OmniNode Edge Fusion (v1.0.0)

Universal, domain-independent OmniNode-encased multi-sensor assimilation, zero-allocation deterministic microkernel guarding, and progressive Matryoshka local-to-remote conformal cascades with a zero-dependency Model Context Protocol (MCP) JSON-RPC 2.0 server.

All operational dimensions, sheaf ranks, Koopman spectral counts, conformal miscoverage targets, static ring-buffer capacities, and progressive feature widths are resolved from a declarative multi-dimensional Configuration Settings Grid (config/edge_fusion_grid.json) enforced mechanically by ZeroHardcodedValuesAstGate and DomainAgnosticNamingGate.

Core Architecture

1. Canonical OmniNode Shell (src/omninode_shell.py)

• Zero-Execution Static CARD Discovery: Brick capabilities, contracts, grid axes, and canonical failure modes (transient, bad_input, capability, data, contract, budget) are extracted from top-level CARD = {...} literals via ast.parse + ast.literal_eval without importing or executing untrusted module code.

• Content-Addressed .omni/cas Store & Cold Lockfile Rebuild: Source artifacts and normalized grid/1 configuration cells are pinned in lockfile.json, enabling deterministic 0-search cold rebuilds into empty directories with automatic refusal on tampered artifacts.

• Algebraic Composition Operators:

  • >> — Sequential envelope pipeline with typed failure short-circuiting.

  • // — Typed-failure waterfall fallback across declared failure classes.

  • | — Parallel fan-out with cellular sheaf coboundary consistency checks across shared output stalks.

  • @ — Disjoint maker_id != checker_id independent check gate (deterministically refuses self-checks).

2. Spatiotemporal Cellular Sheaf, Koopman DMD & Adaptive Conformal Brick (src/bricks/spatiotemporal_sheaf_fusion.py)

• Bitemporal (valid_time, tx_time) Reconciler: Filters multi-rate asynchronous streams strictly to tx_time <= as_of_time, guaranteeing 0.0 future lookahead leakage even when backdated records arrive later.

• Cellular Sheaf Coboundary & Harmonic Projection: Constructs the stalk-space sheaf Laplacian $L_{\mathcal{F}} = \delta^\top \delta$ across overlapping sensor frames, isolating spoofed streams (AUROC = 1.000) and cutting reconstruction MSE by 99.77% under simultaneous sensor spoofing and dropouts.

• Exact Koopman DMD + Adaptive Conformal Inference (ACI): Extracts dominant spectral eigenvalues $(\lambda_k, \omega_k)$ for multi-horizon forecasting while updating $\alpha_t$ online via Gibbs-Candes ACI, maintaining 92.89% empirical interval coverage across non-stationary regime shifts (where static calibration drops to 18.06%).

3. Deterministic Zero-Allocation Microkernel Guard (src/bricks/deterministic_microkernel_guard.py)

• Emits self-contained, zero-heap (0 dynamic allocation calls) ISO C99 source (gcc -std=c99 -O3 -Wall -Wextra -Werror -pedantic) and #![no_std] #![forbid(unsafe_code)] Rust library source (rustc --crate-type lib --edition 2021).

• Executes stateful fixed-capacity ring-buffer drift checks, cross-patch sheaf energy checks, and conformal logit margin guards in ~0.13 µs/step with < 6e-8 float32 discrepancy across Python, C99, and no_std Rust.

4. Conformal Hybrid Cascade, In-Memory WebAssembly, WGSL & MCP Server (src/bricks/conformal_hybrid_cascade_mcp.py)

• In-Memory IEEE-754 WebAssembly & W3C WGSL: Compiles 16d fast Matryoshka weights into raw WebAssembly binary bytes (1,367 bytes, 0.0 float32 diff in Node.js WebAssembly.instantiate without writing binary files to disk) alongside a portable W3C WGSL compute shader.

• Split-Conformal Cascade: Resolves singleton local prediction sets (|C_alpha(x)| == 1) locally with 0 remote tokens and escalates multi-label ambiguous sets to the full 64d tier — reducing remote calls and tokens by 87.08% while lifting accuracy from 92.08% (standalone local) to 96.81% at 96.81% empirical set coverage.

• Zero-Dependency MCP JSON-RPC 2.0 Server: Exposes omninode_fuse_streams, omninode_audit_cross_spec_contradictions, omninode_cascade_route, and omninode_fenced_mutate (optimistic CAS lease_epoch fencing + maker_id != checker_id).

Related MCP server: Living Stack MCP — Retired Beta Listing

Example Workload Mappings (Using Identical Universal Primitives)

Because every module operates on domain-agnostic stalks, envelopes, arenas, and conformal sets, the same OmniNode pipeline serves diverse operational settings without renaming or specializing code:

• Multi-Instrument Earth & Atmospheric Observation: Asynchronous multi-resolution sensor reconciliation (valid_time, tx_time), faulty sensor localization, and Koopman + ACI regime-shift intervals.

• Hard Real-Time Embedded Control Loops: Static ring-buffer ISO C99 and #![no_std] #![forbid(unsafe_code)] Rust microkernel guards with sub-microsecond deterministic execution and zero heap usage.

• Client-Edge Progressive Inference: Browser/device WebAssembly + WebGPU (WGSL) 16d local fast path escalating uncertain envelopes to 64d remote capacity only when |C_alpha(x)| > 1.

• Cross-Artifact Specification Auditing & Fenced State Transitions: MCP JSON-RPC 2.0 cellular sheaf contradiction detection across requirement, ticket, and configuration envelopes plus lease_epoch-fenced mutations.

Quickstart

pip install -e ".[dev]"
pytest -q
python3 -m src.cli

License

Apache License 2.0 (LICENSE).