Show HN: Cerberus MCP – Flash and sandbox beats Argon 100% vs. 50% on ARC-AGI-2

Zero-dependency, Cython-compiled native .pyd/.so MCP governance engine, 3D SNN) memory substrate, and Level-4 conjecture & tensor co-processor for Claude Code, Cursor, Gemini CLI, and VS Code Agent Mode.

pip install https://github.com/MyceliaCognition/cerberus-agi-kernel/releases/download/v7.4.0/cerberus_agi_kernel-7.4.0-py3-none-any.whl

pip install https://github.com/MyceliaCognition/cerberus-agi-kernel/releases/download/v7.4.0/cerberus_agi_kernel-7.4.0-cp314-cp314-win_amd64.whl

Running cerberus-mcp init inside any repository provisions all agent hooks and automatically starts your 7-day bundle:all_access trial:

cerberus-mcp init

• .mcp.json — Claude Code & Universal MCP Server registration ( cerberus-mcp-server --transport stdio )
• .claude/settings.json — Stop &SubagentStop lifecycle hooks (cerberus-mcp stop-guard )
• .cursor/hooks.json — Cursor stop hook (loop_limit: 5 )
• .gemini/settings.json & GEMINI.md — Gemini CLI MCP server + AfterAgent stop-guard hook
• .vscode/mcp.json — VS Code Native Agent Mode stdio MCP server

After completing checkout via any Stripe Payment Link below, you are automatically redirected to the Cloudflare License Claim Portal (/claim?session_id=...) with your signed offline CERB1. token:

cerberus-mcp activate CERB1.<payload>.<sig>
cerberus-mcp status
cerberus-mcp ci-gate --workspace .

Latency-Encoded Leaky Integrate-and-Fire (LIF) Spiking Neural Network: Compresses full repository & episodic corpora into 32-dimensional deterministic spike-latency fingerprints at a hardware-verified**71,029.23:1 compression ratio** (9,091,742 bytes$\rightarrow$ 128 bytes). #

Two-Tier mmap Retrieval: Combines L1-resident centroid routing with zero-copy on-demand$O(1)$ mmap paging in**249.50 µs** .

• Zero-Backpropagation Metabolic Plasticity (cerberus_eval_carbon_economy_plasticity): Dynamically executes synaptic autophagy on low-utility connections and synaptogenesis along high-surprise causal pathways with**0 backpropagation passes** .

• Sub-Millisecond Cross-Node State Transfer (cerberus_serialize_snn_spore_print): Serializes and reconstructs educated SNN weights, refractory thresholds, and epistemic priors via compact JSON"Spore Prints" in**0.5210 ms** (< 1.0 ms SLA).

Algebraic Prompt-Injection Immunity (cerberus_audit_struq_membrane): Enforces a deterministic StruQ attention-masking membrane that clamps cross-boundary attention logits between trusted system instructions and untrusted quarantined data tokens to**0.0 cross-boundary attention weight)** , paired with immediate Security-Tax synaptic autophagy on injection attempts.$-\infty$ (

4. Unassisted Hoare Boundary Slicing MCTS, Calcium-Wave Controller & Active Experimental Design (v7.3–v7.4)

• 1-Shot Maximal-Disagreement Experimental Design (cerberus_design_maximal_disagreement_experiment): Synthesizes boundary probes where competing abductive hypotheses maximally diverge, achieving a verified**9.25x sample-efficiency speedup** over random/grid exploration.

• Minimum Description Length (MDL) Law Distillation ( cerberus_audit_mdl_novelty_gate): Compresses 24 discovered domain laws by**-69.79% MDL bits** while preserving 100% holdout falsification accuracy, governed by theCalcium-Wave Ring Temperature Controller (cerberus_step_calcium_wave_controller ).

• Certified Non-Commutative Matrix Multiplication Tensors (cerberus_hunt_tensor_cse_bounds): Discovers and certifies** 9 exact int64 ternary bilinear tensors** (coefficients in{-1, 0, +1} ) across**10,753 exact Brent equations** with**0 L1 residual** , including**<2,4,4> Rank-26** (vs. 32 naive) and**<2,5,5> Rank-40** (vs. 50 naive).

• Signed Common Subexpression Elimination (CSE): Achieves a verified**-51.16% fleet-wide addition reduction** (538 naive additions reduced to263 hardware-scheduled additions).

Strict Blind-Vault & Zero-DSL-Hint Audit (verify_arc_blind_vault_isolation = True, verify_zero_dsl_code_hints = True): Evaluates gemini-3.8-flash (thinking_level="medium" , Provider-Adapterthought_signature state continuity + Cerberus MCP Sandbox) head-to-head against standalonegemini-4-argon on public ARC-AGI-2 evaluation tasks (benchmarks/cerberus_v7_5_arc_flash_vs_argon_report.json ). #

Contender B (gemini-3.8-flash Medium + Provider-Adapter thought_signature + Cerberus MCP Sandbox): Achieves**8 / 8 (100.0%) official pass@2 accuracy** ,100.0% train-pair zero-error verification ,174 lines of pure standard-library Python synthesized from scratch (0 external DSL helper calls,47,292 bytes of preservedthought_signature state across10 blocks), at**$0.001890 per solved task** . #

Contender A (Standalone gemini-4-argon Frontier Model): Achieves**4 / 8 (50.0%) official pass@2 accuracy** —failing05f2a901 (99.1% pixel match),3906de3d (98.0% ),00d62c1b (97.8% ), and0a938d79 (70.4% ) due to unverified 1-shot grid transcription errors—at**$0.049721 per solved task (26.3x higher cost per solve)** .