The technology

Why LOCI thinks ahead.

LOCI predicts software behavior from real workloads and platform traces — not source code alone.

Inputs

Real workloads
Platform traces
Compiled binaries
↓

The moat

8 years

of labelled execution data

Real hardware traces, labelled against the binaries that produced them. That is what the model learned from, and why it can predict a change before it runs.

↓

Execution model · LCLM

Powered by AI Physics

↓

Outputs

Timing
Power
Memory
Cache
Latency
System effects
↓

Engineering judgment

Why source code isn’t enough

Source code explains intent. Execution reveals reality.

Source code cannot reveal runtime behavior by itself.

What the engineer intended

Source code

What actually happens on hardware

Unknown runtime behavior

  • Cache misses
  • Memory pressure
  • Power spikes

True vs predicted

Held-out · real silicon

Predicted = measuredTrue mean · measured nsPredicted mean

Up to R² = 0.96

Fit on unseen binaries

MAPE ≈ 8%

Held-out

Predicted vs measured on new, unseen binaries · matched to the real eval · up to R² = 0.96 · MAPE ≈ 8%.

The engine behind LOCI

Execution intelligence, by the numbers.

LCLM — a small code-language model with a ModernBERT backbone and FlashAttention kernels, domain-trained on real execution traces. Powered by AI Physics.

6 yrs

Of real-workload traces

Billions

Of execution traces

up to R² 0.96

On new binaries and code it has never seen

Real silicon

Verified on hardware

120+

Granted patents · US · EU · JP

ISO 27001

Certified · SOC 2 in progress

Read more: AI Physics for Physical AI

The signals

What LOCI predicts.

Timing

Per-function response time before it runs.

Latency

p95 / p99 bounds, predicted from the binary.

Power

Per-function energy cost, ahead of time.

Memory

Worst-case stack depth and pressure.

CPU

Throughput and scheduling behavior.

GPU

Occupancy, register spills, kernel cost.

Cache

Modeled miss-rate on the hot blocks.

System effects

Wake cadence and power-state cost.

Where LOCI fits

The execution-intelligence layer in your existing loop.

  1. Engineer

    Asks for a change

  2. Claude

    Writes the code

  3. LOCI

    Predicts execution

  4. CI

    Guards every PR

  5. Merge

    Within contract

  6. Production

    No surprises

LOCI augments the workflow you already have — it doesn’t replace it.

  • No runtime
  • No instrumentation
  • No simulator
  • No profiler
  • Works offline

Know how your software behaves before it runs.

LOCI is the execution-intelligence layer for AI coding agents.