context reject signal), not an audio-recall quality claim; the full audio-cue→recall loop is the open step; one model / one card / one GNA part |
Honest scope: this is a proof-of-mechanism, not a scaling study and not yet independently reproduced — on a single dev host (RTX 2060, 12 GB).
Discussion in the ATmosphere