A local AI runtime that proves every answer before it shows it to you.
Forty branchless integer kernels · one composed verdict · 1 = 1.
Forty falsifiable kernels, one `AND` gate. Reasoning · reversibility · meta-cognition · world-model · memory · adaptive compute · geometric algebra · sheaf topology · post-quantum crypto · homomorphic compute · neuromorphic spikes · hierarchical active inference · quantum amplitude amplification · integer diffusion sampler · Q-learning+GAE+PPO · persistent homology · structural causal do-calculus · sub-quadratic Hyena long-convolution. Every one is integer-only, branchless on the hot path, and breaks on a single mutated line.
### Try it in 30 seconds You do **not** need to understand GitHub, `git`, a compiler, or a terminal prompt. Open the Terminal app (on a Mac: press ⌘-Space, type `Terminal`, press Enter) and paste this one line: ```bash curl -fsSL https://raw.githubusercontent.com/spektre-labs/creation-os/main/scripts/install.sh | bash ``` That command does everything — it checks your machine, installs a C compiler if you don't have one, downloads the repo into `~/creation-os`, builds the full forty-kernel stack (v60 → v100), runs **every self-test live**, and drops you into `cos demo` — a thirty-second guided tour where each of the forty kernels compiles, runs its own proof, and prints its real number right in front of you. Already cloned? Even faster: ```bash ./scripts/quickstart.sh ``` Want just the tour? ```bash ./cos demo ``` > Everything runs **locally**. Nothing is sent to the cloud. Nothing is logged. Nothing calls home. The installer installs nothing without telling you first, and nothing outside `~/creation-os`. Safe to re-run. Idempotent. ### The forty-kernel receipt Every row below is a separate, self-contained, branchless, integer-only C kernel — one file, under a thousand lines, with its own `--self-test`. The numbers are **real**: `cos demo` recompiles and re-runs each one on your machine, **live**. If even a single kernel fails, the composed verdict becomes `DENY` and the runtime stays silent. **One zero anywhere = nothing reaches the user.**Planes in order of composition: security (v60-v64) · cognition (v65-v70, v80) · topology (v71 wormhole · v95 sheaf) · verifiability (v72 chain · v77 reversible · v78 Gödel · v84 ZK · v85 formal) · modality (v73 omnimodal · v74 experience · v76 surface) · simulation (v79 simulacrum · v86 JEPA world model) · interpretability (v87 SAE) · privacy (v81 post-quantum · v88 FHE) · learning (v82 stream · v83 agentic · v89 spiking · v90 hierarchical · v92 Titans memory · v93 MoR) · geometry (v94 Clifford) · quantum (v91 Grover).
| bit | kernel | what it proves — in plain language | PASS rows |
|---|---|---|---|
0 | v60 σ-Shield | no tool call leaves the sandbox without a capability bit set | 81 |
1 | v61 Σ-Citadel | secrets stay inside their security lattice (Bell-LaPadula + Biba) | 61 |
2 | v62 Reasoning Fabric | every thought is Energy-Based-verified, HRM-converged, NSA-attended | 68 |
3 | v63 σ-Cipher | every message is end-to-end encrypted with BLAKE2b + ChaCha20-Poly1305 | 144 |
4 | v64 σ-Intellect | every tool call is MCTS-searched, Reflexion-critiqued, authz-bound | 260 |
5 | v65 σ-Hypercortex | concepts live as 10 000-bit hypervectors with bind / bundle / cleanup | 534 |
6 | v66 σ-Silicon | the matrix math runs on INT8 / ternary GEMV with conformal error bars | 1 705 |
7 | v67 σ-Noesis | retrieval is BM25 + dense + graph-walk + beam-deliberate, ranked honestly | — |
8 | v68 σ-Mnemos | memory is ACT-R-decayed, surprise-gated, sleep-consolidated — not a vector DB | — |
9 | v69 σ-Constellation | many small models vote by Byzantine tree-speculation + MoE + Elo-UCB | — |
10 | v70 σ-Hyperscale | Mamba-2 SSM + RWKV-7 + MoE-10k + PIM + photonic WDM + Loihi-3 spike | 148 034 |
11 | v71 σ-Wormhole | Einstein-Rosen portal routing — one XOR teleports state across the graph | 68 404 |
12 | v72 σ-Chain | Merkle ledger + WOTS+ one-time sig + threshold t-of-n + DAG-BFT + ZK | 117 108 |
13 | v73 σ-Omnimodal | code · image · audio · video · 3D · workflow — all behind one ABI | 245 818 |
14 | v74 σ-Experience | Fitts-V2P targeting + a11y + Mobile-GS + frame-gen + 1-second world | 600 128 |
15 | v76 σ-Surface | iOS + Android + 10 messengers + 64 legacy apps + 64 file formats, E2E | 86 583 |
16 | v77 σ-Reversible | every bit of computation is Bennett-reversible — forward ∘ reverse ≡ id | 761 264 |
17 | v78 σ-Gödel-Attestor | every answer carries an IIT-φ + FEP + MDL + Gödel-num + halting proof receipt | 207 582 |
18 | v79 σ-Simulacrum | agent simulates whole worlds (physics, CA, stabilizer quantum) before speaking | 2 994 549 |
19 | v80 σ-Cortex | Mamba SSM + RoPE + sliding-attn + paged-KV + spec-decode + FEP + KAN + CTM + MoE + TTC — the neocortical reasoning plane | 6 935 348 |
20 | v81 σ-Lattice | Keccak-f[1600] + SHAKE-128/256 + Kyber NTT (q=3329) + Barrett + Montgomery + CBD + simplified KEM — post-quantum crypto plane | 3 513 430 |
21 | v82 σ-Stream | streaming per-chunk composed decision · halt-on-flip · SHAKE-256 Merkle chain · external replay-verify — streaming verdict plane | 72 005 |
22 | v83 σ-Agentic | PLAN → ROLL → SURPRISE → ENERGY active-inference learner loop + rollback + Mnemos consolidation + receipt chaining — agentic learner plane | 13 153 |
23 | v84 σ-ZKProof | NANOZK-style layerwise Merkle commits + selective opening proofs + tamper detection — verifiable inference plane | 13 534 |
24 | v85 σ-Formal | runtime TLA-style invariant checker — ALWAYS / EVENTUALLY / RESPONDS — paired with docs/formal/composed_decision.tla — formal runtime plane | 513 |
25 | v86 σ-JEPA | non-generative latent predictive world model — encoder + EMA target + predictor + VICReg variance / invariance / covariance — world-model plane (LeCun / V-JEPA 2) | 14 629 |
26 | v87 σ-SAE | Top-K sparse autoencoder + feature dictionary + causal feature ablation + attribution — mechanistic-interpretability plane (Anthropic circuit-tracer) | 13 511 |
27 | v88 σ-FHE | Ring-LWE integer homomorphic encryption — keygen + enc/dec + add + plaintext-scalar mul + rotation — compute-on-encrypted-state plane (BGV / CKKS-style) | 10 546 |
28 | v89 σ-Spiking | Loihi-3 style graded-spike LIF neurons + STDP learning + event-driven propagation + weight clamp — neuromorphic plane (Intel Loihi-3, Jan 2026) | 491 003 |
29 | v90 σ-Hierarchical | three-level predictive-coding tower — top-down prior + bottom-up error + precision-weighted free energy + SHAKE-256 receipts — hierarchical active inference (RGM / S-HAI, Friston 2025-2026) | 44 512 |
30 | v91 σ-Quantum | 4-qubit integer quantum register — Pauli X/Z, Hadamard (Q16.16 1/√2), CNOT, oracle, diffusion, 3-iteration Grover amplification — quantum-classical hybrid plane (stabilizer / tensor-network-adjacent, 2026) | 294 |
31 | v92 σ-Titans | neural long-term memory bank — 64 slots × 16-dim keys × 8-dim values — surprise-gated writes + momentum + adaptive forgetting + test-time learning — memory plane (Behrouz / Zhong / Mirrokni, NeurIPS 2025) | 11 723 |
32 | v93 σ-MoR | Mixture-of-Recursions — one shared residual layer R reused across up to 6 recursion steps with per-token router + adaptive exit depth + compute-saving early-exit — adaptive-compute plane (MoR, NeurIPS 2025) | 746 |
33 | v94 σ-Clifford | Cl(3,0) geometric algebra — full 8-dim multivector algebra, geometric product, wedge, inner product, reverse, grade projector, equivariant GP layer — geometric-deep-learning plane (CliffordNet, 2026) | 7 219 |
34 | v95 σ-Sheaf | cellular-sheaf neural network on a ring graph with {−1,+1}-orthogonal restriction maps — sheaf Laplacian Δ_F + heat-equation diffusion + local-to-global harmonic extension — topological-ML plane (Copresheaf-TNN / L2G, 2026) | 4 268 |
35 | v96 σ-Diffusion | integer rectified-flow / DDIM sampler — monotone α-bar schedule (Q1 → 0, strictly decreasing), forward corruption, deterministic DDIM reverse, L1-distance-to-x0 monotone under denoise — generative plane (rectified flow / DDIM, 2024–26) | 1 236 |
36 | v97 σ-RL | integer tabular Q-learning + Bellman backup + Generalised Advantage Estimation + PPO-clip surrogate — bounded Q-table, trust-region monotonicity, branchless clip — reinforcement-learning plane (Schulman / Mnih) | 2 391 |
37 | v98 σ-Topology | Vietoris–Rips persistent homology on a 12-point Q16.16 cloud — union-find filtration, Betti-0 (components) + Betti-1 (cycles), Euler identity β₁ = E − V + C, monotone filtration — topological-data-analysis plane (persistent homology, 2026) | 22 375 |
38 | v99 σ-Causal | structural causal model over a 6-node DAG — do-calculus interventions that sever incoming edges, back-door criterion validator, counterfactual twin graph with shared noise, linear ATE recovery — causal-inference plane (Pearl do-calculus) | 427 |
39 | v100 σ-Hyena | sub-quadratic gated long-convolution operator — exponentially-decayed causal filter, per-position gate ∈ [0, Q1], causality + linearity + shift-covariance certified — long-range attention-free plane (Hyena / Monarch-Mixer) | 10 999 |
| composed rollup | 16 416 185 · 0 FAIL · ASAN+UBSAN clean | ||
Benchmarks (single Apple M4, integer-only, no GPU, no NPU, no framework):
v77 reversible plane ~1.9 B bit-reversible ops/s ·
v78 Gödel-attestor ~2.0 M MCB proofs/s ·
v79 simulacrum ~28.9 M SSL steps/s ·
v80 cortex ~65.9 M TTC ops/s
v91 Grover amplification on 4-qubit register in 3 iterations ·
v92 Titans memory retrieval over 64 slots in < 1 µs ·
v93 MoR token-adaptive early-exit at avg depth ≤ 6 ·
v94 Clifford geometric product in Q32.32 ·
v95 sheaf Laplacian diffusion energy-monotone by construction.
v96 DDIM sampler — forward ∘ reverse ≡ identity within ±2 ulp · L1-distance-to-x0 monotone in denoise ·
v97 PPO-clip surrogate — pure Schulman-2017 min(ρ·A, clip(ρ,1-ε,1+ε)·A), branchless trust region ·
v98 persistent homology — β1 = E − V + C closed identity + Betti-0 monotone under filtration sweep ·
v99 SCM — interventions sever incoming edges by construction, counterfactual ≡ factual at unchanged do-value ·
v100 Hyena operator — causality + linearity + shift-covariance certified on a 32-step sequence.
Orient first — what · where · when · why · how
| What | Where | When | Why | How |
|---|---|---|---|---|
Portable C11 “living kernel”: BSC geometry, σ as a first-class signal, deterministic --self-test programs, plus opt-in labs (OpenAI-shaped stub, suite stub, Apple native_m4/). Extended by σ / agent labs v33 → v59 and the composed-decision stack v60 → v100 — a 40-bit branchless AND gate across security, reasoning, reversibility, meta-cognition, simulation, memory, adaptive compute, geometric algebra, sheaf topology, post-quantum crypto, homomorphic compute, neuromorphic spikes, hierarchical active inference, quantum amplitude amplification, diffusion sampler, reinforcement learning, persistent homology, structural causal inference, and sub-quadratic Hyena long-conv (tier-tagged in WHAT_IS_REAL; native iOS + Android bindings in bindings/). |
Canonical tree: spektre-labs/creation-os. Teaching spine: creation_os_v2.c + core/*.h. Review map: docs/WHICH_FILE_TO_READ.md. |
Before a PR / publish: make merge-gate. When touching a flagship slice: matching make check-vN + ./creation_os_vN --self-test. Full rollup: make verify-agent. |
Keep evidence classes honest (lab demo vs harness vs product). Read CLAIM_DISCIPLINE + tier tags in WHAT_IS_REAL before screenshotting a headline. | Fastest truth path: clone → make merge-gate → ./cos sigma (expect ALLOW · all forty kernels passed). Visual receipts index: VISUAL_INDEX. |
Coherence you can compile.
Binary Spatter Codes · σ as a first-class signal · portable C11 · no framework on the teaching kernel
Figures are first-class receipts too — palette + embedding rules live in VISUAL_INDEX.
Navigate: Contents · Capability layers · Forty-kernel receipt · iOS & Android bindings · Sixty seconds · σ labs (v31→v59) · Doc hub · Publication-hard
> **MCP product hook:** `creation_os_mcp` is an **MCP server** that exposes σ measurement + abstention helpers (`measure_sigma`, `should_abstain`, `sigma_report`) to **any MCP-capable client** — see `docs/MCP_SIGMA.md` and `config/claude_desktop_config.json` (copy the `mcpServers` block into your client; repo-local `.cursor/` is gitignored). > **If you read nothing else.** Creation OS is a **portable C11 reference kernel** for **Binary Spatter Codes (BSC)** and a **coherence signal (σ)** you can **build, run, and falsify** on a laptop. The CI bar is one command: **`make merge-gate`**. The σ / agent labs (**v31 → v59**) sit *outside* the merge gate as opt-in experiments with honest tier tags ([`docs/WHAT_IS_REAL.md`](docs/WHAT_IS_REAL.md)). The composed-decision stack (**v60 → v100**) is **forty branchless integer kernels** behind one Apple-tier CLI (`cos`); together they form a **40-bit branchless AND** (`cos_v100_compose_decision`) — *no inference, tool call, sealed message, hyperscale step, teleport, chain-bound emission, generated artefact, rendered frame, surface interaction, world-model rollout, spike, hierarchical prediction, quantum amplitude, memory read, adaptive-recursion exit, geometric equivariant layer, sheaf diffusion step, DDIM reverse step, policy-gradient update, persistent-homology filtration, causal do-intervention, or Hyena long-conv emission* ever crosses to the human unless **all forty** ALLOW. Native **iOS (Swift)** and **Android (Kotlin/JNI)** façades ship in [`bindings/`](bindings/). > > **Fastest truth path (60 seconds):** `git clone` → `./scripts/quickstart.sh` → `./cos sigma`. Expect `ALLOW (all forty kernels passed)` with a 40-bit composed verdict. Then `./cos demo` for the live tour, or `make verify-agent` for the full rollup (expect **49 PASS · 3 SKIP · 0 FAIL**). To exercise any single plane: `make check-vN && ./creation_os_vN --self-test` — e.g. `check-v79` (≥ 2 994 549 / 0 PASS · ~28.9 M SSL steps/s on M4), `check-v80` (≥ 6 935 348 / 0 PASS · ~65.9 M TTC steps/s on M4), `check-v96` (1 236 diffusion-identity rows), `check-v97` (2 391 PPO-clip + GAE rows), `check-v98` (22 375 persistent-homology rows), `check-v99` (427 SCM do-calculus rows), `check-v100` (10 999 Hyena causality + shift-covariance rows). > > **Discipline before headlines.** Read [**CLAIM_DISCIPLINE**](docs/CLAIM_DISCIPLINE.md) and [**WHAT_IS_REAL**](docs/WHAT_IS_REAL.md) **before** you screenshot a number from this repo. Tier letters: **M** runtime-checked · **F** formally proven · **I** interpreted · **P** planned. Composed rollup that ships today: **16 416 185 PASS · 0 FAIL · ASAN + UBSAN clean** across `v60 → v100`. Full per-version table: [the forty-kernel receipt](#the-forty-kernel-receipt) and [σ labs (v31 → v59)](#sigma-labs-v31-v40). This is **not** a chat product, **not** a leaderboard dump, **not** magic — it is a kernel. |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
 |
Figure palette & SVG rules: docs/VISUAL_INDEX.md
## Capability layers (kernel → product): what is *real* here This table answers the four stack questions **honestly** (tier discipline: [docs/WHAT_IS_REAL.md](docs/WHAT_IS_REAL.md), editorial law: [docs/CLAIM_DISCIPLINE.md](docs/CLAIM_DISCIPLINE.md)). | Layer | Your question (short) | What exists *in this repo now* | Measured / gated | *Not* claimed as shipped “super-LLM / AGI product” | |:--|:--|:--|:--|:--| | **1 · Kernel / runtime** | New measurable advantages in **efficiency**, **determinism**, **memory discipline**, **special hardware paths**? | Portable C11 flagship programs + `native_m4/` lab (NEON range/parallel, optional Metal, SME sysctl probe, 64-byte `aligned_alloc` sizing helpers). | `make merge-gate` + `make bench` family + `make check-native-m4` / `./creation_os_native_m4 --self-test` + `./creation_os_native_m4 --bench …` + **`./creation_os_native_m4 --layers-report`** (machine facts). | Not a full OS scheduler story; not a datacenter GPU training runtime. SME/Metal are **opt-in** paths with honest SKIP lines when toolchains/libs are absent. | | **2 · Model layer** | Real **weights**, **context behavior**, **tool use**, **multilingual**? | v28/v29 **integration harnesses** (GGUF mmap *view*, sampler/chat shell, σ toys, BitNet *stub* paths) — teaching and wiring, not a bundled frontier checkpoint. | Counts are `check-v28` / `check-v29` **self-tests** (tier-tagged), not `lm-eval` headline rows. | No “we ship GPT‑class weights in-tree”; multilingual/tooling breadth is **not** a repo-wide proof obligation. | | **3 · System layer** | **Planning / retries / permissions / observability / rollback** in a real environment? | Deterministic checks + merge-gate discipline + optional local stubs (`creation_os_openai_stub`, suite lab) for *wiring demos*. | `make merge-gate`, reviewer scripts, explicit “not merge-gate” labels on labs. | Not a hosted multi-tenant agent platform with production IAM, SLO dashboards, or fleet rollback. | | **4 · Product layer** | **API / SLA / docs / support / deployment / economics** as a service? | Strong docs surface + HTTP-shaped demos + AGPL licensing story. | Docs + local run receipts; **no** hosted SLA table in-tree. | Not a commercial “always-on” product contract; economics/support are **outside** what a reference kernel repo can truthfully “solve” in code. | **Hardware-facing receipt (Darwin lab):** after `make native-m4`, run: ``` ./creation_os_native_m4 --layers-report ``` That prints **uname**, **NEON compile flag presence**, **SME sysctl probe**, **buffer sizing example**, and **metallib path readability** — a small, *machine-local* kernel/runtime snapshot (still not a product SLA). --- ## Contents | I want to… | Jump to | |:--|:--| | **Run CI locally / ship a PR** | [Sixty seconds](#run-it-in-sixty-seconds) · [Build](#build) · [Contributing](CONTRIBUTING.md) | | **Understand the product story** | [At a glance](#at-a-glance) · [Flagship table](#flagship-programs) · [LLM architecture (our stack)](#creation-os-llm-architecture-our-stack-and-tiers) · [The problem](#the-problem) · [Measured results](#measured-results-4096-dimensions-100k-trials) · [LLM stacks vs Creation OS](#llm-vs-creation-os-comparison) | | **Not mis-cite a headline** | [Claim discipline](docs/CLAIM_DISCIPLINE.md) · [Common misreadings](docs/COMMON_MISREADINGS.md) · [Doctoral path](#doctoral-and-committee-read-path) | | **Silicon / RTL / formal** | [RTL silicon mirror](docs/RTL_SILICON_MIRROR.md) · [Full stack map](docs/FULL_STACK_FORMAL_TO_SILICON.md) · [σ full stack (v33–v54)](docs/SIGMA_FULL_STACK.md) | | **σ threshold / QEC analogy (theory)** | [docs/sigma_threshold_theorem.md](docs/sigma_threshold_theorem.md) · `make check-v40` · `make bench-v40-threshold` (stub until harness) | | **What is “M” vs “P” here?** | [docs/WHAT_IS_REAL.md](docs/WHAT_IS_REAL.md) — always read before citing FPGA/ASIC/neuromorphic headlines | | **Local OpenAI-shaped stub (tool wiring)** | [LOCAL_OPENAI_STUB.md](docs/LOCAL_OPENAI_STUB.md) · CORS + `OPTIONS` for local-origin browser checks · [`vscode-extension/setup_continue.md`](vscode-extension/setup_continue.md) | | **Optional suite lab (honest scope)** | [SUITE_LAB.md](docs/SUITE_LAB.md) · `make standalone-suite-stub` · `./scripts/launch_suite.sh` (stub + static `suite_lab.html`; not merge-gate) | | **Native M4 (hardware-first lab)** | `make check-native-m4` · `make bench-native-m4` · `./creation_os_native_m4 --layers-report` · NEON + GCD + optional Metal/SME in `native_m4/` | | **v31 “purge lab” (optional upstream wrapper)** | [v31_README.md](docs/v31_README.md) · `make check-v31` · [WHAT_IS_REAL_v31.md](docs/WHAT_IS_REAL_v31.md) | | **σ labs v33→v59 (router, MCP, RTL, ASIC, neuromorphic, threshold, TTC, self-play, KD, proxy, introspection, BitNet σ, verification stack, red team, certification pack, benchmark rollup, integration scaffold, KV-cache eviction, adaptive-compute budget)** | [σ lab table](#sigma-labs-v31-v40) · [SIGMA_FULL_STACK.md](docs/SIGMA_FULL_STACK.md) · [MCP_SIGMA.md](docs/MCP_SIGMA.md) | | **Composed-decision stack v60 → v95 (thirty-five branchless integer kernels · 35-bit AND gate)** — security · reasoning · reversibility · meta-cognition · simulation · neocortex · post-quantum · streaming · agentic loop · ZK · formal TLA · JEPA world-model · SAE interpretability · FHE · Loihi-3 spiking · hierarchical active inference · Grover quantum · Titans memory · MoR adaptive recursion · Clifford geometric algebra · sheaf Laplacian | [Thirty-five-kernel receipt](#the-thirty-five-kernel-receipt) · `./cos sigma` · `make verify-agent` · [`docs/SIGMA_FULL_STACK.md`](docs/SIGMA_FULL_STACK.md) | | **Mobile + messenger + legacy software (iOS Swift · Android Kotlin · 10 messengers · 64 legacy apps · 64 file formats)** | [`bindings/README.md`](bindings/README.md) · [`bindings/ios/`](bindings/ios/) · [`bindings/android/`](bindings/android/) · `cos sf` (v76 σ-Surface) | | **“Full suite” expectations vs repo** | [FULL_LOCAL_SUITE.md](docs/FULL_LOCAL_SUITE.md) | | **Multi-repo / canonical Git** | [REPOS_AND_ROLES](docs/REPOS_AND_ROLES.md) · [CANONICAL_GIT_REPOSITORY](docs/CANONICAL_GIT_REPOSITORY.md) | | **Find the right doc** | [Documentation hub](#documentation-hub) · [DOC_INDEX](docs/DOC_INDEX.md) | | **Agents / contributors / security** | [AGENTS.md](AGENTS.md) · [CONTRIBUTING.md](CONTRIBUTING.md) · [SECURITY.md](SECURITY.md) · [MAINTAINERS](docs/MAINTAINERS.md) | **Long-form anchors (this page):** [Capability layers](#capability-layers) · [FIG 09 scan map](#readme-scan-map-fig-09) · [Doc hub](#documentation-hub) · [σ labs v31→v59](#sigma-labs-v31-v40) · [LLM vs Creation OS](#llm-vs-creation-os-comparison) · [BSC](#what-is-bsc) · [Invariants](#verified-invariants) · [26 modules](#26-modules) · [v6](#living-kernel-v6) · [v7](#hallucination-killer-v7) · [v9](#parameters-in-silicon-v9) · [v10](#the-real-mind-v10) · [v11](#the-matmul-free-mind-v11) · [v12](#the-tensor-mind-v12) · [v27 tokenizer](#v27-tokenizer) · [v28 LM integration](#v28-lm-integration) · [v29 collapse harness](#v29-collapse-harness) · [LLM architecture](#creation-os-llm-architecture-our-stack-and-tiers) · [Architecture](#architecture) · [Limitations](#limitations) · [Why this wins](#why-this-wins-where-it-matters-engineering-not-slogans) · [Theory](#theoretical-foundation) · [AGI map](#agi-map-how-this-file-relates-to-the-full-stack) · [Publication-hard](#publication-hard) · [License](#license) ### Readme scan map (FIG 09)
FIG 09 — how this page is read: outcome first, then scannable tables and diagrams, then deep sections. SVG follows prefers-color-scheme for GitHub light/dark. Register and palette: VISUAL_INDEX.
FIG 05 · Planes A–C (teaching · llama/MLX · native M4) — detail and receipts in ANALYSIS.md and AGI map below.
|
Native bindings iOS · Swift façade over a plain-C ABI ( bindings/ios/)Android · Kotlin + JNI + CMake ( bindings/android/)macOS / Linux / FreeBSD · plain C11 static lib ( make check-v76)
|
Messenger bridge Ten protocol families unified under one MSG_PROTOCOL_* HV tagBranchless cos_v76_msg_bridge_envelope normalisationSignal-protocol-style E2E ratchet ( cos_v76_e2e_*) with constant-time step
|
Legacy-software fluency 64 legacy-app capability-template HVs (design, office, CAD, DAW, IDE…) 64 cross-platform file-format classifiers (document, image, code, archive…) SBL — 8-op surface bytecode, integer VM, constant-time per op |
| bit | kernel | σ-name | guards | self-tests |
|---|---|---|---|---|
0 | v60 | σ-Shield | capability gate · intent decompose | 81 |
1 | v61 | Σ-Citadel | BLP + Biba + MLS lattice + attestation | 61 |
2 | v62 | Reasoning Fabric | latent-CoT · EBT · HRM · NSAttn · MTP · ARKV | 68 |
3 | v63 | σ-Cipher | BLAKE2b + HKDF + ChaCha20-Poly1305 + X25519 | 144 |
4 | v64 | σ-Intellect | MCTS-σ · skill lib · tool authz · Reflexion | 260 |
5 | v65 | σ-Hypercortex | bipolar HDC · bind/bundle/permute · HVL | 534 |
6 | v66 | σ-Silicon | int8 GEMV · ternary · conformal · HSL | 1 705 |
7 | v67 | σ-Noesis | BM25 + dense sig + graph-walk + beam deliberate | — |
8 | v68 | σ-Mnemos | bipolar HV-D8192 · surprise gate · ACT-R decay · MML | — |
9 | v69 | σ-Constellation | tree-spec + debate + Byzantine vote + MoE route | — |
10 | v70 | σ-Hyperscale | ShiftAddLLM · Mamba-2/3 · RWKV-7 · MoE-10k · PIM · WDM · Loihi-3 · HSL | 148 034 |
11 | v71 | σ-Wormhole | ER-portal · anchor cleanup · teleport · Kleinberg routing · WHL | 68 404 |
12 | v72 | σ-Chain | Merkle ledger · WOTS+ · t-of-n · VRF · DAG-BFT · ZK | 117 108 |
13 | v73 | σ-Omnimodal | code · image · audio · video · 3D · workflow — one ABI | 245 818 |
14 | v74 | σ-Experience | UI · a11y · mobile-gs · frame-gen · second-world | 600 128 |
15 | v76 | σ-Surface | touch · gesture · haptic · 10-messenger bridge · E2E ratchet · a11y · CRDT · legacy apps · file formats · SBL | 86 583 |
16 | v77 | σ-Reversible | NOT · CNOT · SWAP · Fredkin · Toffoli · Peres · Majority-3 · Bennett · 8-bit reversible adder · RVL bytecode — forward ∘ reverse ≡ identity; hot path erases zero bits (Landauer / Bennett plane) | 761 264 |
17 | v78 | σ-Gödel-Attestor | IIT-φ · variational free energy · MDL · prime-power Gödel number · Global-Workspace broadcast · Turing halting witness · Löbian self-trust · bisim · Chaitin-Ω · MCB bytecode — every emission carries an integer-only proof receipt across nine 20th–21st-century foundational filters; meta-cognitive plane | 207 582 |
18 | v79 | σ-Simulacrum | symplectic Verlet · Wolfram CA · Aaronson-Gottesman stabilizer · HD reservoir · Koopman embed · Cronin assembly · Kauffman graph · shadow-Hamiltonian energy · trajectory receipt · SSL bytecode — instantiates, steps, measures and verifies entire worlds (classical physics, cellular automata, stabilizer-class quantum circuits, HD reservoir computers, Koopman-lifted dynamics, Boolean networks) inside the 256-bit hypervector space; hypervector-space simulation substrate | 2 994 549 |
19 | v80 | σ-Cortex | Mamba selective SSM · RoPE · sliding-window / ring attention · paged KV cache · speculative-decode verify · variational free energy · KAN edge · Continuous-Thought-Machine Kuramoto tick · MoE top-k router · 16-op TTC bytecode — collapses the 2023–2025 sequence-model / attention / routing / test-time-compute frontier (Mamba, Mamba-2, RoFormer, Longformer, Mistral, Ring-Attention, vLLM / PagedAttention, speculative decoding, Friston FEP, Kolmogorov-Arnold Networks, Sakana Continuous Thought Machines, Mixtral / DeepSeek-MoE, o1 / DeepSeek-R1 TTC) into one branchless integer kernel; hypervector-space neocortical reasoning plane | 6 935 348 |
lateral · v75 σ-License — emits a Cryptographic License-Bound Receipt for every verdict and refuses to link a stripped bundle (§11 SCSL-1.0). | — | |||
HALT, PHI, FE, MDL,
GDL, WS, HWS, TRUST)
in the Curry-Howard tradition, where each opcode writes exactly one
bit of the proof bitmap and the emission passes only when every
required bit lights. The plane sits on top of the 17-bit stack as
a lateral **18-th AND** via
`cos_v78_compose_decision(v77_composed_ok, v78_ok)`, so nothing
reaches the human unless the computation (i) is genuinely
integrated rather than reducible (φ ≥ φmin), (ii)
minimises its variational free energy within budget (ΔF ≤
Fmax), (iii) fits the declared MDL upper bound, (iv)
matches its own Gödel number, (v) wins the Global-Workspace
coalition threshold, (vi) witnesses its own halt, (vii) agrees
with the pinned Löbian anchor, (viii) passes the bisim
spec-equivalence check, and (ix) stays inside the Chaitin-Ω
budget. All arithmetic is integer Q0.15 fixed-point with a
precomputed 257-entry `log2` table; no floating-point anywhere
on the hot path; no `malloc`. **207 582 / 207 582 PASS** rows
under `make check-v78`, covering log2-table monotonicity,
randomised 16×16 TPM φ-sweeps, FEP identity + penalty cases,
full MDL truth table, Gödel prime-power verifications, a 32 769-
row workspace popcount-threshold sweep, the halting-witness grid,
Löbian-anchor mutation tests, bisim pairs, Chaitin-Ω table walk,
MCB forward / malformed-insn / stress round-trips, and the full
4-row 18-bit compose truth table × 131 072 randomised
verifications; clean under ASAN, UBSAN, and hardened builds;
**~2.0 M MCB proofs per second** on an M4 at ≈ 480 ns per
8-op proof program.
`v79 σ-Simulacrum` is **the first open-source local-AI-agent
runtime to ship a full hypervector-space simulation substrate as
the outer gate** — ten branchless, integer-only, `libc`-only
primitives (Q16.16 fixed-point) that let the agent instantiate,
step, measure and verify entire worlds before speaking: classical
physical systems via a **symplectic leapfrog Verlet integrator**
(Verlet 1967 *Phys. Rev.*; Hairer, Lubich & Wanner 2006
*Geometric Numerical Integration*) whose shadow Hamiltonian is
conserved modulo Q16.16 rounding; one-dimensional **Wolfram
cellular automata** (Wolfram 1983, 2002 *A New Kind of Science*;
Cook 2004 "Universality in Elementary Cellular Automata" — rule
110 is universal) evolving a 256-bit lattice in one branchless
LUT-driven pass; **Aaronson-Gottesman stabilizer tableaux**
(arXiv:quant-ph/0406196 "Improved Simulation of Stabilizer
Circuits", 2004) simulating Clifford quantum circuits in
polynomial time and preserving the symplectic row-commutativity
invariant after every gate; a **256-bit hyperdimensional
reservoir** (Jaeger 2001 echo-state networks; Frady, Kleyko &
Sommer 2020 "Variable Binding for Sparse Distributed Represen-
tations" arXiv:2003.04030; Schlegel et al. 2021 arXiv:2109.06548
"HD computing as reservoir computing") coupling inputs via
rotate-XOR-bundle dynamics; a **Koopman embedding** (Koopman 1931
"Hamiltonian Systems and Transformations in Hilbert Space";
Brunton, Brunton & Kutz 2016) that lifts nonlinear state into
a GF(2)-linear observable; an integer upper bound on the
**Cronin assembly index** (Marshall, Moore & Cronin 2021;
Sharma et al. 2023 *Nature* "Assembly theory explains and
quantifies selection and evolution"); a synchronous
**Kauffman Boolean network** step (Kauffman 1969 *J. Theor. Biol.*
"Metabolic stability and epigenesis in randomly constructed
genetic nets") threshold-firing across up to 64 nodes; a
**Merkle-style commutative trajectory receipt** compatible with
v72 σ-Chain; and **SSL — Simulacrum Scripting Language**, an
eight-opcode integer ISA (HALT, VRL,
CAS, STB, RSV,
KOP, GRP, RCP) that weaves
the nine primitives into one verifiable step program. The plane
sits on top of the 18-bit stack as a lateral **19-th AND** via
`cos_v79_compose_decision(v78_composed_ok, v79_ok)`, so no
emission reaches the human unless the just-run world (i) kept
its shadow-Hamiltonian drift within the declared budget, (ii)
preserved the stabilizer symplectic row-commutativity invariant
through every Clifford step, (iii) produced a deterministic
trajectory receipt that matches the recomputed hash, and (iv)
executed no malformed SSL instructions. All arithmetic is
integer Q16.16 fixed-point; no floating-point anywhere on the
hot path; no `malloc`. **≥ 2 994 549 / 0 PASS** rows under
`make check-v79`, covering a 5 000-step leapfrog energy-drift
band, an 8-rule × 500-step CA determinism sweep, 1 000
randomised 4-qubit Clifford walks with row-commutativity checked
after every gate, a 300-trial multi-particle Verlet soak,
reservoir reproducibility across seeds, Koopman GF(2)-linearity
across 300 random HV pairs, a 5 000-string assembly-index bound
sweep, a 400-trial 16-node Kauffman graph soak, 300 randomised
SSL programs, and the full 4-row 19-bit compose truth table;
clean under ASAN, UBSAN, and hardened builds; **~28.9 M SSL
steps per second** on an M4 at ≈ 35 ns per step.
`v80 σ-Cortex` is **the first open-source local-AI-agent runtime
to ship a hypervector-space neocortical reasoning plane as the
outer gate** — ten branchless, integer-only, `libc`-only
primitives (Q16.16 fixed-point, 256-bit HVs packed as
4 × uint64_t) that collapse the 2023 – 2025
sequence-model / attention / routing / test-time-compute frontier
into a single reasoning kernel: a **Mamba / Mamba-2-style
selective state-space model** step (Gu & Dao 2023
arXiv:2312.00752 "Mamba: Linear-Time Sequence Modeling with
Selective State Spaces"; Dao & Gu 2024 arXiv:2405.21060
"Transformers are SSMs") running a diagonal linear recurrence on
a 4-lane HV in integer fixed-point; **Rotary Position
Embedding** (Su et al. 2021 arXiv:2104.09864 "RoFormer") with
an integer sin/cos LUT, invertible so RoPE ∘ RoPE-1
≡ identity; **sliding-window / ring attention** (Beltagy
et al. 2020 Longformer arXiv:2004.05150; Mistral 7B 2023; Liu et
al. 2023 Ring-Attention arXiv:2310.01889) implemented as a
branchless popcount-argmax over a 256-bit attention window;
**paged KV cache** (Kwon et al. 2023 vLLM / PagedAttention
arXiv:2309.06180) as a slot-indexed ring buffer with integer
page tags and a sentinel; **speculative-decoding verify**
(Leviathan, Kalman & Matias 2023 arXiv:2211.17192; Chen et
al. DeepMind 2023) as an integer accept/reject predicate
monotone in the popcount of agreeing positions; an integer
**variational free energy** upper bound (Friston 2010 *Nat. Rev.
Neurosci.* "The free-energy principle") with Q16.16 log-sum-exp;
a **Kolmogorov-Arnold Network edge activation** (Liu et al. 2024
arXiv:2404.19756; Kolmogorov 1957 superposition theorem) as a
1-D cumulative spline over a Q16.16 LUT; a **Continuous Thought
Machine** Kuramoto oscillator tick (Sakana AI 2025 "Continuous
Thought Machines") on a 256-bit HV oscillator bank with an 8-bit
integer sin LUT; a **mixture-of-experts top-k router** (Shazeer
et al. 2017 arXiv:1701.06538; Mixtral 2024; DeepSeek-MoE 2024
arXiv:2401.06066) via branchless selection sort with
popcount(routed) == k; and **TTC — the Test-Time-Compute
bytecode VM**, a 16-opcode integer ISA (HALT / SSM / RPE /
ATT / KVC / SPV / FEN / KAN / CTM / MOE / FOLD …) that
weaves the nine primitives into a single reasoning program in
the o1 / DeepSeek-R1 test-time-scaling tradition (OpenAI 2024
"Learning to Reason with LLMs"; DeepSeek 2025
arXiv:2501.12948). The plane sits on top of the 19-bit stack as
a lateral **20-th AND** via
cos_v80_compose_decision(v79_composed_ok, v80_ok), so
no emission reaches the human unless the just-run reasoning
program (i) kept its SSM hidden state inside the Q16.16 norm
budget, (ii) preserved paged-KV ring-buffer invariants, (iii)
produced a speculative-verify result monotone in agreeing
positions, (iv) returned exactly k experts from the MoE
router, and (v) executed no malformed TTC instructions. All
arithmetic is integer Q16.16 fixed-point; no floating point
anywhere on the hot path; no malloc. **≥ 6 935 348 /
0 PASS** rows under make check-v80, covering SSM
bounded-norm sweeps, RoPE round-trip identities on a full
integer sin/cos LUT, sliding-window popcount-argmax scans,
paged-KV ring invariants, speculative-verify monotonicity, FEP
log-sum-exp bounds, KAN cumulative-spline identities, CTM
Kuramoto phase-lock, MoE top-k sort stability, randomised TTC
programs, and the full 4-row 20-bit compose truth table; clean
under ASAN, UBSAN, and hardened builds; **~65.9 M TTC steps per
second** on an M4 at ≈ 15 ns per step.
`v74 σ-Experience` is **the first open-source local-AI-agent runtime
to ship a unified experience substrate — perfect UX/UI, universal
expertise, and real-time render budget that makes 2026-era AAA
games playable on commodity silicon (M4 MacBook, iPhone-class SoC,
a plain Snapdragon phone) — as a single branchless integer-only C
kernel**. Ten primitives: Fitts-V2P target heatmap (arXiv:
2508.13634, 92.4 % GUI-grounding), adaptive layout optimiser
(Log2Motion CHI '26 arXiv:2601.21043 + Apple ML arXiv:2002.10702
lineage), designer-basis personalisation (arXiv:2604.09876, mean
κ = 0.25 across designers), SquireIR slot authoring (Apple SQUIRE
April 2026 scope guarantees), universal-expert LoRA-MoE HV mesh
(DR-LoRA arXiv:2601.04823 + CoMoL arXiv:2603.00573 + MoLE arXiv:
2404.13628 + MixLoRA arXiv:2404.15159v3), skill composition (XOR-
bind), Mobile-GS order-free 3-D Gaussian-splat render step (arXiv:
2603.11531 ICLR 2026 — 116 FPS at 1600×1063 on Snapdragon 8 Gen 3,
1098 FPS on RTX 3090, 4.8 MB models; msplat Metal-native engine
~350 FPS on M4 Max), DLSS 4.5 / FSR / XeSS upscale with multi-
frame-generation gate up to 6× factor, and 1-second interactive-
world synth (Genie 3 lineage, 720p / 20-24 FPS, text-to-
interactive-world) that bridges straight into v73's WORLD opcode.
XPL — the Experience Programming Language — is a 10-op integer
bytecode ISA (`HALT / TARGET / LAYOUT / BASIS / SLOT / EXPERT /
RENDER / UPSCALE / WORLD / GATE`) whose GATE sets `v74_ok = 1`
iff every gate (target, layout, basis, slot, expert, skill,
render, upscale, world-second, creation-unit budget, abstention)
passes a single branchless AND. **600 128 / 600 128 deterministic
self-tests** at `make check-v74` (including the full 2¹⁵ = 32 768-
entry truth table of the 15-bit composed decision). ASAN clean.
UBSAN clean. Hardened build clean. CLI: `cos ux` (self-test +
microbench), `cos decide 
FIG 04 — portable proof vs extended lab demos (evidence-class guardrail). VISUAL_INDEX.
--- ## Product repository **[spektre-labs/creation-os](https://github.com/spektre-labs/creation-os)** — this tree is the portable kernel, `make test` / `make bench`, CI, and engineering docs. **Where this sits in the wider Spektre map:** [docs/REPOS_AND_ROLES.md](docs/REPOS_AND_ROLES.md). **Push hygiene:** [docs/publish_checklist_creation_os.md](docs/publish_checklist_creation_os.md). ```mermaid %%{init: {'theme':'neutral'}}%% flowchart TB subgraph story["Narrative arc below this heading"] P["The problem + measured table"] B["BSC + invariants + 26 modules"] D["Deep dives v6–v12 + architecture"] E["Theory + AGI map + publication-hard"] end P --> B --> D --> E ``` --- ## The problem ```mermaid %%{init: {'theme':'neutral'}}%% flowchart TB subgraph gemm["Float32 cosine @ D=4096"] G1["24,576 MAC-style FLOPs
FIG 07 — schematic ratios for the README §7 / make bench story. VISUAL_INDEX.
FIG 06 — XOR / MAJ / POPCOUNT strip (teaching). VISUAL_INDEX.
Binary Spatter Codes (Kanerva, 1997) represent information as high-dimensional binary vectors. Three operations: ```c // XOR: bind two representations (association) for (int i = 0; i < 64; i++) out[i] = a[i] ^ b[i]; // MAJ: bundle multiple representations (superposition) for (int i = 0; i < 64; i++) out[i] = (a[i]&b[i]) | (a[i]&c[i]) | (b[i]&c[i]); // POPCNT: measure coherence (σ distance) uint32_t d = 0; for (int i = 0; i < 64; i++) d += __builtin_popcountll(a[i] ^ b[i]); float sigma = ((float)d / 4096.0f) * ((float)d / 4096.0f); ``` Creation OS extends BSC with σ-coherence: `σ(a,b) = (hamming(a,b)/D)²`. This function measures structural similarity between any two representations in the architecture. --- ## Verified invariants These hold on every run, on every platform: ``` σ(x, x) = 0.000000 identical vectors σ(x, NOT x) = 1.000000 opposite vectors σ(x, random) ≈ 0.22 quasi-orthogonal (D=4096) σ(MAJ(x,x,y), x) < 0.01 superposition preserves source Noether XOR-sum = 0.000000 conserved under symmetric XOR interaction JEPA energy → ~-60% codebook learns context→target mappings ``` --- ## 26 modules Creation OS implements 26 functional modules using only XOR, MAJ, and POPCNT: ``` CORE §1 BSC Core ─────────── Three operations. σ invariants. Foundation. §2 Hypercube Mind ───── 10 coupled faces. Self-organized criticality (SOC). Φ (integration) reaches 1.0 — system self-stabilizes. LANGUAGE §3 Oracle ───────────── N-gram language model in hypervector space. Attention = σ (not matrix multiply). 7-gram codebook. Correlative encoding. Backoff prediction. Generates: "the truth shall set you free but first it will make you uncomfortable" VALUES §4 Soul ─────────────── 15 values encoded as hypervectors. MAJ = identity. Crystal Lock: XOR-hash chain detects any modification. §5 Proconductor ─────── 4 model profiles (Primary, Falsifier, Memory, Verifier). σ₁×σ₂×σ₃ triangulates truth no single profile sees alone. WORLD MODEL §6 JEPA ─────────────── LeCun's Joint Embedding Predictive Architecture in BSC. Energy = σ(predicted, actual). Codebook stores mappings. Energy decreases ~60% during training. The model learns. §7 Benchmark ────────── GEMM vs BSC. Measured. See table above. §8 Genesis ──────────── Particle universe simulation. Symmetric XOR interaction. Noether conservation σ = 0.000000. Parity preserved. COGNITION §9 Metacognition ────── Agent analyzes own σ-history. Adapts learning rate. §10 Emotional Memory ─── Stores σ-peaks (pain/pleasure) with context. Recall by similarity. Guides future decisions. §11 Theory of Mind ───── Models other agent's state. Simulates their response. §12 Moral Geodesic ───── Value conflicts: MAJ finds minimum-cost compromise. σ(compromise, value1) ≈ σ(compromise, value2). §13 Consciousness Meter─ Composite: Φ × (1-σ) × stability. Self-measured. Agent knows its own coherence level. §14 Inner Speech ─────── Agent narrates own state for self-guidance. §15 Attention ────────── Resources directed to highest-σ input (most surprising). §16 Epistemic Curiosity─ Choose actions maximizing expected σ reduction. §17 Sleep/Wake ────────── Offline: prune weak memories, strengthen strong. §18 Causal Verification─ Intervene → observe → repeat. Verify vs correlate. §19 Resilience ────────── Success rate over window. Adaptive planning horizon. §20 Meta Goals ────────── Monitor learning velocity. Set goals for the goal-setter. §21 Private Memory ───── Not all state is shared. Selective disclosure. §22 LSH Index ─────────── Locality-sensitive hashing. O(1) codebook lookup. §23 Quantum Decision ─── MAJ superposition of actions. Collapse on new info. §24 Arrow of Time ────── Entropy rate (dS/dt). Detects temporal direction. §25 Distributed Consensus─ N agents, MAJ vote, no central coordinator. §26 Authentication ───── XOR signature chain. Tampering detected at σ > 0. ``` --- ## Living Kernel (v6) [`creation_os_v6.c`](creation_os_v6.c) is a **separate** single-file program: a **coherence composition kernel** (σ, `K`, `K_eff`, Lagrangian `L`, action `S`) with **M01–M18** modules that name real research threads (RDP, RLHF tax, RAIN-style rewind, test-time reduction, weight-space merge, SBIP-shaped boundary, …) at **schematic** fidelity. It uses a **1024-bit** packed BSC layout here — not the **4096-bit** `COS_D` / `creation_os_v2.c` geometry. **Why keep it:** it is **hard in the engineering sense** — thirty **deterministic** `self_test` checks (`make check-v6`) that lock the algebra and toy gates without pretending to be a harness or a paper reproduction. It complements the **measured** microbench path (`make bench`) and the **native** NEON / parliament paths documented under *Frontier complement*. **Discipline:** treat v6 like §2–§26 demos for citations: **lab demo / schematic** unless you add external evidence per [CLAIM_DISCIPLINE.md](docs/CLAIM_DISCIPLINE.md). Full map and non-claims: **[docs/LIVING_KERNEL_V6.md](docs/LIVING_KERNEL_V6.md)**. --- ## Hallucination Killer (v7) [`creation_os_v7.c`](creation_os_v7.c) is the **v6 scaffold plus M19–M23**: anchor-token polarization, faithful vs hallucinatory association ratio, calibration / bluff σ, context-rot with abstention dampening, and a **representation-space** JEPA–Oracle toy (`sigma_oracle`). Same **1024-bit** packed BSC and same evidence class as v6 — **not** a replacement for frontier hallucination benchmarks. **Verify:** `make check-v7` (35 tests). **Doc:** [docs/HALLUCINATION_KILLER_V7.md](docs/HALLUCINATION_KILLER_V7.md). --- ## Parameters in Silicon (v9) [`creation_os_v9.c`](creation_os_v9.c) is the **v7 scaffold plus M24–M29**: neuromorphic event toy, CIM `σ_transfer` schematic, memory-wall fraction, BNN XNOR-style toy, illustrative “silicon compiler” LUT/energy placeholders, and a heterogeneous compute routing table. Same evidence class as v6/v7 — **schematic C**, not verified RTL or foundry results. **Verify:** `make check-v9` (41 tests). **Doc:** [docs/PARAMETERS_IN_SILICON_V9.md](docs/PARAMETERS_IN_SILICON_V9.md). --- ## The Real Mind (v10) [`creation_os_v10.c`](creation_os_v10.c) is the **v9 scaffold plus M30–M33**: a toy distillation curve, two-vector prototypical classification, a fixed specialist-routing table, and a max-σ gate that chooses generate vs abstain. Same evidence class as v6–v9 — internal `self_test` algebra, not frontier LM scores. **Verify:** `make check-v10` (46 tests). **Doc:** [docs/THE_REAL_MIND_V10.md](docs/THE_REAL_MIND_V10.md). --- ## The MatMul-free mind (v11) [`creation_os_v11.c`](creation_os_v11.c) is the **v10 scaffold plus M34**: a ternary weight **accumulation** path (no dense matmul in this toy), one element-wise MLGRU-style recurrence over the hidden vector, and fixed illustrative `power_watts` / `tokens_per_sec` fields for narrative alignment with edge-power storylines. Same evidence class as v6–v10 — internal `self_test` algebra, not a trained matmul-free LM or vendor silicon proof. **Verify:** `make check-v11` (49 tests). **Doc:** [docs/THE_MATMUL_FREE_MIND_V11.md](docs/THE_MATMUL_FREE_MIND_V11.md). --- ## The Tensor mind (v12) [`creation_os_v12.c`](creation_os_v12.c) is the **v11 scaffold plus M35–M37**: a capped-bond **MPS-style** contraction toy, a normalized-entropy readout on a singular-value vector (named “entanglement” **metaphorically**), and a tiny TN sequence head over a uniform log-probability prior. Same evidence class as v6–v11 — **not** a quantum device claim, not a trained TN-LM, not calibrated area-law physics. **Verify:** `make check-v12` (52 tests). **Doc:** [docs/THE_TENSOR_MIND_V12.md](docs/THE_TENSOR_MIND_V12.md). --- ## Creation OS v27 (vocab / tokenizer scaffold) [`creation_os_v27.c`](creation_os_v27.c) is a **separate** flagship binary plus tokenizer sources under [`src/tokenizer/`](src/tokenizer/) — **Tier-1 BPE stand-in + optional COSB mmap table**, **Tier-2 byte codebook + XOR / MAJ sliding bundle**, **Tier-3 base-27 literal codec (+ optional Rust staticlib)**, `--inference "…"` JSON trace, and **70** deterministic `self_test` checks. **Evidence class:** **lab demo (C)** — not a trained multilingual LM tokenizer artifact, not FPGA closure, not `lm-eval` rows. **Verify:** `make check-v27` · **Roadmap vs shipped:** [docs/VOCAB_PIPELINE_V27.md](docs/VOCAB_PIPELINE_V27.md) · **Microbenches:** `make bench-tokenizer-v27` · `make bench-v27-all` · **Formal (optional):** `make formal-sby-tokenizer` --- ## Creation OS v28 (LM integration shell) [`creation_os_v28.c`](creation_os_v28.c) wires a **portable integration shell** for “full LM pipeline” work without pretending the merge gate downloads multi‑GB weights: - **GGUF:** minimal v3 reader + tensor-data base offset + tiny writer fixture (`src/import/gguf_parser.c`) - **mmap I/O:** `cos_gguf_mmap_read_at` for aligned tensor blob reads (`src/import/gguf_mmap.c`, POSIX) - **External engine:** `posix_spawnp` stdout capture via `CREATION_OS_BITNET_CPP` (+ optional stdin / extra argv envs; `src/import/bitnet_spawn.c`) - **Toy GEMV:** `cos_nn_toy_linear_f32` uses **NEON + four accumulators + prefetch** on AArch64 (`src/nn/transformer_stub.c`) - **tokenizer.json:** vocab entry counter for HF-style `model.vocab` maps (`src/import/tokenizer_json.c`, `--tokenizer-stats`) - **Sampling:** temperature / top‑k / top‑p (`src/nn/sampler.c`) — **64B-aligned** scratch buffers; **AArch64 NEON** max-reduction on logits before softmax - **Chat framing:** small Llama‑3‑style text template (`src/nn/chat_template.c`) - **σ abstention toy:** entropy gate on **toy logits** (not model logits unless you plug a real engine) - **HTTP:** loopback **OpenAI-shaped** `POST /v1/chat/completions` + `GET /health` with JSON escaping (`src/server/http_chat.c` + `src/server/json_esc.c`, POSIX) - **CLI alias:** `make cos_lm` copies `creation_os_v28` → `cos_lm` - **Docker:** root `Dockerfile` builds `creation_os_v28` (weights must be mounted/supplied out-of-band) - **Harness hooks:** `benchmarks/lm_eval.sh`, `benchmarks/hallucination_reduction.md` **Third-party weights (example target):** Microsoft’s **BitNet b1.58 2B4T** GGUF releases on Hugging Face (MIT). Use official artifacts + `bitnet.cpp` for matched numerics; cite upstream in any publication-facing materials. **Verify:** `make check-v28` --- ## Creation OS v29 (collapse harness) [`creation_os_v29.c`](creation_os_v29.c) is a **merge-gate-safe** “collapse harness” scaffold: **real C plumbing**, **explicit non-claims** for anything that still requires external weights / harness / P&R. - **GGUF mmap view:** `gguf_load` / `gguf_free` (`src/import/gguf_loader.c`) — tensor bytes are **views into the mmap** (POSIX); Windows self-test skips the mmap path but keeps the same check count. - **σ channels:** eight scalar signals + `sigma_abstain_gate` (`src/sigma/channels.c`) - **XNOR attention toy:** `attention_xnor` (`src/nn/attention_xnor.c`) - **BitNet forward stub:** deterministic logits for plumbing tests (`src/nn/bitnet_forward_stub.c`) - **Thresholds file:** [`config/sigma_thresholds.json`](config/sigma_thresholds.json) (JSON numbers; gate wiring in-tree is still minimal) - **Benchmark stubs:** `benchmarks/truthfulqa_sigma.sh`, `benchmarks/attention_ab.sh` (SKIP until harness + weights exist) - **FPGA smoke (optional):** `hdl/synthesis/xnor_binding_4096.sv` + `hdl/synthesis/synth_yosys.sh` (Yosys `stat` if installed) **Truth pass:** [docs/WHAT_IS_REAL.md](docs/WHAT_IS_REAL.md) · **Verify:** `make check-v29` --- ## Creation OS LLM architecture (our stack and tiers) This section is the **single map** for “our LLM story” in **this repository**: what is **shipped as C**, what is **wired for external engines**, and what stays **honestly out-of-tree** (weights, `lm-eval` archives, P&R). It is written to pair with the **evidence ladder** ([FIG 03](#publication-hard)) and [docs/CLAIM_DISCIPLINE.md](docs/CLAIM_DISCIPLINE.md). ### What landed recently (v27 → v29 + optional σ spine), in one view | Layer | What it is | Where it lives | Merge gate | |:--|:--|:--|:--| | **Text boundary + tokenizer scaffold** | Tiered tokenizer story (BPE stand-in, byte codebook / XOR–MAJ bundles, optional COSB mmap table, inference trace JSON) | [`creation_os_v27.c`](creation_os_v27.c) + [`src/tokenizer/`](src/tokenizer/) | `make check-v27` (**70**) | | **LM integration shell** | GGUF v3 subset + **tensor-data base** + **mmap reads**; **external engine** stdout capture (`CREATION_OS_BITNET_CPP` + optional stdin/extra argv); **tokenizer.json** vocab counting (`--tokenizer-stats`); **sampling** (temperature / top‑k / top‑p) with **64B-aligned** buffers + **AArch64 NEON** max-reduction; **Llama‑3-ish chat framing**; **loopback HTTP** (`/v1/chat/completions`, `/health`) with **JSON escaping**; **σ toy** on logits; **`make cos_lm`** alias; Docker image builds **v28** without bundling weights | [`creation_os_v28.c`](creation_os_v28.c) + [`src/import/`](src/import/) + [`src/nn/`](src/nn/) + [`src/server/`](src/server/) | `make check-v28` (**29**) | | **Collapse harness (LM “hard parts” without lying)** | **mmap GGUF tensor views** (no multi‑GB `malloc` memcpy); **eight σ scalar channels** + abstention gate on **real-shaped logits**; **XNOR / Hamming-style attention toy** for alternative similarity geometry; **BitNet-shaped forward stub** (deterministic logits for plumbing); **threshold JSON**; **benchmark shell stubs** + optional **Yosys** SV smoke | [`creation_os_v29.c`](creation_os_v29.c) + [`src/import/gguf_loader.c`](src/import/gguf_loader.c) + [`src/sigma/channels.c`](src/sigma/channels.c) + [`src/nn/attention_xnor.c`](src/nn/attention_xnor.c) + [`config/sigma_thresholds.json`](config/sigma_thresholds.json) + [`hdl/synthesis/`](hdl/synthesis/) | `make check-v29` (**22**) | | **σ / agent / silicon labs (v31–v57)** | MCP server, σ decomposition + spec hooks, router/schema, `σ_hardware` + crossbar SV sim, RTL σ-pipeline + ASIC tile drivers, independence / syndrome / threshold story, **σ-guided test-time compute** scaffold, **σ-guided self-play** scaffold, **σ-guided distillation** scaffold, **σ-native inference proxy** (`creation_os_proxy`), **σ-introspection** (`creation_os_v45`), **σ-optimized BitNet pipeline** (`creation_os_v46`), **verification stack** (`creation_os_v47`, `make verify`), **red-team stack** (`creation_os_v48`, `make red-team`), **certification pack** (`make certify`), **v50 benchmark rollup** (`make v50-benchmark`), **v51 integration scaffold** (`creation_os_v51`, `make check-v51` 13/13), **v53 σ-governed harness scaffold** (`creation_os_v53`, `make check-v53` 13/13; structural critique of the Claude Code harness — see `creation.md`, `docs/v53/POSITIONING.md`, `docs/v53/paper_draft.md`), **v54 σ-proconductor scaffold** (`creation_os_v54`, `make check-v54` 14/14; multi-LLM orchestration policy — no network from `src/v54/`; σ as the missing routing signal vs MoA / RouteLLM / MoMA / Bayesian Orchestration — see `docs/v54/POSITIONING.md`, `docs/v54/paper_draft.md`), **v55 σ₃-speculative scaffold** (`creation_os_v55`, `make check-v55` 29/29; σ₃ decomposition + EARS adaptive acceptance + EASD entropy-aware quality gate; NEON hot path; wires Taparia 2603.24967, Sun 2512.13194, Su 2512.23765 — see `docs/v55/POSITIONING.md`, `docs/v55/paper_draft.md`), **v56 σ-Constitutional scaffold** (`creation_os_v56`, `make check-v56` 56/56; rule-based VPRM verifier + σ-gated IP-TTT budget controller + grokking commutator-defect σ-channel + ANE `matmul→1×1 conv` layout helper; NEON hot path; wires VPRM 2601.17223, IP-TTT 2604.06169, SLT grokking 2603.01192 + 2603.13331, 2026 ANE RE; one invariant — *any inference-time self-modification must strictly lower σ* — see `docs/v56/POSITIONING.md`, `docs/v56/paper_draft.md`), **v57 The Verified Agent convergence** (`creation_os_v57`, `make check-v57` 49/49 + `make verify-agent` live aggregate; 5 invariants × 9 composition slots tagged honestly **M / F / I / P**; no new σ math, `static const` registry, no socket; live aggregate dispatches each owning `make` target and reports **PASS / SKIP / FAIL** per slot — never silent downgrades; response to ad-hoc agent-sandbox field — see `docs/v57/THE_VERIFIED_AGENT.md`, `docs/v57/POSITIONING.md`, `docs/v57/paper_draft.md`) | [`#sigma-labs-v31-v40`](#sigma-labs-v31-v40) · [docs/SIGMA_FULL_STACK.md](docs/SIGMA_FULL_STACK.md) | **Not** `merge-gate` — `make check-v31`, `check-v33` … `check-v48`, `check-v51`, `check-v53`, `check-v54`, `check-v55`, `check-v56`, `check-v57`, `make verify-agent`, `check-mcp`, HDL targets in `make help` | | **OpenAI-shaped localhost stub (optional)** | Loopback-only **`/v1/models`**, **`/v1/chat/completions`**, **`/v1/completions`** + **`GET /health`**; deterministic stub strings; **no SSE streaming** (`stream:true` → **501**) | [`creation_os_openai_stub.c`](creation_os_openai_stub.c) + [`docs/LOCAL_OPENAI_STUB.md`](docs/LOCAL_OPENAI_STUB.md) + [`vscode-extension/setup_continue.md`](vscode-extension/setup_continue.md) | `make check-openai-stub` (**5**; **not** part of `merge-gate`) | For a **tier-tagged** “what is real vs imported vs not claimed” table, see [docs/WHAT_IS_REAL.md](docs/WHAT_IS_REAL.md). ### How we think about “our LLM” (three planes, one discipline) - **Plane A — portable spine (this repo’s merge gate):** `creation_os_v2.c` + `core/*.h` teach the **BSC / σ algebra** with **stdlib + libm only**. Standalone programs **v6–v29** are **separate binaries** that extend the same *receipt language* (σ, POPCOUNT/Hamming, abstention patterns) with increasing **integration surface** — still mostly **lab / harness plumbing classes**, not a productized chat server mandate. - **Plane B — MLX / Python paths (extended checkout):** described in [docs/ANALYSIS.md](docs/ANALYSIS.md) and the **AGI map** below — these are **not required** to pass `make merge-gate` here, but they are where full transformer forward passes typically live in practice. - **Plane C — native M4 / heterogeneous dispatch (extended checkout):** NEON/SME/Metal/CoreML style composition appears in project rules and native trees; **this README’s merge gate** intentionally stays **portable C11** unless a target is explicitly optional. **Discipline (why this is “its own level” without magic):** 1. **Claims are typed:** internal consistency checks (**lab demo**), integration shells (**integration / collapse harness**), and future **harness rows** are not interchangeable in prose — see [docs/CLAIM_DISCIPLINE.md](docs/CLAIM_DISCIPLINE.md). 2. **σ is structural, not decorative:** v28 adds σ on **logits-shaped toys** and serving boundaries; v29 adds **multi-channel σ readouts** and a real abstention **gate API** on the same numeric types you would feed from a real LM. 3. **Weights meet memory honestly:** v28/v29 prefer **mmap views** and **small verified fixtures** in the merge gate, instead of pretending CI clones multi‑GB checkpoints. 4. **Hot paths respect silicon habits where we touch them:** AArch64 NEON for toy GEMV / logit max reduction + aligned scratch in v28; extended “all units at once” composition remains **documented** primarily outside this portable gate ([docs/ANALYSIS.md](docs/ANALYSIS.md), [docs/RTL_SILICON_MIRROR.md](docs/RTL_SILICON_MIRROR.md)). 5. **Alternative attention geometry is first-class as an experiment hook:** the **XNOR/BSC-style** attention path in v29 exists to make “softmax vs POPCOUNT-like similarity” a **testable fork**, not a tweet — optional SV/Yosys smoke is intentionally small and local. **Non-goals (still true):** this repository does **not** ship a full in-process **BitNet b1.58 2B4T** forward in portable C as the merge-gate default; it **does** ship the **interfaces + receipts** that let you bolt one on without confusing “compiled” with “evaluated on TruthfulQA”. --- ## Architecture
FIG 08 — module stack (dark editorial). VISUAL_INDEX.
``` ┌─────────────────────────────┐ │ creation_os_v2.c │ │ ~1246 lines · 26 modules │ └──────────────┬──────────────┘ │ ┌─────────────────────┼─────────────────────┐ │ │ │ ┌───────┴───────┐ ┌───────┴───────┐ ┌───────┴───────┐ │ HYPERCUBE │ │ ORACLE │ │ WORLD MODEL │ │ 10 faces │ │ 7-gram │ │ JEPA+Genesis │ │ SOC / Φ≈1 │ │ correlative │ │ Noether = 0 │ └───────┬───────┘ └───────┬───────┘ └───────┬───────┘ │ │ │ └─────────────────────┼─────────────────────┘ │ ┌───────────┴───────────┐ │ BSC CORE │ │ XOR / MAJ / POPCNT(σ) │ │ 4096 bits / 512 B │ └───────────┬───────────┘ │ ┌─────────────────────┼─────────────────────┐ │ │ │ ┌───────┴───────┐ ┌───────┴───────┐ ┌───────┴───────┐ │ SOUL │ │ PROCONDUCTOR │ │ COGNITION │ │ 15 values │ │ 4 profiles │ │ §9–§26 │ │ Crystal Lock │ │ σ₁×σ₂×σ₃ │ │ 18 modules │ └───────────────┘ └───────────────┘ └───────────────┘ ``` --- ## Build Hand `cc` (minimal; flags are yours): ```bash # Any platform cc -O2 -I. -o creation_os creation_os_v2.c -lm # Apple Silicon (M1–M4), native ISA cc -O2 -I. -march=native -o creation_os creation_os_v2.c -lm # Apple Silicon — optional SME experiment (may SIGILL without streaming context) cc -O2 -I. -march=armv9-a+sme -o creation_os creation_os_v2.c -lm # x86-64 cc -O2 -I. -march=native -o creation_os creation_os_v2.c -lm ``` With **Make**, the repo default is **`CFLAGS = -O2 -march=native -Wall -std=c11`** and **`LDFLAGS = -lm`** (see root `Makefile`). Teaching kernel + structural tests: ```bash make help # full target list (labs, RTL, benches) make check # `standalone` + `tests/test_bsc_core` (good before a small PR) make merge-gate # `check` + `check-v6` … `check-v29` (maintainer / CI bar) ``` Flagship **`creation_os_vN`** binaries (each is its own `standalone-vN` + `test-vN`): ```bash make check-v6 # Living Kernel (`creation_os_v6.c`) + `--self-test` (30 checks) make check-v7 # Hallucination Killer (`creation_os_v7.c`) + `--self-test` (35 checks) make check-v9 # Parameters in Silicon (`creation_os_v9.c`) + `--self-test` (41 checks) make check-v10 # The Real Mind (`creation_os_v10.c`) + `--self-test` (46 checks) make check-v11 # MatMul-free mind (`creation_os_v11.c`) + `--self-test` (49 checks) make check-v12 # Tensor mind (`creation_os_v12.c`) + `--self-test` (52 checks) make check-v15 # Silicon mind (`creation_os_v15.c`) + `--self-test` (58 checks) make check-v16 # Unified field (`creation_os_v16.c`) + `--self-test` (66 checks) make check-v20 # Ship mode (`creation_os_v20.c`) + `--self-test` (86 checks) make check-v21 # AGI sovereign stack (`creation_os_v21.c`) + `--self-test` (99 checks) make check-v22 # Twenty colossal insights (`creation_os_v22.c`) + `--self-test` (120 checks) make check-v23 # AGI affordances (`creation_os_v23.c`) + `--self-test` (141 checks) make check-v24 # arXiv echo latches (`creation_os_v24.c`) + `--self-test` (162 checks) make check-v25 # Enterprise pain ledger (`creation_os_v25.c`) + `--self-test` (183 checks) make check-v26 # Global 500 echo orbit (`creation_os_v26.c`) + `--self-test` (184 checks) make check-v27 # v27 tokenizer scaffold (`creation_os_v27.c` + `src/tokenizer/*.c`) + `--self-test` (70 checks) make check-v28 # v28 LM integration shell (`creation_os_v28.c` + import/nn/server helpers) + `--self-test` (29 checks) make check-v29 # v29 collapse harness (`creation_os_v29.c` + mmap GGUF view + σ + XNOR + BitNet stub) + `--self-test` (22 checks) make standalone # build `creation_os` from `creation_os_v2.c` only ./creation_os ``` **Optional (not `merge-gate`):** σ / MCP / M4 / RTL labs — [σ labs (v31→v59)](#sigma-labs-v31-v40), [Composed-decision stack (v60→v95)](#composed-decision-stack--v60--v95-thirty-five-branchless-integer-kernels), `make check-mcp`, `make check-native-m4`, `make formal-sby-v37`, `make verify`, `make verify-agent`, `make red-team`, `make certify`, `make v50-benchmark`, `make check-v51` … `check-v74`, `make check-v76`, `make check-v77` … `make check-v95`; see `make help`. Requirements: C11 compiler + libm. --- ## Limitations This is a research prototype. Specific limitations: - **Oracle** generates text from a 15-sentence corpus via n-gram codebook. It demonstrates that attention can be implemented as σ, not that it matches LLM-quality text generation. - **JEPA learning** is codebook memorization with correlative blending. Energy decreases because the codebook stores training pairs, not because the model has learned to generalize to unseen data. - **GEMM benchmark** compares computational cost of the same geometric task (vector distance) at different precision levels. The 192× ops ratio is measured and real. Whether binary precision is sufficient for a given application is an empirical question. - **Cognitive modules** are BSC implementations of cognitive primitives. They demonstrate that these computations can be expressed in three bit operations. They are not validated against cognitive science benchmarks. - **Living Kernel (`creation_os_v6.c`)** is a **second** program: schematic σ–K–L composition + M01–M18 *toys*. The 30 `self_test` checks are **internal consistency**, not clinical consciousness proof, not COGITATE reproduction, and not a substitute for `make bench` or NEON/HV receipts. See [docs/LIVING_KERNEL_V6.md](docs/LIVING_KERNEL_V6.md). - **`creation_os_v7.c`** is a **third** program: v6 **plus** M19–M23 hallucination-*shaped* σ channels; 35 `self_test` checks. Still **not** measured LM hallucination rates — see [docs/HALLUCINATION_KILLER_V7.md](docs/HALLUCINATION_KILLER_V7.md). - **`creation_os_v9.c`** is a **fourth** program: v7 **plus** M24–M29 stack/silicon-*shaped* σ toys; 41 checks — not tape-out or vendor TOPS/W claims — see [docs/PARAMETERS_IN_SILICON_V9.md](docs/PARAMETERS_IN_SILICON_V9.md). - **`creation_os_v10.c`** is a **fifth** program: v9 **plus** M30–M33 distillation / routing / abstention toys; 46 checks — see [docs/THE_REAL_MIND_V10.md](docs/THE_REAL_MIND_V10.md). - **`creation_os_v11.c`** is a **sixth** program: v10 **plus** M34 matmul-free LM **schematic**; 49 checks — not a trained BitNet-class model or published throughput reproduction — see [docs/THE_MATMUL_FREE_MIND_V11.md](docs/THE_MATMUL_FREE_MIND_V11.md). - **`creation_os_v12.c`** is a **seventh** program: v11 **plus** M35–M37 classical tensor-train / entropy / sequence-head **toys**; 52 checks — not quantum hardware, not TN-LM harness rows — see [docs/THE_TENSOR_MIND_V12.md](docs/THE_TENSOR_MIND_V12.md). - **`creation_os_v27.c`** is an **eighth** program: **M177–M186** vocab / tokenizer / mmap COSB / inference-trace **scaffold** with `src/tokenizer/*.c`; 70 checks — **not** a trained multilingual LM tokenizer product, not FPGA timing proof, not “coherent LM” quality — see [docs/VOCAB_PIPELINE_V27.md](docs/VOCAB_PIPELINE_V27.md). - **`creation_os_v28.c`** is a **ninth** program: **M190–M199** LM **integration shell** (`src/import`, `src/nn`, `src/server`); 29 checks — **not** an in-process BitNet b1.58 2B4T forward, not `lm-eval` rows by itself, not a weights bundle — see [#v28-lm-integration](#v28-lm-integration). - **`creation_os_v29.c`** is a **tenth** program: **v29 collapse harness** (`src/import/gguf_loader.c`, `src/sigma/channels.c`, `src/nn/attention_xnor.c`, `src/nn/bitnet_forward_stub.c`); 22 checks — **not** a downloaded 2B checkpoint, not harness rows by itself — see [#v29-collapse-harness](#v29-collapse-harness). --- ## What this demonstrates 1. **Transformer attention can be implemented as σ** — no matrix multiply required for the similarity computation at the core of attention. 2. **JEPA-style world models work in BSC** — energy-based learning where energy = σ. 3. **Noether conservation holds under symmetric XOR** — a formal invariant, not an approximation. 4. **26 cognitive primitives fit in one ~1.25k-line C file** (`creation_os_v2.c` as of this tree) — the algebra is compact. 5. **The entire architecture runs on any hardware** — no GPU, no framework, no dependencies. 6. **Living Kernel v6** packages cross-domain σ narratives (alignment, RDP, rewind, ghost boot) behind one **executable** gate — useful for thesis structure and for separating *proved in this file* from *cited externally* ([LIVING_KERNEL_V6.md](docs/LIVING_KERNEL_V6.md)). 7. **Hallucination Killer v7** adds **five** more σ-shaped readouts (anchors, association, bluff, context rot, JEPA–Oracle) on the same deterministic gate ([HALLUCINATION_KILLER_V7.md](docs/HALLUCINATION_KILLER_V7.md)). 8. **Parameters in Silicon v9** extends the same gate with M24–M29 stack- and silicon-shaped σ toys ([PARAMETERS_IN_SILICON_V9.md](docs/PARAMETERS_IN_SILICON_V9.md)). 9. **The Real Mind v10** adds M30–M33 distillation, few-shot distance, swarm routing, and max-σ abstention schematics ([THE_REAL_MIND_V10.md](docs/THE_REAL_MIND_V10.md)). 10. **The MatMul-free mind v11** adds M34 — a ternary accumulation + MLGRU **toy** forward path with zero `sigma_matmul` in this file’s definition of “no matmul” ([THE_MATMUL_FREE_MIND_V11.md](docs/THE_MATMUL_FREE_MIND_V11.md)). 11. **The Tensor mind v12** adds M35–M37 — MPS contraction, entropy readout, and sequence-head **schematics** on classical `double` math only ([THE_TENSOR_MIND_V12.md](docs/THE_TENSOR_MIND_V12.md)). 12. **Tokenizer / vocab pipeline v27** ships a **real tokenizer scaffold** (BPE stand-in + byte bundles + optional COSB mmap + inference trace) as **lab demo (C)** — not a shipped multilingual LM tokenizer product ([VOCAB_PIPELINE_V27.md](docs/VOCAB_PIPELINE_V27.md)). 13. **LM integration shell v28** ships **GGUF + mmap I/O + sampling + HTTP + JSON escape + external engine hook** as **integration harness (C)** — merge-gate safe without downloading weights ([§ v28](#v28-lm-integration)). 14. **Collapse harness v29** ships **mmap tensor views + eight σ channels + XNOR attention toy + BitNet forward stub + tier table** as **collapse harness (C)** — explicit about what is still external ([§ v29](#v29-collapse-harness), [WHAT_IS_REAL.md](docs/WHAT_IS_REAL.md)). --- ## Theoretical foundation **Papers & DOIs** (~80, CC BY 4.0): [Zenodo community — Spektre Labs](https://zenodo.org/communities/spektre-labs/). This repository holds the **portable kernel** and measured claims; theory citations and uploads are anchored on **Zenodo** under that community. - Paradigm: Distortion Theory of Intelligence - Core: `K(t) = ρ·I_Φ·F`, `Keff = (1−σ)·K`, `1=1` invariant **External literature and evaluation norms (vetted links, English brief):** **[docs/EXTERNAL_EVIDENCE_AND_POSITIONING.md](docs/EXTERNAL_EVIDENCE_AND_POSITIONING.md)** — Kanerva (binary spatter coding; HDC introduction), Schlegel–Neubert–Protzel (*Artificial Intelligence Review* / arXiv:2001.11797 VSA comparison), EleutherAI `lm-evaluation-harness`; separates **field-level consensus** from **in-repo measurements** (`make bench`, invariants, harness rows in ANALYSIS). **Why the HDC line matters now (literature-backed, no hype):** **[docs/HDC_VSA_ENGINEERING_SUPERIORITY.md](docs/HDC_VSA_ENGINEERING_SUPERIORITY.md)** — Ma & Jiao (2022) HDC vs neural trade-offs; Aygun et al. (2023) encoding survey; Springer AIR HDC classification review (2025); Yeung et al. (2025) robustness estimation; FAISS Hamming / popcount precedent — each row mapped to **evidence class** vs **this repo’s demos**. **Extended narrative:** full three-plane map (llama.cpp + superkernel, MLX, native M4), evidence classes (harness vs microbench vs lab demo), AGI `cos_*` batches, and publication gates — **[docs/ANALYSIS.md](docs/ANALYSIS.md)** (same technical story as this README, with file-level anchors; some paths are forward references when optional trees are not on disk). **Claim discipline (what you may merge in one headline):** **[docs/CLAIM_DISCIPLINE.md](docs/CLAIM_DISCIPLINE.md)**. --- ## Why this wins where it matters (engineering, not slogans) **One geometry for coherence.** In the Creation OS map (see ANALYSIS), σ / Hamming / POPCOUNT is the same language for kernel state, GDA codebooks, oracle prediction, JEPA energy, and native receipt fields. That reduces “coherence as vibes across ten tools” to **one measurable quantity** you can gate on before spending GPU on a full forward pass. **Cost shape.** The reference benchmark is explicit: for the **same 4096-bit task shape**, the GEMM path pays **24,576 multiply-add style FLOPs** in the proxy used here; the BSC path pays **128 bit-ops** (XOR + POPCOUNT per word lane). Memory drops **32×** for the two vectors in the **§7 / `make bench` microbench** pair buffers (`16 KiB` vs `512 B`). Throughput gap is **measured** (`make bench`); the headline **192×** ops and **32×** RAM are **not** host-dependent — they come from the chosen `D` and `W`. **Checkable structure.** §8 shows XOR-sum conservation after symmetric interactions; §4 / §26 show tamper sensitivity on identity chains. That is a different failure mode than silent numeric drift in an unconstrained float pipeline: you get **discrete, replayable** violations. **Deployment surface.** `creation_os_v2.c` plus **`core/*.h`** (same tree; `cc … -I.`) is **stdlib + libm only** — no framework, no CUDA graph, no Python import tax for the teaching kernel. NEON hypervector ops live in headers; the same algebra wires into native / MLX / llama paths in extended checkouts. **AGI-relevant boundary.** This single file does **not** claim benchmark parity with frontier chat models. It **does** show that a broad slice of cognitive primitives (metacognition, ToM, moral compromise, consensus, sleep consolidation, …) can live in **one** small C program built only from XOR / MAJ / POPCOUNT — which is the point of the **26-module** layout: **composition under one algebra**, not a second hidden stack. --- ## AGI map (how this file relates to the full stack)FIG 05 — same planes-abc.svg as under At a glance (kept once above the fold to avoid duplicate renders).
The public **`creation_os_v2.c`** kernel is the **pedagogical spine** (Plane “teaching”: one TU, LOCs quoted in this README). The **production** Creation OS stack (Planes A–C in ANALYSIS) adds, without replacing the algebra: | Plane | Role (summary) | |:--|:--| | **A — llama.cpp + superkernel** | GEMM inference stays here; SK8 superkernel + GDA bridge steer logits and masks with σ / Hamming paths. | | **B — MLX / Python** | Orchestration, receipts, harness vs native evaluation modes, ARC / policy tooling. | | **C — native M4 dylib** | NEON σ batches, optional Metal living weights, dispatcher — `cos_agi*` / `cos_mega*` style primitives for receipts and audits. | **Evidence discipline (from ANALYSIS):** never mix **harness table scores** with **`./creation_os` demo output** in one headline number. Report **two rows** — official harness vs internal native — when comparing to published LLM tables. **Why that matters for AGI work:** long-horizon autonomy needs **contracts** (what was measured, on what hardware, with what receipt). A bit-geometry first pipeline gives you a place to attach those contracts **before** the expensive forward pass — the same design move as “lookup / kernel / transformer last” in Creation OS dispatch rules. --- ## Paradigm shift (what changes — quoted discipline) From the analysis doc: the repository **does not** claim that 4096 bits replace QFT or that MMLU moves without harness runs. **What changes** is engineering + epistemology: | Dimension | Typical LLM-only story | Creation OS map | |:--|:--|:--| | Unit of measure | Loss / logits scattered | **σ / Hamming** one receipt language | | Priority | “Call the big model first” | **Cheap structure first** (LSH, codebook, σ gates) then generation | | AGI primitives | Float Python only | **Native `cos_agi*` / `cos_mega*`** plus optional **4096-bit HV receipts** for audit (`cos_agi_hv_*` family in full tree) | This README’s benchmark table is the **microbench / lab** class; cite it as such next to any frontier row. --- ## Road from this README to production receipts 1. Run **`make test`** and **`make bench`**; archive stdout if you publish numbers. 2. Read **ANALYSIS** sections *Parity program* and *Evaluation modes* before claiming MMLU / ARC parity. 3. Use **`creation_os_v2.c`** as the **portable** artifact for “here is the algebra”; use **Planes A–C** for “here is how it wraps real inference.” 4. Keep **AGPL + dual license** on shipped sources; commercial path stays in `COMMERCIAL_LICENSE.md`. --- ## Publication-hard (what that phrase means here)
FIG 03 — evidence ladder for numbers and headlines. VISUAL_INDEX.
```mermaid %%{init: {'theme':'neutral'}}%% flowchart LR B["Microbench: make bench + host log"] I["Invariants: make test / creation_os"] H["Harness: lm-eval + SHA + model id"] B -.->|separate headline row| H I -.->|separate headline row| H ``` **Not** marketing volume. **Yes** — a standard of argument that many peer-reviewed ML systems papers do not meet on **baseline hygiene**: mixed eval modes, appendix-thin reproducibility, and “task-defined-after-results” tables are common; this repository names those failure modes and blocks them by construction where possible. | Stricter than typical write-ups | How this tree enforces it | |:--|:--| | Baseline separation | Harness vs native vs C demo = **different evidence classes**; ANALYSIS and **CLAIM_DISCIPLINE** require **two-row** reporting when both appear. | | Reproducibility | One TU (`creation_os_v2.c` + `core/*.h`, `cc -I.`) compiles with **stdlib + libm**; invariants print to stdout; `make bench` regenerates throughput on your metal. | | Bounded language | **Limitations** lists what the Oracle, JEPA toy, and benchmark are *not* — no silent upgrade from “demonstrates mechanism” to “beats frontier.” | | Falsifiable core | Algebraic and conservation statements are **discrete**; a counterexample is a log line, not a vague “worse loss.” | | AGI-relevant honesty | Full-stack receipts (`cos_*`, Planes A–C) are mapped in ANALYSIS; this README’s file is the **portable spine**, not the entire production claim. | If a sentence cannot point to **(a)** a line of C, **(b)** a command, or **(c)** an evidence-class label, it does not belong in a “results” paragraph — that single editorial rule is already **stricter than most paper abstracts** in applied ML. **Canonical discipline doc:** [docs/CLAIM_DISCIPLINE.md](docs/CLAIM_DISCIPLINE.md). **Dissertation- and committee-grade map (research questions, contributions C1–C6, threats to validity, suggested chapter outline):** [docs/RESEARCH_AND_THESIS_ARCHITECTURE.md](docs/RESEARCH_AND_THESIS_ARCHITECTURE.md). **Academic citation metadata:** [CITATION.cff](CITATION.cff) (include commit SHA + evidence class when citing numbers). --- ## License This project is **dual-licensed**. The choice between the two licenses is **not** at Your discretion — see [`LICENSE`](LICENSE) §0 for which one binds You — and a third option (paid Commercial License) is available **only** from the Licensor. | Path | Cost | Document | |-----------------------------------------------------|-----------------------|-------------------------------------------------------------------------------------------------| | **Spektre Commercial Source License v1.0** (primary) | free for non-commercial | [`LICENSE-SCSL-1.0.md`](LICENSE-SCSL-1.0.md) | | **GNU Affero GPL v3.0-only** (fallback after 4-yr Change Date, and for any AGPL-derived portions) | free | [`LICENSE-AGPL-3.0.txt`](LICENSE-AGPL-3.0.txt) | | **Commercial License** (closed-source / SaaS / OEM / Sovereign / Strategic) | paid (Startup → Sovereign → Strategic) | [`COMMERCIAL_LICENSE.md`](COMMERCIAL_LICENSE.md) | | **Contributor License Agreement** (binding on every PR author) | n/a | [`CLA.md`](CLA.md) | **TL;DR** - Private individuals · academia · non-profits · journalism · reproducibility/security audits · 30-day commercial evaluation (under EUR 1 M revenue) → **FREE** under SCSL-1.0. - For-profit > EUR 1 M revenue · all hosted SaaS / model-as-a- service / agent-as-a-service (unless You publish the **complete service-stack source** per SCSL §5, including hypervisor, FPGA bitstream and attestation pipeline) · all OEM closed-source redistribution → **paid Commercial License required**. - All government / military / intelligence / law-enforcement **operational** use (SCSL §9.1(b) — targeting, kinetic effects, lethal autonomy, signals-intelligence collection, mass communications interception, biometric mass identification, predictive policing, election-influence operations) → **DENIED at any price**, no commercial license available. Civilian Sovereign deployments by EU CFR / ECHR / ICCPR-bound states are the only government route, under SCSL §9.3. - Sanctioned Persons (EU / UN / OFAC / UK HMT / Finland) and parties credibly accused of Aggression (Rome Statute Art. 8 *bis*) → **categorical denial**, SCSL §10. **Sole holder of all paid commercial rights:** **Lauri Elias Rainio** (natural person, ORCID [0009-0006-0903-8541](https://orcid.org/0009-0006-0903-8541)) **and Spektre Labs Oy**, jointly and severally. **No other person or entity may grant a Commercial License**; any attempted grant is void *ab initio* (SCSL §4.3). Cryptographic License-Bound Receipt (SCSL §11): every Receipt emitted by Creation OS carries the **SHA-256 of `LICENSE-SCSL-1.0.md`**. The pinned reference hash lives in [`LICENSE.sha256`](LICENSE.sha256) and is independently verifiable: ```bash shasum -a 256 LICENSE-SCSL-1.0.md # macOS sha256sum LICENSE-SCSL-1.0.md # POSIX bash tools/license/license_sha256.sh # bundled helper make license-attest # full kernel: 11 KAT + bundle + sample receipt ``` This makes silent forks, license-stripped redistribution, and "laundered through an LLM" derivation patterns **detectable from the published artefact alone**. A Receipt that does not carry a matching `license_sha256` is admissible evidence of a §6 / §8 / §11 breach in any forum. See [`docs/LICENSING.md`](docs/LICENSING.md) for the human-readable explainer, [`docs/LICENSE_MATRIX.md`](docs/LICENSE_MATRIX.md) for the who-may-do-what compatibility matrix, and [`NOTICE`](NOTICE) for copyright, trademark and patent notices. Lauri Elias Rainio · Spektre Labs Oy · Helsinki, Finland ORCID: [0009-0006-0903-8541](https://orcid.org/0009-0006-0903-8541) · licensing: `spektrelabs@proton.me` --- **End of README.** Quick re-entry: [Contents](#contents) · [`make merge-gate`](#run-it-in-sixty-seconds) · [σ labs v31→v59](#sigma-labs-v31-v40) · [Composed stack v60→v95](#composed-decision-stack--v60--v95-thirty-five-branchless-integer-kernels) · [Thirty-five-kernel receipt](#the-thirty-five-kernel-receipt) · [iOS + Android bindings](bindings/README.md) · [DOC_INDEX](docs/DOC_INDEX.md) · [VISUAL_INDEX](docs/VISUAL_INDEX.md) · [LLM vs Creation OS](#llm-vs-creation-os-comparison) · [FIG 09 scan map](#readme-scan-map-fig-09) · [Publication-hard](#publication-hard) · [Claim discipline](docs/CLAIM_DISCIPLINE.md) *2026 · Spektre Labs · Helsinki · Creation OS — coherence you can compile.*