KHALM

Deterministic neural compute

AI that doesn't guess. It learns, reasons, and cites its work.

A neural engine that runs on edge silicon — zero GPU required, byte-identical from one run to the next, and trained on unlabeled data. No hallucination. No drift. No cloud tax.

See KHALM-VMR The problem
Scroll
KHALM-VMR

Verifiable records for AI models, checked offline

“Trust us” is not evidence.
A record you can check is.

KHALM-VMR is KHALM's implementation of the Verifiable Model Record (VMR), an open standard for recording what an AI model is and who stands behind it. A verifier checks a record offline against its own trust store, and reference policy packs grade it rule by rule. No cloud service. No ledger. No custom cryptography.

0network calls
Verification runs offline against a trust store the operator controls. No cloud service to trust.
5reference packs
EU AI Act, NIST AI RMF 1.0, ISO/IEC 42001, C2PA AI Disclosure 2.2 and IETF RATS. Any authority can author its own.
3answers per rule
Compliant, non-compliant or indeterminate, with the clause each rule encodes. Reference results, not legal advice.
See KHALM-VMR The open standard
Next: NullTick
NullTick

A neural server engine for
multiplayer games.

NullTick replaces the fixed 60 Hz tick loop with a neural scheduler that only fires when something changes. The same deterministic engine, applied to multiplayer networking: fewer CPU cycles, lower bandwidth, identical game state on every peer.

Visit NullTick →
2,160×fewer ticks
In an idle Minecraft server scene (60 Hz baseline vs event-driven). The engine sits silent when nothing moves; it spikes only on real state changes.
98% CPU idle
Headroom on a dedicated server during normal play. The neural scheduler consumes compute only when the game state actually demands it — no spin-waiting for a clock tick.
3engines
Unity, Unreal, and Godot — NullTick ships SDKs for all three. Same deterministic brain, same plugin API, one line to enable.

Built to drop into your engine

Unity P/Invoke · Burst ECS Shipping Unreal Engine UE plugin · NullTickRuntime Shipping Godot Godot · GDScript Shipping Custom engine (C ABI) 16-symbol C ABI C ABI ready
Next: Pulsar
Pulsar

The training platform for game AI

Brains, not behavior trees.
Game AI that's trained, not scripted.

Pulsar trains the AI behind your characters — navigation, decisions, reactions — on the same deterministic KHALM substrate that powers NullTick. You supply the game world. Pulsar supplies the brain.

See Pulsar Get notified In development
45KB brains
A trained brain ships in a few kilobytes of frozen state — no model zoo, no runtime framework, no Python anywhere near the player's machine.
99.7% nav
Learned crowd navigation in our demo — measured headless on held-out episodes, no hand-written navigation rules.
0GPU
Inference runs on plain CPU, byte-identical to the trained pass. A shipped game carries the brain, not the datacenter.
Next: Rift
Rift

Passive network anomaly detection

Byte-identical alerts.
No packets touched.

Rift is a fixed-point spiking neural network that watches network telemetry, classifies it against a trained baseline, and emits per-class confidence scores — without ever touching the packet path. Same brain, same input, same answer: on x86, on ARM, in eBPF.

See Rift Get in touch
8classes
SYN flood, UDP amplification, ICMP flood, DNS amplification, FIN scan, HTTP flood, exfiltration — and normal. Each class neuron fires independently, so two attacks can be flagged at once.
0floats
Q16.16 fixed-point everywhere in the tick path. Byte-identical output across x86_64, aarch64, Cortex-A, and eBPF — no drift, no rounding surprises, replayable forever.
0packets touched
Passive observer only. No drop, redirect, or modify anywhere in the engine. The host firewall stays authoritative. Zero blast radius.
Next: The wall
01

The industry's AI wall

Today's dominant AI paradigm — large stochastic models on server GPUs — has three fundamental problems that no amount of scale solves:

Hallucination
Non-deterministic by design
A stochastic model can give three different answers to the same input. For safety-critical decisions — autonomous driving, industrial monitoring, financial trading — maybe is not acceptable.
Drift
No two deployments are identical
Model updates, A/B tests, and cloud-side changes mean the "same" system behaves differently tomorrow. A frozen, deterministic brain eliminates this failure class entirely.
Cloud tax
Inference belongs on the edge
Every inference round-trips through a GPU datacenter, burning watts and adding latency. The future is silicon that runs AI where the data lives — at mW power, with no network dependency.
Next: The substrate
02

The substrate: a neural engine
that keeps its word.

Every KHALM learning product — NullTick, Pulsar, Rift, and what comes next — runs on the same substrate: a learned neural model that is deterministic by construction, not by capping a stochastic system.

Deterministic
Byte-identical, every run, forever
Same input produces the same output — every time, on every machine, cold-loaded from a frozen brain. An alarm from March replays identically in December. No RNG, no version drift, no "it worked on my machine."
No train-time labels
Learns normal; anomaly is emergent
The engine watches unlabeled data and builds a model of what is normal. Novelty is detected as deviation from that learned baseline — no manual labelling of faults or anomalies required, which means zero bottleneck for deployment.
Edge-native
No MAC units, no GPU, no cloud
The engine uses only integer arithmetic — no multiply-accumulate operations. This makes it deployable on microcontroller-class silicon at milliwatt power, with predictable real-time latency and zero cloud dependency.
Get in touch
Contact

Built something that needs a brain that keeps its word?

We're heads-down building. If your application demands deterministic, auditable, edge-deployable neural inference — let's talk.