LATENT SPACE / TOPOLOGY VIEW

DSCR / Neuro-symbolic AI research

Intelligence.
Under control.
By design.

DSCR is the core research technology developed by Adiktiv Technologies Inc. Nexdatas is its current presentation brand. We are seeking industrial partners to experimentally evaluate steering large language models (LLMs) through latent injection.

Read the technical specs Prototype status

Explore an industrial partnership

RESEARCH PROTOTYPE / LOCAL-FIRST INFERENCE

ENGINEERED FOR CONTROL Rust-native core Latent-space steering Portable .dscr archives Local model execution

EXPERIMENTAL PROTOTYPE

Experimental LLM steering.

Native Rust inference engine in active development. We are validating latent-space steering in a research prototype; production readiness and deterministic behavior remain to be established.

Review the technical scope

A different foundation

Neural capability.
Symbolic precision.

Work closer to the model. Make the structures that guide inference explicit, inspectable, and yours to configure.

Steer the latent space

Explore targeted latent injection. Choose a model layer and adjust the influence of structured representations on generation.

Keep execution local

Keep sensitive data on-premise and reduce dependence on metered inference APIs. Local execution gives industrial teams control over where models run and how infrastructure costs scale.

Inspect the foundation

Package topological structures in .dscr archives. Inspect and validate them with a focused, Rust-native command-line toolkit.

From structure to inference

A small surface.
A deeper level of control.

The DSCR toolkit connects portable topological archives to local inference. Start by inspecting a structure, then experiment with how it shapes model output.

RUST CORE GGUF MODELS .DSCR FORMAT

Prototype / technical scope

What we are validating.

A concrete starting point for researchers and industrial partners evaluating the prototype.

Models tested: Experiments include Qwen and SmolLM2. DSCR focuses on LLM steering; applicability to other models remains to be evaluated.
Steering controls: Target layer, latent influence (α), and semantic injection input.
Structure format: .dscr native archives for knowledge, experiments, and steering configuration or data.
Execution: Rust CLI and inference core, with Python tooling for latent preparation.
Validation stage: Research prototype. No published production benchmarks or deterministic-output guarantees.

CLI / NATIVE FILES

dscr                   # Rust CLI
knowledge.dscr         # Native archive
experiment.dscr        # Experimental archive
model-steering.dscr    # Steering configuration or data

Native .dscr naming convention for the research prototype.

DSCR / Experimental validation

Industrial Research Partnerships

We are seeking industrial partners to experimentally evaluate steering LLM behavior through latent injection.

Our objective is to test DSCR against real-world use cases, measure its effectiveness, reproducibility, and limitations, then evaluate its potential for industrial application.

Proposed experimental conditions

A / Baseline

LLM without steering

The task prompt, without additional steering instructions or DSCR latent injection.

B / Textual instructions

LLM with textual instructions

The task prompt with explicit steering instructions supplied as text.

C / DSCR latent injection

LLM with DSCR latent injection

The task prompt with a separate DSCR latent steering signal.

Evaluation criteria Influence on responses, response quality, reproducibility, and execution overhead.

Proposed comparison using the same model and evaluation tasks. Outcomes remain to be measured. Research prototype · Active development.

Propose a pilot use case

Founder & technical lead contact@adiktiv.ca

Bring a real-world use case.

Help evaluate DSCR against an industrial problem and define an experimental protocol together.