The Moat

A model trained on five years ofreal-world execution.

LOCI's execution guardian doesn't rely on rules or heuristics. It's powered by LCLM — small, specialized code language models trained on billions of ASM blocks and real-time traces from open-source projects spanning IoT, networking, safety-critical, and industrial systems. ~220× cheaper per query than general LLM tokens. It reads binary files. It predicts before code runs.

How LOCI's model is built — and what it runs on

Inputs

5 Years of Data

Real-time traces from open-source projects

CPU & GPU traces

Binary Input

ELF · Mach-O · PTX · Wasm

Any compiled target

Execution Model (LCLM)

Trained on real hardware behavior

Not heuristics. Not rules.

Outputs

5 Signals

Timing · Throughput · Stack · Power · Risk Paths

Fires before code ships

Agent-Ready

Ground any AI coding agent

Higher first-pass accuracy. Fewer debug cycles.

Not logsNot samplingNot static analysisReal execution traces5 years · open-source projects
Data Foundation

Five Years of Open-Source Execution Traces

LCLM is a small, specialized model built on something no heuristic can replicate — five years of real-time traces collected from open-source projects running on real hardware. Billions of ASM blocks across IoT, networking, safety-critical, and industrial systems. ~220× cheaper per query than general LLM tokens.

  • Real-time hardware performance counters
  • CPU and GPU execution traces from open-source projects
  • Timing, power, and throughput measurements at instruction level
  • Collected across diverse workload types and hardware generations
Why this is different
Most tools are built on synthetic benchmarks or hand-crafted rules. LCLM is trained on real open-source software, running on real hardware, over five years. No customer binaries. No proprietary code.
Binary-First

Binary Files as Input — Not Source Code

LOCI reads compiled binaries directly — ELF, Mach-O, PTX/SASS, and Wasm. The source language is an input, not a constraint.

  • ELF — C, C++, Rust, Go, Zig
  • Mach-O — Swift, C/C++ on Apple platforms
  • PTX / SASS — CUDA kernels and GPU compute
  • Wasm — WebAssembly targets
  • JVM bytecode, CPython bytecode
Key insight
The binary encodes how code will execute — control flow, instruction sequences, memory layout — without needing to run it.
The Model

LCLM — A Small, Fast Execution Model Trained on Reality

LCLM is a small, specialized model trained on real execution behavior — not a general LLM. ~220× cheaper per query. Predictions reflect how hardware actually runs code, not how engineers think it should. Not a rule engine. Not static analysis.

  • Trained on measured instruction throughput and latency
  • Learns branching, divergence, and memory hierarchy effects
  • Models CPU/GPU scheduling behavior from real workloads
  • Trained on boundary measurements — upper and lower execution bounds per function
  • Outputs are numeric, physically meaningful values (ms, watts, instructions/cycle)
  • No hallucination — signals are bounded by binary structure, not generated from text
No hallucination — by design
LLMs generate tokens. LCLM predicts within measured execution bounds. Every signal has a floor and a ceiling derived from real hardware traces — a value outside those bounds is structurally impossible.
5 Signals

Five Execution Signals — Before the Code Runs

Every signal is a prediction from the model — fired from the binary, available before a single test runs or a line ships.

  • Timing & Latency — response time and p95/p99 bounds per function
  • Throughput — instructions and operations per cycle
  • Stack & Memory — worst-case stack depth and memory pressure
  • Power & Energy — per-function energy cost estimate
  • Risk Paths — CFI, worst-case execution paths, and attacker-reachable branches
When it fires
As you code (incremental .so), after full build, during test, and at PR merge — each stage independently useful.
Guardian Agent (Gartner: Multiagent Systems)

Execution Guardian for Any AI Coding Agent

AI coding agents reason from source code alone — they have no sense of how code actually executes. LOCI is the guardian agent that fills that gap — shifting execution signals into the workflow before code ships.

  • Preflight audits the agent's plan before code is written
  • Post-edit catches quality regressions after code changes
  • Guard every PR — approve or block with evidence
  • Works with Claude Code, Cursor, and Copilot via MCP. Available for 15+ agents via skills.sh. Listed on AWS Marketplace.
The outcome
Higher first-pass accuracy. Lower token burn. Human-on-the-loop — you review and approve.
Zero Overhead

No Instrumentation. No Runtime Required.

LOCI runs entirely from the binary artifact. No agents to deploy, no profilers to configure, no runtime hooks.

  • No code changes or annotations required
  • No runtime overhead — analysis happens offline
  • No new build steps — reads existing CI artifacts
  • Works on object files (.so) during development, full binaries at build
Incremental

Signals From the First Line Written

LOCI doesn't wait for a full build. It compiles incrementally — isolated object files per function or module — so signals are available as code is written.

  • Analyzes small .so files compiled per function or module
  • Similar to Compiler Explorer — continuous, lightweight feedback
  • fn-level signal while you type, full program signal after build
  • No binary yet? No problem — LOCI fires from the first compiled object
Analogy
Think of Compiler Explorer — but instead of showing assembly, LOCI shows execution signals: timing, throughput, stack, power, risk.
Production-Grade

Built Also for Performance-Critical Systems

LOCI works across any compiled target — and is especially suited for teams where performance, power, and correctness are non-negotiable: AI inference, networking, HPC, embedded, and data center workloads.

  • AI training and inference systems (LLM, diffusion, vision)
  • Networking and infrastructure software
  • High-performance computing and scientific workloads
  • Embedded and edge systems with power constraints
Principle
Correctness, predictability, and trust over speculative reasoning. Real data. Real hardware. Real signals.

A guardian agent that runs from the compiled binary with zero overhead. But if it's autonomous — who guards the guardian?

Who guards the guardian?

Physics, verifiability, and human oversight.

  • Deterministic

    Same binary in, same signals out. Every time. No probabilistic variation, no prompt sensitivity, no temperature knob. The output is a function of the binary — not a guess.

  • Bounded by physics

    Every prediction has a measured floor and ceiling from real hardware. LCLM cannot hallucinate a value outside observed execution bounds.

  • Verifiable on hardware

    Every prediction can be validated by running the binary on real hardware. No black box — inspect, compare, confirm.

  • Trained on reality

    Five years of real execution traces from open-source projects on real CPUs and GPUs. Not synthetic benchmarks, not hand-crafted rules.

  • Human-on-the-loop

    Decision layers are opt-in. Start with full human review, hand off to agentic when you're ready. Every level keeps a full audit trail.

  • 120+ granted patents

    The core technology is peer-reviewed through the patent process and defensible across US, EU, Japan, and additional markets.

You decide when to let go

From human-on-the-loop to fully agentic.

Not in-the-loop — that pauses the agent on every step and doesn't scale. On-the-loop means LOCI runs automatically and the human reviews the verdict, not the code. Each decision layer is opt-in. Hand off when you're ready.

1
Full human reviewDay one default

LOCI surfaces verdict on every PR. Human approves or blocks each one.

2
Threshold-based alertsAfter calibration

Set regression budgets. LOCI auto-passes PRs within budget, flags only exceptions for human review.

3
Auto-guard with overrideTrusted teams

LOCI auto-approves or auto-blocks based on your quality contract. Human can override any decision.

4
Fully agenticFull handoff

LOCI runs as an autonomous guardian inside the agentic workflow. Human is notified, not required. Full audit trail retained.

Every level keeps a full audit trail. Every decision is logged, traceable, and reversible. You choose the level — per repo, per team, per signal.

Ready when you are

Start guarding your code. First verdict on your next commit.

Five signals. One execution guardian. Install in minutes — no instrumentation, no runtime overhead.


LCLM · Tech Sheet

The model under the gate. By the numbers.

Reference card for the execution model that powers every LOCI signal.

Architecture
Small, specialized code-language model · purpose-built for compiled binaries
Foundations
FlashAttention IO-aware kernels · ModernBERT encoder backbone · domain-tuned for assembly
Training corpus
Massive assembly blocks · real-CPU/GPU execution traces · growing
Supported ISAs
ARM Cortex-A/M/R · RISC-V (RV32, RV64) · x86_64 · NVIDIA SASS · AMD GCN
Output signals
WCET · worst-case stack depth · throughput · power/energy · memory
Verification
Predictions match measured hardware — not confidence scores or probabilities
Per-query cost
~220× cheaper than general-purpose LLM tokens (internal benchmark)
Runtime requirement
None · no instrumentation · no profiler · no GPU run required
Deployment
SaaS · on-premises · air-gapped · AWS Marketplace VPC
Compliance
ISO 27001 certified · SOC 2 in progress
IP
120+ granted patents (US · EU · JP) · five years of training on real workloads
ISO 27001 Certified
Try the LOCI app