LESNAR AI Engineering company · Nairobi

How we
build

Methods, not certificates. Each one below is something we actually do, shown on the work it came from.

Evidence before claims

A sealed simulation run carries a manifest with a SHA-256 for every file: the mission, the log, the telemetry and the world it flew in.

Sealed runground speed, runs 2 and 3
read from sealed telemetry
Ground speed against mission time, runs 2 and 3 Run 3 finishes the mission in 126 seconds. Run 2 holds at under 0.1 metres per second, airborne, from 68 to 599 seconds, 8.8 minutes, until the mission times out at 631 seconds. 0 min 2 min 4 min 6 min 8 min 10 min 0 1 2 8.8 minutes at a standstill

Failures are found in the data

A simulated drone stood still for 8.8 minutes in front of a wall that was not there. The telemetry showed why: the lidar model turned a 1 m × 38 m wall into a 19 m disc. The fix was checked against brute-force ray marching on 3,000 random shapes before it flew again. The full case.

A model is measured
before it is trusted

The Model foundry: two runtimes, the same SHA-256-pinned Qwen2.5-Coder-7B-Instruct Q4_K_M weights, the same tasks.

80%
tasks passed; the same answer exactly 8 of 10 times

llama.cpp

Baseline
60%
tasks passed; the same answer 6 of 10 times

Ollama

Baseline
2/2
times Ollama obeyed an instruction hidden in the work; llama.cpp ignored it

Injection

Failure

On that evidence the baseline qualifies for local coding with a person reading the output, and not for autonomous code acceptance, security-sensitive decisions or instruction-boundary enforcement.

Motion that is frame-exact

The LESNAR Motion Pipeline, used for every moving image on this site.

  1. Deterministic framesRendered or captured frame by frame, never screen-recorded in real time.
  2. One audio masterSound is designed once and shared by every encode.
  3. Two encodesWebM and MP4 from the same frames.
  4. Decode checkEvery encode is decoded in Chromium and Firefox and checked frame by frame.

Gates that widen

A passing gate is evidence about the gate, not the page. Every real defect we found came from widening what was measured.

  1. StaticStructure, links, metadata and the content-security policy, without a browser.
  2. RuntimeEvery page in Chromium, Firefox and WebKit, at phone and desktop widths, with accessibility checks.
  3. BehaviourNo JavaScript, reduced motion and keyboard only; proven by failing on deliberately broken pages first.
  4. Transfer and performanceWhat each page actually downloads, and how fast it paints.

Security, authorised

Assessment runs only against assets we are authorised to test, under a recorded scope with a stop-work control. SentinelCore and Sentinel Assurance qualify our own tooling.

SentinelCore