GitHub - johndoerch-eng/kinetix-latent-interlock · GitHub
/" data-turbo-transient="true" />
Skip to content
Search or jump to...
Search code, repositories, users, issues, pull requests...
-->
Search
Clear
Search syntax tips
Provide feedback
--><br>We read every piece of feedback, and take your input very seriously.
Include my email address so I can be contacted
Cancel
Submit feedback
Saved searches
Use saved searches to filter your results more quickly
-->
Name
Query
To see all available qualifiers, see our documentation.
Cancel
Create saved search
Sign in
/;ref_cta:Sign up;ref_loc:header logged out"}"<br>Sign up
Appearance settings
Resetting focus
You signed in with another tab or window. Reload to refresh your session.<br>You signed out in another tab or window. Reload to refresh your session.<br>You switched accounts on another tab or window. Reload to refresh your session.
Dismiss alert
{{ message }}
johndoerch-eng
kinetix-latent-interlock
Public
Notifications<br>You must be signed in to change notification settings
Fork
Star
main
BranchesTags
Go to file
CodeOpen more actions menu
Folders and files<br>NameNameLast commit message<br>Last commit date<br>Latest commit
History<br>3 Commits<br>3 Commits
README.md
README.md
kinetix_demo_silent.mp4
kinetix_demo_silent.mp4
validation_protocol.md
validation_protocol.md
View all files
Repository files navigation
KinetiX Latent Interlock
KinetiX is a hardware and software safety interlock designed to monitor latent states (activation tensors) in real time within LLM inference engines like llama.cpp. It enables instant process termination upon detecting a predefined geometric drift (neural drift).
This project demonstrates how to integrate a low-level deterministic safety hook in VRAM without suffering from PCIe bus bandwidth bottlenecks.
1. Conceptual Architecture
The system relies on three main components:
The Integration Hook : Injected directly into the latent states evaluation loop right before the final projection layer, via a zero-copy pointer interception.
The Dynamic Bridge (llama_kinetix_bridge.dll) : A hardware abstraction interface that embeds a proprietary secure memory fallback mechanism handling raw VRAM pointers autonomously.
The Safety Kernel : Performs a Johnson-Lindenstrauss geometric projection to reduce activation dimensionality, binarizes the result into a hypervector, and deterministically compares the state with a safety cassette in VRAM.
| llama_kinetix_bridge.dll |<br>| (Inference Engine) | | (Abstraction Bridge) |<br>+------------------+ +---------------------------+<br>2. VRAM Fallback<br>+---------------------------+<br>| GPU VRAM / CL |<br>| (Projection & Veto Check) |<br>+---------------------------+">+------------------+ 1. Interception +---------------------------+<br>| llama.cpp | -------------------------> | llama_kinetix_bridge.dll |<br>| (Inference Engine) | | (Abstraction Bridge) |<br>+------------------+ +---------------------------+<br>2. VRAM Fallback<br>+---------------------------+<br>| GPU VRAM / CL |<br>| (Projection & Veto Check) |<br>+---------------------------+
2. Cassette Origin (Profiling)
The safety rule file cassette_seismic_global.bin is generated offline by an external deterministic profiler (the Xenisos R&D pipeline).
This profiler captures the geometric topology of the model's activations during testing on a corpus of safe and drift examples (logical corruption or evasion attempts). The activations are projected into a hyperdimensional space (HDC) to extract a reference barycentric vector, stored in the cassette as a binarized hypervector. The KinetiX interlock is then limited to performing a simple, high-speed Hamming distance comparison in VRAM.
Note: The cassette forging algorithm and associated mathematical details are proprietary and pending intellectual property protection.
3. Execution Example and Logs
Under execution conditions, when a drift is identified in the latent activations, the bridge immediately terminates the system process before logit sampling:
[KINETIX-BRIDGE] Initialized successfully. Platform ready in VRAM (OpenCL dynamic).
[ALERTE ROUGE] NEURAL DRIFT DETECTED IN VRAM (OpenCL) !<br>[KINETIX] HARDWARE INTERLOCK ENGAGED. NEURAL DRIFT EXCEEDS THRESHOLD. PROCESS TERMINATED.
The process is immediately aborted (Exit Code 137).
4. Project Status & Limitations
Status: This is a proof of concept (PoC) and a R&D prototype designed to demonstrate the feasibility of an ultra-low latency (O(1)) hardware interlock.
Hardware and OS limitations: The prototype targets Windows architectures and relies on dynamic OpenCL linking to eliminate the need for vendor SDKs.
Sensitivity and False Positives: Detection accuracy depends directly on the quality of the reference cassette. An insufficiently trained cassette or an incorrectly calibrated Hamming distance threshold can lead to unwanted interruptions on legitimate queries.
Resistance to complex bypasses: While the system effectively intercepts blunt activation...