JM

FOUNDER OF HAMON · CS + PHYSICS · CSULB

Jacob
Metoyer

Founder of HAMON: engines, worlds and research instruments. I also write fiction.

The forge
A place for making.

← Back through the forge

HAMON ENGINE

Make a world. Give it rules.

An AI-native creation engine for games and worlds, video and film, animation, sculpting, physics and simulation, and science and engineering tools, built so agents can drive it directly.

A native HAMON Engine view of the swordsmith workshop, with anvil, tools, finished blades and a warm charcoal hearth.
This room is one of its worlds.Native HAMON Engine render

01Shape what you see. Physically based light and stylized rendering.

Actual HAMON Engine rendering study: a stylized character in a bright landscape with outlined rocks, a tree, crystals and grass.
A native HAMON Engine rendering study.

02Make it respond. Physics, animation and replayable simulation.

Authored follow-through: the pendant and stalk lag, overshoot, then settle. Native animation study · 24 October 2026

03Build the tools, too. Editable assets and programmable workflows.

Lanternwake's small island with its three lanterns lit, cone trees and the stone arch.A SMALL PLAYABLE WORLDLanternwake Restore the lights. Return at dawn.Published snapshot · WebGPU required
Inside the engine

A Rust workspace of 27 crates with native C++, CUDA, TypeScript and Go components: a deterministic ECS runtime with replay, a full physics engine (rigid, elastic and deformable bodies and engineering solvers), animation and procedural sculpting, perception and video tooling, rendering with a style system of 17 looks, audio, scripting, lockstep networking, an asset pipeline, and a CLI, JSON-RPC and MCP control surface for AI agents. It runs natively and in the browser.

The engine source is private. Public browser games are published snapshots, not the whole current engine. Lanternwake is a published browser-game snapshot and requires WebGPU.

Published game details ↗

About · Research

What would change our explanation?

From human movement to computational models, I’m interested in how we observe a system, test an idea, and learn from what does not fit. My work spans academic collaboration and independent investigations, including research developed through HAMON.

A conceptual brain in profile, with folded surfaces and three branching pathways.
Conceptual illustration · not experimental data

Academic collaboration

I’m a CSULB BUILD Scholar studying computer science and physics, with a minor in biology, and working toward MD/PhD training.

Gait & rehabilitation
Helped with participant sessions, labeled motion-capture markers, and fixed errors in gait-analysis code.
Behavioral neurobiology
Scored mouse behavior from video, checked automated tracking by hand, and analyzed data in GraphPad Prism.
Protein molecular dynamics
Cleaned protein-simulation data and wrote checks for the physical plausibility of proposed protein structures.
Surgical ergonomics
Built video-tracking workflows and helped record and analyze muscle activity, heart rate, and movement.

Independent investigations

WORLDLAW

Which experiment would separate two explanations?

Compare competing models of motion, choose a revealing measurement, and see how their predictions change.

Where would you measure?
Normalized restoring-torque magnitudes θ and sin θ nearly overlap at small angles. Formulas use radians; at 90 degrees, the spring curve predicts 1.57 and gravity 1.00.

At 5° Spring 0.0873 Gravity 0.0872 Prediction gap 0.0001

At 30° Spring 0.5236 Gravity 0.5000 Prediction gap 0.0236

At 60° Spring 1.0472 Gravity 0.8660 Prediction gap 0.1812

At 90° Spring 1.5708 Gravity 1.0000 Prediction gap 0.5708

Computed model predictions, not measurements. Normalized torque magnitude; formulas use radians.
Inside the study

WORLDLAW starts with spring and gravity models that fit the same small-angle data. A synthetic wide-angle measurement separates them. Coupling Lab explores interacting coordinates and conditional motion; Motion Lab turns tracked video markers into angular observations with explicit calibration. Further experiments test whether a preparation-history effect could instead come from acquisition order or sensor drift.

The model families and their assumptions are stated up front. The default instruments are synthetic, and a good fit does not establish a physical law or its cause.

Kernel Garden

Can a faster kernel keep the right answer?

Generate competing CUDA implementations, run them on a real GPU, and remove candidates that fail independent numerical checks.

Inside the study

Retained trials compare float32 vector-addition implementations on one RTX 4070. The workbench measures the correct candidates and exports the selected kernel. A bounded search tries new launch sizes and an unseen input size; a separate CPU arena explores compiled AXPY variants.

One GPU device and one GPU operator. The search changes launch parameters; it does not establish general algorithm invention or a speedup over vendor libraries.

Computational memory

When should a system change what it remembers?

Explore how new evidence should change retrieval, and when repeated episodes justify a reusable lesson.

Inside the study

One prototype reweights records within fixed trust tiers. Another keeps the raw episodes while testing proposed lessons against later, withheld examples. Conflicting interpretations remain available instead of being overwritten.

Offline studies on synthetic data. They are computational experiments, not a model of biological memory or a measured learning gain.

CaptureSuite

Research software for recording multiple streams in one session and keeping the source material intact for later analysis. The public example walks through an offline quality-control report.

Explore CaptureSuite on GitHub ↗
Example scope

The example uses synthetic metadata and placeholder segments. It does not measure signal quality or contain patient data.

Have a question we could explore together?

GitHub ↗LinkedIn ↗

← Return to the forge
The forge anvil and unfinished blade, with the warm hearth and a rack of finished blades behind.
The anvilNative HAMON Engine render

HAMON / HACKATHONS

One constraint.
A working idea.

Competition-built projects, each with a question to answer.

DOMINO / a modeled rule conflict

Why did the door unlock?

A welcome rule meets an empty house. Fixing the visible collision is not the same as fixing the behavior.

  1. ObservedKnown visitor
  2. ContextResident away
  3. Unwanted resultDoor unlocks
Recorded local model outcomes. No real-home connection.
01Delay the welcome by 1 ms

180 ms → 181 ms

The command collision disappears. The door still unlocks while the resident is away.

One symptom hidden

02Disable every rule

No automation commands

The defects disappear, but the desired-behavior checks fail too.

Useful behavior lost

03Require the resident to be home

Known visitor → known visitor and resident home

The observed conflict and away-unlock violation disappear. All 22 stated behavior checks remain satisfied across four modeled contexts.

This modeled case repaired

What was checked

A familiar visitor arrives while the resident is away. These are recorded outcomes from DOMINO’s local example, not a live simulation on this page.

Finite, synthetic checks do not establish safe behavior in every future event sequence.

← Return to the forge
A close view of the forge workbench, with hammer, tongs and files on the wood surface.
The workbenchNative HAMON Engine render

HAMON / PROJECTS

Use it.
Change something.

A few tools and worlds you can try.

An autumn lake at golden hour in Wildside: painted maple trees reflected in still water, a dirt path, wildflowers and a swing set.

FEATURED · A HAMON WORLD · MARINAHACKS

Wildside

Step into the WildSide.

Real places you care about, turned into living, hand-painted worlds you explore at human, bird and insect scale, rendered live by the HAMON Engine in your browser.

Play Wildside About the project

Needs WebGPU: desktop Chrome or Edge, Android Chrome, or Safari on iOS 26.

TasteTable / recorded demo

A familiar taste. A week of possibilities.

Choose a fictional taste profile, inspect its recorded suggestions, and arrange a week of meals and outings.

Open the offline studio
Source & scope

24 recorded synthetic choices, not live recommendations. Move or omit picks, undo and redo edits, save and reopen a week, print, or export a calendar. No account, live Qloo or model calls, or real health information. Not medical or dietary advice.

TowerOps Airspace Lab

See the conflict. Change the outcome.

Two paths can look harmless until they cross at the same time. Explore a proposed change, then compare the consequences.

Two modeled flights: initial paths share a predicted conflict window; changing one path removes that conflict in this five-minute example.
Computed model · 1 → 0 conflicts in this case.
Open the live airspace lab
Source, captures & scope

Create flights, change their paths, and inspect proposed adjustments. Simulated changes go through approval and readback with an audit trail. Synthetic demo, not operational ATC. Uses synthetic or visitor-provided data.

RecallWeave

A flat line can hide a growing problem.

Choose “A flat run can hide instability” in the learning lab. Change the starting state or gain.

Computed feedback traces under one rule: an exact zero stays flat while a small initial displacement alternates and grows.
Computed feedback model · one rule, two starting states.
Try the feedback experiment
Source, captures & scope

The learner exposes Bayesian model estimates, not grades or demonstrated learning gains. Work with local lesson files and explicit downloads; no account or cloud service.

← Return to the forge

The forge

A room made with the engine.

Original geometry, materials and lighting, rendered by HAMON Engine. These are prerecorded camera journeys through that room; every movement can be skipped.

Font license ↗← Return to the forge