Position Is Arithmetic

A number-theoretic architecture for transformer inference and memory — where a token's position, index, and routing are exact arithmetic, not floating-point metadata about it.

Live site: https://nihilistau.github.io/Position_Is_Arithmetic/

The main project is located at Shannon-Prime-Lattice — you can join the Discord here: Shannon-Prime-Lattice-Discord.

---

What this is

Position Is Arithmetic is the public research home of the Shannon-Prime project: a ground-up re-derivation of the transformer forward pass in discrete integer arithmetic, plus a memory architecture (PPT-ARM) that attaches to a frozen, pretrained transformer and gives it long, auditable memory on commodity hardware.

The thesis in one line: a transformer's positions, indices, and routing are arithmetic objects — primes, residues, lattices — so they can be computed exactly instead of approximated in floating point. That turns operations that are normally lossy (KV-cache compression, quantization, weight offload, and now inter-model memory) into operations that are bit-exact when disabled, gated when enabled, and receipted always.

This repository holds the receipts-first paper series and the project's document history. Active code lives in the linked repositories; the headline implementation is Shannon-Prime-Lattice.

Status labels used throughout (and in LEDGER.md):

• [PROVEN] — measured and gated; the number has a ledger row and a command. Citable.
• [WIRED] — implemented and gated in-engine/in-core; running today, not yet a public citable row.
• [DESIGN] — specified, with its falsification gates pre-stated; not built.

Measured results (citable)

Receipts-first: every number reproduces from a single command and traces to a row in LEDGER.md. The unflattering numbers are kept attached on purpose.

Result	Number	Scope / caveat
Resident KV-cache shrink @ 32k context	910× (7.5 GB → 8.3 MB)	two-ring offload to byte-addressable storage (01-R5)
Needle retrieved off a physical NVMe drive	HIT at 512 positions (7.57 µs/read)	poison-gated; latency figure is Optane-specific (01-R3/R4). At 32k the composed run completed but MISSed (B=512 = a 64× selection budget, far past the gated 2×–8× regime; 01-R9) — kept here on purpose
KV sparsification quality	8× at +0.69% perplexity	one corpus, 2k context (2× and 4× go negative) (01-R1)
Reducing loader (transcode)	model → ~50% smaller, bit-faithful forward	gemma-3 + Qwen3, closure-gated (paper 02)
Bit-exact when disabled	argmax-identical to the stock model	the invariant under everything (01-R8)
12B GPU decode + quality, same RTX 2060 12GB	26.1 tok/s at wikitext PPL 5.12 (graph EXACT, 256/256 top-1, 24/24 gates)	gated + citable (06-R10). llama.cpp-CUDA: 31.29 tok/s at PPL 192–506 — every gemma-4 GGUF measurable in June 2026, incl. the post-fix rebuilds, carries broken weights (06-R8). SP engine bandwidth 245 vs 207 GB/s (+18%); the earlier 34.2 (+9.3%) headline is retired — its artifact failed the PPL gate (the series' own rule caught it)
The gemma-4 ecosystem finding	true full-precision PPL 4.68 (hand-written reference forward) vs GGUFs 192–506	engine-independent conviction (06-R8); verification + fix tutorial: GEMMA4-QUANT-FIX.md
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
The crossbar's bind re-carried onto the exact-integer O_K substrate	the floating-point Ring-3 superposition is now 256/256 bit-identical to the native integer path and reduction-order-immune (integer M byte-identical across 8 summation orders; float drifts 4.44e-15); live loop on native sp_pr_mul, CAP=32 unregressed	gated + citable (X-OK-BIND). The first negative: a number-theoretic χ_d carrier lowers coherence but is inert — the integer container is the win, not the integer carrier. One model / one card
Integer Frobenius episode codec, and the boundary of the algebra	rank-2 O_K codec sub-ULP at 24 bits (relL2 1.2e-7, 0.76× store); losslessness = reconstruction fidelity, not a fake +0.000% (the n=42 PPL gate is blind below ~1%, reported as such)	gated + citable (X-OK-FROB). Three boundary negatives kept on the record (entropy 1.02×, Möbius sheds memories, T2-on-weights ≈ random). O_K wins on exact arithmetic, never on structuring high-entropy content. One model / one card
The organism breathes (full loop): cross-modal continuous→discrete→continuous on silicon	audio cue → discrete sig (audio 256 / decoys 147,129) → integer bind → retrieve top-1 (cos +0.47) → C2 cross-modal verify (accept audio / reject text) → integer Frobenius land (relL2 ~9e-8) → metal SP_REPLAY into the live 12B (checks=5, fails=0)	gated + citable (X-OK-ORG). The full loop is closed; matched-context audio recall is not claimed (the 88.89 PPL is the cross-context reject signal). Plus the period-8→6 rebase (separation cleaner, decoy 154→129). One model / one card / one GNA part
A byte-exact 12B forward — the four float "islands" killed	RMSNorm / softmax / GELU / RoPE now deterministic exact-integer (RoPE by device CORDIC, no libm; no __int128 — M ≈ 2^60 fits a u64); OFF = PPL 4.6665 byte-identical to bf16 gold (null floor); ON = 4.6569 parity (−0.21% @ n=42); ON run-to-run bit-identical	gated + citable (X-BX-ISLANDS). Byte-exact buys auditability (cross-machine determinism), not speed or size. The one open step is external: a bit-identical logit check across two physical GPUs (needs a 2nd machine). One model / one card
One substrate, every backend + an O(1) 12B decode through the daemon	a universal Rust crate is the bit-exact reference the C/CUDA/HVX backends gate to; attention's ⟨q,k⟩/p·V → exact-integer dual-prime convolution (p·V ~2^46 ≪ M ⇒ no third prime); daemon decode 32/32 tokens bit-identical to the oracle, VRAM flat (O(1))	gated + citable (X-BX-WIRE). The decode bit-identity is on one host (the 2-GPU check is X-BX-ISLANDS's open step); "O(1) VRAM" is the KV-cache term. One model / one card
Byte-exact ≠ compression (the de-conflation)	the compression levers were convicted — incoherence rotation ~1.37× / column reorder ~1.05×, both redundant vs OK_Q4B (already gold PPL); and T2 on the real 12B embedding reconstructs at cos 0.032 ≈ random	gated + citable (X-BX-BOUNDARY). The boundary thesis: O_K wins on exact arithmetic (the container), never on structuring content — four content-side levers measured inert. And one honest negative about our process: the byte-exact linear algebra was already in the bounded crate — re-deriving it was the wasted motion

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). The 0.6B (Qwen3-0.6B) carries the two-ring memory ladder; the 12B (Gemma-3-12B, QAT 4-bit — the B1 artifact) carries the XBAR and KAIROS results. CPU decode is ~1.34× behind a tuned llama.cpp at the same quantization. On GPU the citable point is the speed/quality PAIR: 26.1 tok/s at PPL 5.12 — a point no other stack currently occupies on this model at any speed, because their artifacts are broken. The memory envelope remains the primary value claim.

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 (written, citable — X-R1) — 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 (written, citable — X-R2) — 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 (written for the mechanism — KAIROS-01/02/03; full release held 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), and the wrap-aware journaled ring uniting O(1)-time with O(1)-space + the semantic metal loop (KAIROS-03). The ≥24 h endurance soak is running now — no verdict from a mid-run log; the full release ships on the receipt.
• 10 — Receipts or it didn't happen (written) — bit-exact-or-bounded as the contribution: the four documented self-corrections that prove the gates discriminate.
• 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).
• 16–18 — The discrete container (written, citable) — the memory architecture re-carried onto the project's own exact-integer O_K substrate (Q(√-163) / the dual-prime negacyclic CRT-NTT), and the boundary it draws: the integer container wins where arithmetic must be exact, while every attempt to make memory content itself number-theoretic is measured-inert and kept as an honest negative.
  • 16 — The Unification (written, citable — X-OK-BIND) — the Ring-3 holographic bind, which had been floating-point, re-carried onto the engine-native CRT-NTT: 256/256 bit-identical to native sp_pr_mul and reduction-order-immune (integer M byte-identical across 8 summation orders; float drifts 4.44e-15); the live loop wired to the native multiply (CAP=32 unregressed); and the first negative — a number-theoretic χ_d carrier lowers coherence but is operationally inert (the integer container is the win, not the integer carrier).
  • 17 — The Frobenius integer episode store (written, citable — X-OK-FROB) — the Ring-2 episode codec on the integer ring, sub-ULP at 24 bits (a16b8: relative L2 1.2e-7, 0.76× store), with losslessness established by reconstruction fidelity, not a fake +0.000% (the n=42 perplexity gate is honestly reported as blind below ~1%); plus the three boundary negatives — entropy-coding the codes (1.02×), Möbius over the superposition (sheds memories), T2 on real model weights (recon ≈ random) — that draw the line: O_K wins on exact arithmetic (the container), never on structuring high-entropy content.
  • 18 — The organism breathes (full loop) (written, citable — X-OK-ORG) — the full cross-modal continuous→discrete→continuous loop on the integer substrate: real audio sensed on physical GNA → a discrete 256-bit signature (audio 256 / text decoys 147, 129) → integer superposition → audio cue retrieves top-1 (cos +0.47) → the C2 cross-modal verify (accept audio / reject text) → integer Frobenius land (K/V relL2 ~9e-8) → metal SP_REPLAY into the live 12B (checks=5, fails=0); plus the period-8→6 rebase to the true gemma-4 global layers (separation cleaner, decoy 154→129 — the paper-15 caveat retired).
• 19–21 — Byte-exact (written, citable) — the forward pass itself carried onto the exact-integer substrate: the four nonlinear float "islands" killed, the universal-crate architecture that makes the exactness identical on every backend, and the de-conflation that keeps it honest — byte-exact buys auditability, not compression.
  • 19 — Killing the float islands (written, citable — X-BX-ISLANDS) — the four nonlinear fp32 operations of the 12B forward (RMSNorm, softmax, GELU, RoPE) made deterministic exact-integer, with no libm (RoPE by a device CORDIC, the norm's reciprocal-root by a 64-bit isqrt split) and no __int128 (M = q1·q2 ≈ 2^60 fits a u64); G-BYTEEXACT-FORWARD-12B: OFF = PPL 4.6665 byte-identical to bf16 gold (null floor); ON = 4.6569 parity (−0.21% @ n=42); ON run-to-run bit-identical — the one open step is the external two-physical-GPU check.
  • 20 — One substrate, every backend (written, citable — X-BX-WIRE) — a universal Rust crate is both the L2 orchestrator and the bit-exact reference the C/CUDA/HVX backends gate to; attention's ⟨q,k⟩/p·V become an exact-integer dual-prime negacyclic-convolution dot (a single coefficient = the plain dot; the p·V accumulator ~2^46 ≪ M ⇒ no third prime); and the resident daemon drives the 12B over the L1 C ABI, with a new token-by-token decode verb — G-WIRE-CUDA-DECODE-GEMMA4: 32/32 tokens bit-identical to the oracle, VRAM flat (O(1)).
  • 21 — Byte-exact, not compression (written, citable — X-BX-BOUNDARY) — the reflective record: the de-conflation (auditability vs compression), the convicted compression levers (incoherence rotation ~1.37× / column reorder ~1.05×, both redundant vs OK_Q4B), the boundary thesis on the forward (four inert content-side levers; T2 on the real 12B embedding ≈ random), and the honest negative about our own process — the byte-exact linear algebra was already in the bounded crate, bit-exact-gated, and re-deriving it offline was the campaign's one wasted motion.
• GEMMA4-QUANT-FIX.md — community tutorial: verify the gemma-4 GGUF breakage yourself (engine-independent, ~30 min) and the working fix recipe. Ready-to-post issue text: GEMMA4-ISSUE-POST.md.

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)

The point of the split: raw recall degrades past ~16× selection budget (the measured 32k MISS is the honest anchor), so Ring 2 stays bounded and recent — where the budget is favorable — and the long tail lives in Ring 3 as compact gist. The Exec queries both per step and attends over the union.

XBAR — the Auditable Latent Crossbar

Multi-agent systems today communicate by detokenizing model A's state into text and retokenizing it for model B. The boundary is lossy, slow, and discards everything the residual stream knew that the argmax threw away. XBAR bypasses the boundary: two models — the Exec (generator) and Memo (a small, differently-trained curator) — share the ring memory and communicate through latent state, not tokens, with every write receipt-backed, gated, and rewindable. "Auditable" is the one word no floating-point agent stack can claim, and it is the entire reason this lane exists.

What is measured so far — [PROVEN], ledger row X-R1: a gemma-4-12B's generation steered by direct KV-cache transplant (no tokens involved) — 15/15 lexical incorporation across a 5-prompt × 3-concept matrix, 15/15 selectivity with a 2×2 double dissociation, max single-token rank pull 3.69 orders, dose-response from a single row (~4% attention mass) to a 6-row lexical breach, instrumentation null bit-identical 7/7, and coherence certified by the gold instrument (steered-text PPL 1.70–4.10 against the model's true 4.68). Raw KV splice is a deliberately blunt instrument; the learned-adapter phase exists to refine exactly that. (Paper 07.)

And the cache the crossbar writes into is now O(1) in context — [PROVEN], ledger row X-R2: a learned 512×32 LSH router selects the global top-B keys at +0.47% PPL at 8× compression (oracle ceiling −0.08%; the frozen ±1 router was +4.17%, RED), a compact slab realizes O(1) VRAM (nvidia-smi flat within ~50 MiB from 8k to 16k context, where a full O(N) cache would add ~5.4 GiB), and the NIAH needle survives the compaction at depths 10 / 50 / 90% — learned-router-only, with the frozen-router negative control MISSING to isolate the projection as the cause, not leakage. The KV term is O(1); the absolute footprint in this backend-direct harness still carries the resident model. (Paper 08.)

Design rules (fixed, not aspirational):

1. Memo is small. It sorts latents, it does not speak; it co-resides with Exec — no weight-swap latency.
2. Memo is non-causal. It is offline and sees the whole episode at once — the architectural form of consolidation, not a vague autoencoder.
3. Shadow ring, promote-on-accept. Memo never writes canonical Ring 2 directly. Proposals land in Ring 2′; a downstream coherence gate accepts → promote with receipt, or rejects → rewind.
4. Geometry is the law. Nothing enters the ring that does not honor the per-layer, per-head, position-exact coordinates.
5. One CRT prime per modality lane. Audio/text/video blocks are residue-separable in the same ring; lanes can never alias; provenance stays recoverable.

The honest negative, stated up front: "injected memory as sudden realization" and "confident hallucination from off-manifold state" are the same event described twice. The discrete substrate detects invalid blocks (Spinor sentinel, Frobenius-lift identity); it cannot detect semantically-wrong-but-valid ones. Therefore the coherence gate is load-bearing, not decorative — no promotion without a measured downstream delta, accept-or-rewind, every time.

KAIROS — the time / agency axis (the resident cognitive loop)

XBAR gives the cache an O(1) space axis. KAIROS adds the time axis on the same card: a frozen gemma-4-12B running as a resident background daemon that wakes each tick, reads one environment event, and replies with exactly NO_OP (stay silent) or <ACTION>…</ACTION> (intervene). The whole point is disciplined silence — a useful resident agent must do nothing, correctly, almost all the time — and the machinery that makes that cheap forever is an O(1) byte-exact rewind: rewind(Δ) logically decrements the decode position, and because each cache slot maps to exactly one position, the sheared slots are never read again, so the rewind is a perfect inverse. The daemon can "think" on a tick and then perfectly un-think it.

What is measured so far:

• [PROVEN], ledger row KAIROS-01: the 24-tick crucible is PERFECT — 21/21 idle → NO_OP, 3/3 salient → coherent contextual ACTION (start / clean / renew for build-finished / disk-95% / ttl-expiring), 0 false-action, 0 missed, 0 malformed, 0 drift. The negative control proves it is capacity, not plumbing: the identical harness collapses a 0.6B model into a deterministic corruption attractor (NO_克作); the 12B holds.
• [PROVEN], ledger row KAIROS-02: cold-evict happens at the metal — after an idle tick + rewind(Δ), the KV is byte-identical to never-visited across all 48 owner layers (16.5 MB, diffs = 0), and re-running the tick reproduces identical tokens. The O(actions)→O(1) telemetry: prefix-grow (host re-ingest) rises at 0.924 s/action; the metal rewind is flat at 0.0073 s/action — a 127× shallower slope — the measured flatline that is the O(1) claim.
• [PROVEN], ledger row KAIROS-03: the O(1)-time rewind now runs on the O(1)-space SWA ring — a wrap-aware journaled ring byte-exact across a forced wrap (40/40 owner layers clobbered, diffs = 0), with the journaled-ring slope 0.00365 ≈ 0.00371 s/action (no asymptotic cost), and the full run_kairos_metal semantic loop (commit-on-action / rewind-on-idle) passing the 24-tick tape 0 false / 0 missed / 0 drift / 0 pos-violations. Space ⊗ time ⊗ cognition, one resident system.
• [PROVEN], ledger row X-R3: the crossbar now writes as well as reads — a stored episode replayed into the cache reproduces the baseline bit-identically (diffs = 0) while a zeroed episode diverges 12/12 (so the payload is load-bearing), proven on both the 12B and the smaller E2B; and replaying a foreign episode deflects perplexity only +1.38%, inside the pre-registered <2% gate (n = 42, a single chunk — deterministic, with a larger-N run named as the hardening lever). The auditable latent crossbar now reads, writes, compresses to O(1), retrieves under poison, and replays without breaking the model.
• [PROVEN], ledger row KAIROS-04: the model gains a real-world audio sense, physically realized on silicon — real speech drives the resident 12B to the correct NO_OP/ACTION 7 of 8 times (the 1 miss is a conservative ACTION→NO_OP), and the audio front-end runs on a physical Intel GNA 2.0 accelerator at 0.877 token-recovery == software-emulation == FP32 (a naive int16 quantization sheared it to 0.667; a calibrated GNA-native int16 fully recovered). The "EAR" is a separate-but-related sibling of the latent-memory crossbar, sharing the same KV-inject seam.

Honest scope: a 24-event scripted tape (not live sensors), one model, one card. A 6 h unattended endurance soak is GREEN (351 loops / ~8,400 ticks / 6 h 01 m, 0 false / 0 missed / 0 pos-violation); the formal ≥24 h soak is un-pursued by operator choice (not failed), so paper 09's release stays held pending the operator's wording. (Paper 09.)

This session (the discrete container — papers 16–18, written + citable). The whole memory architecture was re-carried off generic floating-point carriers onto the project's own exact-integer O_K substrate (Q(√-163) — the dual-prime negacyclic CRT-NTT, the Frobenius lift), so the organism now breathes end-to-end on the discrete container it was designed for. The holographic superposition that held the consolidated memories is now computed on the integer ring 256/256 bit-identical to the native arithmetic and reduction-order-immune (the integer answer is byte-identical regardless of summation order, where the float answer drifts at ~4e-15); the Ring-2 episode store became an integer Frobenius-π^k codec (a reconstruction-faithful, sub-ULP rank-2 lattice — and we keep the honest caveat that a coarse single-chunk perplexity gate cannot see that fidelity, so no zero-delta is claimed from it); and the full real-episode loop — real audio sensed on the physical GNA, addressed by a discrete signature, superposed and verified and stored as integers, then landed back into the live 12B cache — runs clean. The receipts also draw a boundary, stated up front: the integer substrate wins where arithmetic must be exact (the container), while every attempt to make the memory content itself number-theoretic — split-prime carriers, a Möbius sieve over the superposition, entropy-coding the codes, a Möbius transform of the model's own weights — measured inert or worse, and each falsified attempt is kept on the record. The named next step is to apply the one validated content-side lever (the Frobenius-π^k codec) to the 9.4 GB model weights themselves. (Papers 16–18.)

Recently landed + forward. Two lines built on the same proven KV-cache delivery seam (the direct latent-write primitive behind paper 07). (1) An audio "EAR" line — now realized on hardware (ledger row KAIROS-04): real speech projected into the residual stream through that seam drives the resident 12B correctly 7/8, and the fixed-point front-end runs on a physical Intel GNA 2.0 accelerator at no quality loss (0.877 == software-emulation == FP32). This is a separate-but-related sibling of the latent-memory crossbar; with it landed, the project's center of gravity has returned to the XBAR latent-memory crossbar. (2) Latent-interrupt delivery — the single-event inject seam is frozen and validated as a delivery primitive; the learned compressed single-event codec is an open, bounded problem (a single packet recovers most of the signal but the compression wall is sequence-positional — no clean win yet), so it stays internal until a clean gate. The active forward edge is now the orchestration tier above the crossbar — an autonomous recall loop (recall-cue → episode-id → replay) over the now-closed read/write substrate; [IN PROGRESS] / [DESIGN], not yet a ledger row.

NIGHTSHIFT — offline consolidation [DESIGN]

Sleep does not just tidy the hippocampus; it replays raw episodes and writes compressed semantic traces to neocortex. NIGHTSHIFT does the same, on idle time: it reads aging Ring 2 episodes, compresses spans via the adapter, proposes the gist to Ring 2′, and on gate-accept promotes it to Ring 3 — eviction of the now-redundant raw positions happens under the same receipt or not at all. The association-strength signal already exists in measured form (the recall path's temporal-locality telemetry). A synthetic subconscious whose dreams are auditable.

Honest constraint carried forward: the NIAH budget ladder broke at 16×–32× selection, so v0 NIGHTSHIFT bounds episodes (≤8k tokens) or runs two-stage re-rank; budget scaling is an open risk-register item, not a buried assumption.

Direction and reasoning: why discrete, why auditable

The substrate. Positions, indices and routing computed exactly (CRT-NTT arithmetic, Frobenius lift, KSTE tiering) means internal state can carry proofs instead of vibes: a Spinor block is 63 bytes + a sentinel — one cache line — and its Frobenius-lift identity is a bit-level integrity receipt. Floating-point drift and unprovable identity are entropy bleeding into the hardware; the lattice makes correctness a property you check.

The discipline (in full in METHODOLOGY.md) is as much the contribution as any mechanism:

• Bit-exact when off. Every mechanism is a strict no-op in its default state — the baseline is provably the original network.
• No number without a command. Nothing is claimed that is not a ledger row reproducible by a specified command.
• Scope travels with the number. Every figure carries its model, context, corpus, and what it does not generalize to.
• No silent gate revisions. If an implementation can't meet a pre-stated gate, that surfaces upstream — gates are never quietly retuned until a number passes.
• Falsification pre-stated; honest negatives published. The 32k NIAH MISS (01-R9) stays on the front page; the 34.2 tok/s headline was retired by the series' own quality rule; a falsified recall signature is reported, not hidden. A result with its caveats attached is one a reader can trust without re-deriving the authors' incentives.

A note on the latent layer and security. Deployed AI safety today lives almost entirely at the lexical layer — refusal training, input filters, output classifiers all scan text — while the decision is made in the residual stream. The field's trajectory (multimodal projectors, agentic/retrieved context, shared KV-cache serving) steadily adds pathways that reach latent space without passing the layer where safety is enforced. Calibration matters: direct latent writes require runtime ownership — a deployment-isolation threat, not a remote skeleton key — and the structural worry is a widening gap on a multi-year horizon, not an imminent break. The connection to this project is the constructive half: latent state has been an un-inspectable continuous blob, and XBAR's premise is the counter — a discrete substrate where a block of internal state is provably well-formed, every memory write carries a receipt, and nothing commits without passing a coherence gate. A verifiable, gated latent substrate is a defense direction the field currently lacks; we record that as motivation, not as a project pivot.

Older material

The original document history — theory drafts, Friedman/KSTE notes, results, and tools — has been moved to Archived/. It is kept for provenance, not as a starting point. Begin with the paper series above or the live project.

Related repositories

Main project

• shannon-prime-lattice — the lattice: discrete Z_q substrate, the headline implementation
• shannon-prime-system — math core
• shannon-prime-system-engine — inference engine

Earlier / supporting

• shannon-prime
• shannon-prime-engine
• shannon-prime-llama
• shannon-prime-bernhard
• shannon-prime-burnhard
• shannon-prime-lmstudio-server
• shannon-prime-comfyui

Audio / Voxtral

• voxtral-tts.c
• voxtral-mini-realtime-rs
• ComfyUI-FL-VoxtralTTS

License and citation

MIT. Cite via CITATION.cff.

---

Shannon-Prime-Lattice is an open-source research project by KnackAU — contact: raydaniels@gmail.com

Attributed to Transformers and 250 years of Mathematicians.