Scalar builds deterministic AI orchestration systems —
the execution-verification layer that proves an action is safe *before* it runs.
Solo-built. Two patent applications filed (KR). Verified in the open.

Seung-hyeon Lee — Founder, SCALAR (incorporation in progress) · Seoul, Korea · l0architect@scalar-inc.com

Orchestration, robotics, finance, copyright, kernel verification — every project answers one question: can you trust the result of an execution? Verify before you trust.
_{오케스트레이션·로보틱스·금융·저작권·커널 검증 — 모든 프로젝트가 답하는 하나의 질문: 실행 결과를 신뢰할 수 있는가. 신뢰하기 전에 검증한다.}

LED manufacturing R&D  →  overseas auto production line (full-lot quality inspection)  →  ERP · logistics · call-center CS  →  Korea's largest manufacturing floors (Hyundai · Kia)  →  independent development  →  2 patent applications · LKML RFC · open source

Across every floor — production lines, quality inspection, logistics, and customer support — the same problem kept surfacing three ways: equipment broken by an unverified command, customer trust broken by an unverifiable answer, audits impossible on an unrecorded execution. Scalar is the answer, built from that floor up.
_{생산라인·품질 전수검사·물류·CS까지 — 전 공정에서 같은 문제를 세 번 만났다: 검증 없는 명령이 부순 설비, 검증 불가한 답이 무너뜨린 고객 신뢰, 기록 없는 실행에 불가능한 감사. 스칼라는 그 현장에서 출발한 답이다.}

Enterprises don't stall on AI because the models are weak. They stall because no one can prove what an action did — so every AI action carries a trust cost: human review, audit, rollback, liability. Verification removes that cost. It is the layer that turns an impressive demo into something deployable in production — and the spend is moving from "a better model" to "an action you can prove."
_{AI 도입이 막히는 건 모델이 약해서가 아니라 "무엇을 실행했는지 증명할 수 없어서"다 — 검토·감사·롤백·책임이라는 신뢰비용. 검증은 그 비용을 제거해 데모를 프로덕션으로 바꾼다. 자본은 "더 좋은 모델"에서 "증명 가능한 실행"으로 이동 중.}

"Agents don't fail in production because the model is wrong. They fail because no one can prove what an agent did." — Alex Salazar, Arcade.dev CEO

Where Scalar sits — the execution-guarantee layer, not another framework:

L1  Orchestration   LangChain · CrewAI            connect models & tools          →  complement
L2  Guardrails      Guardrails AI · NeMo          output format / schema check     →  adjacent
L3  Deterministic   PromptPort · Lean-Agent        prove constraints BEFORE run     →  Scalar combines
    execution       SCALAR MTP-L2                  · reproduce · audit                 routing + execution
                                                                                       verification

Scalar combines a deterministic L3 execution guarantee with multimodel routing — two patent applications filed (KR) and the same verification methodology submitted to the Linux kernel (LKML RFC). Honest weakness, stated up front: as a new entrant, ecosystem and mindshare are early-stage.
_{L1 연결·L2 형식검증이 못 푸는 "실행 전 결정론적 보장(L3)"을 멀티모델 라우팅과 결합 — 특허 2건 출원(KR), 같은 검증 방법론을 리눅스 커널에 RFC로 제출. 정직한 약점: 신생 팀이라 생태계·인지도는 이제 시작.}

A local, single-user, BYOK AI orchestration app (Tauri 2). Every model call is deterministically routed and verified at the boundary before it leaves the device — each egress commits an AdmissionRecord to an Ed25519-signed local audit hash-chain (proof candidate). Your keys, conversations, and files stay on your machine; Scalar's servers can't see inside. Same routing patent application and Kani-proof methodology as MTP-L2, applied to a consumer desktop app.
_{로컬 단일유저 BYOK AI 오케스트레이션 macOS 앱. 모든 모델 호출은 기기를 떠나기 전 경계에서 결정론적으로 라우팅·검증되고, Ed25519 서명 로컬 감사 해시체인에 기록됩니다. 키·대화·파일은 전부 기기 안에 — 서버는 들여다보지 못합니다.}

A deterministic layer between the LLM and the executor: tokenize → validate → clamp → compile. Every downstream command is schema-conform and within physical bounds by construction — not by hoping the model behaves.

Patent applications (KR, filed):

• Routing Layer — KR 10-2026-0105850 (reasoning-profile vector routing)
• Execution Layer — KR 10-2026-0084899 (DSL parser-verified execution)

Input → [Routing Layer (application ①)] → [Execution Layer (application ②)] → Verified Output

Routing and execution are covered by two separately filed applications.

Fair benchmark: identical fuzzed input to both paths, reproducible offline (python kibo_fair_test.py --n 2000). Scope: 100.00% is bounds/schema conformance by construction; 90.9% is intent match — separate properties. MTP-L2's edge is determinism & safety, not latency.

→ Repo + benchmark · Live Demo

Real-time BTC/USDT derivatives + on-chain + whale tracking. The signal engine runs continuously on Hetzner VPS; trading and auto-tuning currently run in shadow mode (signals computed, logged, and evaluated — not live-applied) while the public endpoint is being rewired.

Hybrid audio-similarity engine for AI-music copyright disputes, producing forensic reports for B2B legal use.

A complete 16-volume psychological-SF novel released as a source-available interactive novel — the deterministic philosophy applied to storytelling: same choices, same story, every time. No AI, no server, no build, zero runtime dependencies. The full manuscript ships in the repo; non-commercial culture is free, commercial use is licensed.
_{완결 심리 SF 장편을 오픈 IP 인터랙티브 노벨로 공개 — 비영리 향유는 자유, 상업적 이용(출판·굿즈·영상화)은 라이선스.}

Four short stories (비효율 · 기억 라우터 · 차이 · 입술의 무게) running on the same SNZ 5-layer engine — proof the engine is a reusable noncommercial substrate any author can build their own interactive novel on. Korean · English · Japanese, with theme-native interactions (silence as record, manipulated choice, refrains as echo).
_{SNZ 엔진을 그대로 재사용한 결정론 단편집 — 엔진은 비영리 작가 누구나 자기 작품에 쓸 수 있는 토대(PolyForm Noncommercial). 한·영·일 3개 언어.}

A cross-domain study of when verification should stop — the methodology behind the deterministic-verification work, made falsifiable. A non-refutation overhead boundary with a provable risk bound (miss_rate ≤ ε), and one assumed-not-earned law tested across number theory, cosmology, control, statistics, and LLM evaluation. Pre-registered; reports its own negative results (headline: windowed IDE falsified, MTP-MetaEval survived). DOI-archived on Zenodo.
_{검증을 언제 멈춰야 하는가 — 비반증 오버헤드 경계의 위험 한계 증명과 assumed-not-earned 법칙, 5개 도메인 교차 검증. 사전등록·정직한 negative 보고. Zenodo DOI 아카이브.}

Rust model of the Linux BPF verifier's scalar abstract domain (tnum × cnum) with Kani soundness proofs and differential tests against the unmodified kernel C (N=2000 × 11 ops, 0 mismatches). Companion LKML RFC:

RFC      : [PATCH RFC] bpf: add DAG fast-path in verifier to skip redundant state pruning
Sent to  : bpf@vger.kernel.org · linux-kernel@vger.kernel.org        2026-05-29
Reply    : Alexei Starovoitov (BPF subsystem maintainer · Meta) — direct reply in 17 hours
Language : Rust (model + proofs) · C (kernel patch)

→ LKML thread 2026-05-29

Windowed interacting-dark-energy toy model fit to real DESI DR1 BAO data (emcee MCMC). Honest negative result, reported as such: coupling bounded at β₀ < 0.27 (95%), no advantage over ΛCDM (Δln Z = −1.9) — the falsification that MTP-MetaEval's stopping rule was tested against.
_{DESI DR1 BAO 실데이터 피팅 — ΛCDM을 이기지 못했다는 정직한 negative result. MetaEval 중단 규칙의 실전 반증 사례.}

Monetary issuance anchored to verifiable physical energy. Unauthorized inflation structurally impossible.

Simulation (seed=42): Conservation 100%  ·  False issuance blocked 53.7%  ·  SNR +94.2%

Full BOM/production tracking · 41 functions · Zero infrastructure cost.

Two-domain stress test: music industry vs. social discourse. When μ > 1.0, echo chamber lock-in is the rational response.

Cross-domain Δβ: +0.141 / +0.135  ·  Music golden-age: 8% ↓  ·  Social: 98% intact
(v0.5 snapshot — structured argument, not a proof; author-reported limits in-repo)

Month 2  →  Multi-LLM production system       FastAPI · pm2 · Tailscale VPN · Hetzner
Month 4  →  LKML BPF DAG verifier RFC         Alexei Starovoitov · 17h direct reply
Month 6  →  7-project portfolio               2 patent applications filed · funding track active
Month 7  →  SCALAR Atelier RC                 notarized macOS build · Kani-proven routing · 700+ tests

 

SCALAR · Seoul, Korea · 2026 — incorporation in progress