# Krusch Dev — Kevin Ruschman

> AI Systems Architect & Local-First Infrastructure Engineer building deterministic agent control planes, zero-distillation working memory (#1 Official DolphinBench at 91.17% / $24.54 with Gemini 4 Argon Low & #1 Open-Weights Gemma 4 31B at 54.17% / $0.00 across 8,400 traces), sub-15µs CPU cascade routers (#1 RouterArena, ACL 2026 LLMRouterBench & Martian RouterBench), 35-brick certified code substrates (@krusch/toolkit & krusch_toolkit), PostgreSQL 16 ACID state planes, and air-gapped statutory/contract RAG.

Machine-readable full digest with executive TL;DRs and section outlines: [llms-full.txt](https://krusch.dev/llms-full.txt).

## Core Pages

- [Home — 5-Layer Sovereign AI Tech Stack](https://krusch.dev/): Architecture hub covering bare-metal GPU topology, empirical benchmarks, 3-Tier MCP substrates, and open-source systems.
- [Systems Architecture Publications (26 Whitepapers)](https://krusch.dev/articles/): Searchable catalog of all 26 empirical monographs with Technical and Plain-English reading modes.
- [On-Premise Sovereign AI Specifications](https://krusch.dev/sovereign/): Hardware, air-gapped security, and deployment blueprints for enterprise sovereign AI infrastructure.
- [Technical Resume — Kevin Ruschman](https://krusch.dev/resume): Systems engineering background, verified leaderboard records, and production open-source deployments.
- [Engage & Architecture Advisory](https://krusch.dev/contact): Technical consulting, sovereign AI architecture pilots, and engineering inquiries.
- [Return on Memory Preprint (PDF)](https://krusch.dev/papers/return-on-memory-dolphinbench.pdf): 14-page peer-style preprint evaluating 14 controlled 600-task Official DolphinBench runs (8,400 traces across 5 model families).
- [Full Whitepaper Digest (llms-full.txt)](https://krusch.dev/llms-full.txt): Complete abstracts, Plain-English TL;DRs, section outlines, and topic metadata for all 26 whitepapers.
- [Publications RSS 2.0 Feed](https://krusch.dev/feed.xml): Syndicated XML feed of all 26 systems architecture publications.

## Flagship Benchmarks & Open-Source Repos

- [krusch-stack](https://github.com/kruschdev/krusch-stack): Unified 1-command installer bundling 39 annotated MCP tools across 7 pillars, deferred schema loading (-68.5% bytes), and the generalized 2PC Tool-Action Harness.
- [krusch-context-mcp](https://github.com/kruschdev/krusch-context-mcp): 5-verb hierarchical working-memory MCP server (#1 Overall 91.17% & #1 Open-Weights 54.17% on Official DolphinBench; 2,000/2,000 Official ICLR 2026 BEAM 100K–10M).
- [@krusch/toolkit & krusch_toolkit](https://github.com/kruschdev/toolkit): 35 certified JS/TS and Python runtime Lego bricks plus 7 zero-LLM project blueprints enforcing 100% zero-boilerplate compliance.
- [krusch-cascade-router](https://github.com/kruschdev/krusch-cascade-router): Sub-15µs syntactic CPU gate, 8K Arctic centroid router, and live SSE logprob abort (#1 Martian RouterBench, #1 ACL 2026 LLMRouterBench, 0.7883 RouterArena CI).
- [krusch-git](https://github.com/kruschdev/krusch-git): Git DAG indexer, multi-language AST call-graph engine, and 15.25ms PostgreSQL 16 HNSW/GIN hybrid code search validated on 47,998 RepoBench-R queries.
- [krusch-fs-mcp](https://github.com/kruschdev/krusch-fs-mcp): Multi-node OS filesystem page-tree indexer with SHA-256 drift verification and secret-redacted configuration search.
- [krusch-coding-harness](https://github.com/kruschdev/krusch-coding-harness): Bubblewrap sandboxed execution control plane with PostgreSQL 2PC apply journals (lifting Aider Polyglot 2.64x from 19.56% to 51.56%).
- [Jean SRE](https://github.com/kruschdev/Jean-SRE-by-kruschdev): Autonomous 24/7 code audit daemon with Cross-Family Referee Consensus and 18,131 Zero-Leak SFT+DPO distillation trajectories.
- [krusch-law & krusch-nexus](https://github.com/kruschdev/krusch-law): Air-gapped Poppler TSV citation spines (100% citation accuracy on 6,889 LegalBench-RAG queries) and 0.00% False Support Rate statutory precedence graphs.
- [Sovereign G4 Benchmarks Dataset](https://huggingface.co/datasets/krush420/sovereign-g4-benchmarks): 36 SHA-256-verified raw JSON audit logs across 11 external peer-reviewed benchmark suites (317,500+ queries).
- [Sovereign Router Arena](https://huggingface.co/spaces/krush420/sovereign-router-arena): Interactive Pareto cost-versus-accuracy telemetry explorer across routing and memory benchmarks.
- [Sovereign Code Audit SFT+DPO Dataset](https://huggingface.co/datasets/krush420/sovereign-code-audit-sft): 18,131 test-verified, zero-leak SFT and DPO preference rows powering Wave-4 120B, 30B, and 8B QLoRA adapters.

## Track 01 — Sovereign Tech Stack, Cross-Stack Benchmarks, Closed-Loop Audit & Routing

- [The 2026 Sovereign AI Tech Stack](https://krusch.dev/articles/the-sovereign-ai-tech-stack): End-to-end systems architecture blueprint of the 2026 KruschDev tech stack: 5 bare-metal GPUs + 96GB Blackwell G4 burst, 8K snowflake-arctic-embed2, krusch-stack (39 MCP tools, BM25 deferred loading & INV-STACK-10 2PC action harness), 11 hermetic external benchmarks (317,500+ queries, 73/73 SHA-256 gates: #1 Martian RouterBench, #1 RouterArena, #1 LLMRouterBench, #1 91.17% Official DolphinBench @ $24.54, #1 Open-Weights 54.17% Gemma 4 31B & 42.67% local 30B @ $0.00, 0.00% ASR AgentDojo, #1 56.40% ACL 2025 CoIR, 40.20% BIRD-SQL, 66.28% BFCL Live, 51.56% Polyglot, RepoBench-R, LegalBench-RAG & ICLR 2026 BEAM), @krusch/toolkit, and 3-tier 120B/30B/8B QLoRA+DPO distillation.
- [The 15 Production Agentic Systems: Reference Architecture, State-Machine Invariants, and Benchmark Proofs Across a Sovereign Stack](https://krusch.dev/articles/fifteen-production-agent-systems): Complete engineering teardown of the 15 production agentic systems—mapping async 100+ task concurrency, strict 3-bucket context budgeting, Pydantic v2 tool routers, 2PC Verify-on-Stop loops (19.56% to 51.56% Aider Polyglot Pass@2), 72-hour HMAC Slack HITL gates, 0% AgentDojo prompt-injection ASR, $0.50 cost kill-switches, and 5% shadow trajectory comparators to open-source repositories.
- [How an Open-Weights Sovereign Stack Stacks Up Across 317,000+ Benchmark Queries](https://krusch.dev/articles/how-krusch-stack-stacks-up): Empirical cross-stack comparison of krusch-stack across 317,500+ queries on 11 external suites: Official DolphinBench (14 runs / 8,400 traces: #1 Overall 91.17% Argon Low @ $24.54, #1 Open-Weights 54.17% Gemma 4 31B & 42.67% local 30B Wave-4 @ $0.00, 8 queued), RouterArena (#1 CI 0.7883, PR #169 open), ACL 2026 LLMRouterBench (#1 Pareto 0.4269), Martian RouterBench (#1 ID AIQ 0.7488/0.7695 & #1 OOD 0.7268/0.7455 across 72,980 rows), AgentDojo v1.2.2 (19.28% to 0.00% ASR across 1,046 episodes), ACL 2025 CoIR (56.40% #1 NDCG@10 across 2.39M docs & 162,216 Qs), BIRD-SQL Mini-Dev (26.00% to 40.20% EX), BFCL v3 Live (66.28% Micro, -80.67% schema tokens), INV-STACK-10 2PC Action Harness, ICLR 2026 BEAM, RepoBench-R, LegalBench-RAG, and Aider Polyglot (51.56% Pass@2)—with 73/73 SHA-256 clean-room gates and honest frontier gaps.
- [The Closed-Loop Audit & Distillation Pipeline: How We Hydrate Context via 3-Tier MCPs, Gate Fixes with Cross-Family Referees, and Fine-Tune Models on Our Own Stack](https://krusch.dev/articles/how-we-use-jean-for-code-audits): How Jean SRE and the KruschDev audit pipeline combine 3-Tier MCP context hydration (krusch-context-mcp, krusch-git AST graphs, krusch-fs page trees), 11-checkpoint adversarial steel-manning, cross-family Referee convergence gates, and 120B/30B/8B Wave-4 QLoRA + DPO fine-tuning across 18,131 test-verified SFT + DPO records.
- [Multi-Tier Cascade Routing Architecture: Engineering Microsecond CPU Gating, Neural Centroids, and Speculative Hedging for High-Throughput Agent Swarms](https://krusch.dev/articles/multi-tier-cascade-routing): Systems architecture paper: Optimizing enterprise LLM inference budgets using hierarchical multi-tier cascade routing.
- [Postgres as the State Plane: Why Side Effects Require a Transaction Manager, Not a Prompt](https://krusch.dev/articles/postgres-as-operating-plane-for-ai): A language model is a probabilistic sampler; side effects require a transaction manager.

## Track 02 — Working Memory, Return on Memory, BEAM Benchmarks & AST/FS Graphs

- [Return on Memory: 91.17% (#1 Overall) & 54.17% ($0.00 Open-Weights) Across 8,400 Official DolphinBench Traces](https://krusch.dev/articles/return-on-memory-dolphinbench): Across 14 controlled 600-task Official DolphinBench (openai/gpt-5.6-sol) runs (8,400 traces, 5 model families) on byte-identical 0o444 SQLite Raw-RAG checkpoints ($0.00 ingest), Sovereign Raw-RAG + Gemini 4 Argon (Low) hits 91.17% (547/600, #1 Overall @ $24.54, +20.50 pp over #1 Cloud Mem0 70.67% @ $96.21) while Google Gemma 4 31B shatters the 50% open-weights barrier at 54.17% (325/600, $0.00), beating all 8 non-Luna cloud stacks.
- [What 5,535 DolphinBench Runs Taught Us About Local AI Memory](https://krusch.dev/articles/what-dolphinbench-taught-us-about-local-ai): What 5,535 local ablation traces (9 runs) + 8,400 official gpt-5.6-sol traces (14 runs across 5 model families) on DolphinBench taught us about zero-LLM Raw-RAG memory (+12.50 pp Alex, +10.50 pp Riley), dense general-domain post-training (Gemma 4 31B 54.17% @ $0.00 vs.
- [Benchmarking Agent Memory Without Cheating: Closed-System ICLR 2026 BEAM (100K–10M)](https://krusch.dev/articles/iclr-2026-beam-closed-system-benchmark): Closed-system evaluation of krusch-context-mcp (v1.9.0 & v2.0.0) across all 100 conversations (~178.34M tokens) and 2,000/2,000 Official ICLR 2026 BEAM probes (100K–10M) on SQLite inside a 96GB RTX PRO 6000 Blackwell box: 92.58% 100K & 93.00% 10M IE recall, zero E2E drop to 1M (72.86%), and beating cloud past.dev on Contradiction Resolution (4/4 splits) & 10M Multi-Session Reasoning.
- [What Makes krusch-context-mcp Special](https://krusch.dev/articles/what-makes-krusch-context-mcp-special): A comprehensive engineering deep dive into krusch-context-mcp (v2.0.0, 82★): why flat vector RAG fails coding agents, the radical 5-verb discipline (~350 prompt tokens), closed write taxonomy, temporal lineage, hierarchical episode rollups, 2-hop residual bridges, 2,000/2,000 Official ICLR 2026 BEAM validation across 100K–10M (~178.34M tokens, up to 92.58% 100K & 93.00% 10M IE recall) with 8K snowflake-arctic-embed2, zero-Docker SQLite, and krusch-git.
- [krusch-git: Git DAG Indexing, Relational AST Symbols & Hybrid Recency Search](https://krusch.dev/articles/krusch-git-architecture): A deep systems architecture analysis of krusch-git (v1.2.1): indexing Git DAG objects into PostgreSQL + pgvector, zero-dependency AST chunking, multi-hop symbol dependency graphs, exponential temporal decay (e^-0.01t), and hybrid Reciprocal Rank Fusion.
- [krusch-fs-mcp: OS Filesystem Page Memory, SHA-256 Drift Detection, and Multi-Node Fleet Indexing for AI Agents](https://krusch.dev/articles/krusch-fs-architecture): Systems architecture deep dive into krusch-fs-mcp (v0.3.0-alpha): transforming unstructured OS directories, PDFs, DOCX, systemd units, and multi-node SSH server fleets into a deterministic, page-indexed, SHA-256 drift-verified MCP memory substrate powered by KruschNexus Library Mode.
- [Architecting Code-Agent Retrieval & Memory: Beyond Naive Vector Search to Single-Store PostgreSQL 16 Hybrid Systems](https://krusch.dev/articles/code-agent-retrieval-memory-postgres): A practitioner's deep dive into code-agent retrieval architectures — single-store PostgreSQL 16 + pgvector HNSW, structural symbol lexing, Reciprocal Rank Fusion, dependency graph walks, token budget packing, and empirical retrieval evaluations across 12k symbols.
- [Sovereign Context & Safe Mutation Architecture for Coding Agents: Decoupling Tool Density, Stratified Decay, and Transactional State](https://krusch.dev/articles/enterprise-mcp-hub-architecture): Systems architecture reference paper: Decoupling Enterprise Access Gateways from the Sovereign Context & Working-Memory Plane (krusch-context-mcp) and Safe Execution Harness (krusch).

## Track 03 — Certified Lego Substrates, 2PC Sandboxes, Swappable Storage & Personal AI

- [Why Coding Agents Need Certified Lego Bricks, Not Prompted Boilerplate: Inside a 35-Brick, Self-Evolving Code Substrate](https://krusch.dev/articles/lego-system-for-code): How we eliminated LLM-regenerated systems bugs across 100 repositories (3,150+ files) using 35 dual-runtime JS/TS & Python Lego bricks (@krusch/toolkit v1.7 & krusch_toolkit v1.6), 7 zero-LLM blueprints, a 10-rule pre-commit MCP boilerplate guard, a deterministic Rule-of-3 anti-bloat scout (8/8 clusters promoted), and 2-stage QLoRA + DPO calibration (25% to 100% brick adoption).
- [The Krusch Coding Harness: Why Autonomous Code Agents Require a Transaction Manager, Not a Prompt](https://krusch.dev/articles/krusch-coding-harness): A systems architecture deep dive into krusch-coding-harness (v0.1.0): why unconstrained file writing is a flawed paradigm for autonomous agents, pre-commit diff staging, Bubblewrap container sandboxes, storage-engine apply journals, row-locked catalog FSMs, and the architectural trade-offs of a relational state plane.
- [Zero-Friction Agent Substrates: Powering KruschContext & KruschGit with Swappable Local & Polygres Drives](https://krusch.dev/articles/zero-friction-sovereignty-polygres): Codespaces cannot see your local GPU — how AI coding agents eliminate amnesia with ACID decision logs (krusch-context) and navigate code with relational AST graphs (krusch-git)—running on local SQLite or Polygres Cloud without changing agent tools.
- [Local-First Core, Hybrid Cloud Elasticity: The Architecture, Failure Modes, and Trade-offs of Coding-Agent Memory](https://krusch.dev/articles/cloud-native-memory): A practitioner's analysis of coding-agent memory topologies: SQLite edge caching (sub-5ms reads), PostgreSQL durability, embedding parity pitfalls, two-stage HNSW decay indexing, and the realities of data sovereignty.
- [Introducing Krusch Companion: A Sovereign, Voice-First Personal AI Assistant & Dual-Axis Daily Execution Engine](https://krusch.dev/articles/introducing-krusch-companion): Complete guide to Krusch Companion (v1.2): Why you need a 100% local-first personal AI assistant, how the Dual-Axis Priority Board solves the Priority Paradox, voice-to-text capture, the 14-tool holodata agentic proxy, step-by-step setup, and a practical playbook for daily efficiency.
- [Introducing G-Krusch: A Sovereign Google Workspace Hub, Hybrid Drive RAG & MCP Bridge](https://krusch.dev/articles/introducing-g-krusch): Complete architectural deep dive into G-Krusch: a sovereign Google Workspace Hub and 12-tool MCP server with cloud-synced agent.md steering, 1,024-dim hybrid Drive RAG, 3-tier PDF extraction, and AES-256-GCM security hardening.

## Track 04 — Air-Gapped Document Ingestion, Statutory Precedence & Contract Intelligence

- [Citation Spine, Statutory Graph & Working Memory: Systems Architecture Note on Bounding Generative Error in Legal Workflows](https://krusch.dev/articles/the-sovereign-triad): A systems architecture note on bounding generative error in legal RAG: layout-true ingestion geometry, deterministic jurisdiction filtering, claim-level entailment, and episodic memory hygiene.
- [Citations Die in Local RAG: Why We Built KruschNexus and What It's For](https://krusch.dev/articles/why-we-built-krusch-nexus): Ingestion destroys locator fidelity; this is the spine we built to keep page and span coordinates intact.
- [Why We Built KruschLaw: Local-First Legal RAG, Ordinance Intelligence, and Grounded Statutory Analysis](https://krusch.dev/articles/why-we-built-kruschlaw): Local-first statutory intelligence: 0.00% False Support Rate assertion grounding, versioned statutory precedence DAG with temporal time-travel, tenant defense checklists, and cryptographic purge.
- [Why Semantic Similarity Isn't Governing Precedence: What Makes KruschBiz Different From Generic Business RAG](https://krusch.dev/articles/why-krusch-biz-differs-from-business-rag): Vector search fails on enterprise contracts because cosine distance has zero legal precedence.
- [Authority Packs: Controlled Rulebooks for Enterprise AI — Sovereign Domain Scoping vs. Vector Bleed](https://krusch.dev/articles/authority-packs): Why enterprise buyers don't buy retrieval — a systems architecture and commercial blueprint for Authority Packs: bounded, versioned rulebooks, span-grounded slots, and preemption graphs.
- [The Ingestion Imperative: Why Retrieval Quality Is Bounded Before the Model Reads a Single Word](https://krusch.dev/articles/the-ingestion-imperative): A systems architecture note on document engineering, vector embedding space geometry, measured failure modes of ingestion, and the failure modes of the structured antidote.