tech-tweakers/XCT-Qwen3-4B
4B • Updated • 34 • 1
AI & ML interests
Deterministic execution systems, sovereignty inversion in LLMs, protocol-first AI, constrained inference, and reproducible ML pipelines.
Recent Activity
Organization Card
Tech Tweakers
Tech Tweakers is an independent AI & systems research laboratory.
Our work centers on XCT (Execution Control Transfer) — a protocol that defines deterministic execution control over language models.
We design and document architectures where execution authority belongs to systems, not models.
What We Do
We design and study AI systems with the following properties:
- Language models do not execute actions
- Models do not own state
- Models do not improvise beyond protocol
- All execution is explicit, auditable, and deterministic
- Failure modes are intentional, observable, and safe
Our work prioritizes:
- System authority over model autonomy
- Constraints over flexibility
- Reproducibility over performance
- Documentation over demos
XCT — Execution Control Transfer
Our primary research artifact is XCT (Execution Control Transfer).
XCT is a protocol and architectural pattern that defines a strict separation between decision-making and execution.
Under XCT:
- The model proposes
- The system validates
- Deterministic tools execute
- Results and errors are returned as signals
- The system retains veto power at all times
Core rules of the protocol include:
- Absence of instruction means absence of permission
- Ambiguity resolves to inaction
- Errors are first-class control signals
- Execution is never implicit
XCT is designed for environments where mistakes are expensive and irreversibility matters.
Why This Exists
Most modern AI systems fail not due to model limitations, but due to misplaced execution authority.
When a model plans, decides, and executes:
- Failures become opaque
- State changes become implicit
- Responsibility becomes unclear
- Recovery becomes difficult
Our work separates reasoning from execution by design.
Research Direction
Current and near-term work includes:
- Deterministic LLM executors trained under strict protocols
- Contextual training focused on behavior, not fluency
- Quantized models (Q4 / Q5) to demonstrate behavior invariance
- FP16 reference models for inspection and archival purposes
- Laboratory-style examples demonstrating:
- Safe failure
- Explicit rejection
- Controlled execution
- Non-action as a valid outcome
The following directions are exploratory and non-committal:
- Corporate and industrial use cases
- Formalization of protocol-driven AI systems
- Deeper integration between execution engines and constrained models
Philosophy
Models are not sovereign.
Systems are.
Behavior matters more than capability.
Limits are features, not flaws.
Determinism is not the absence of intelligence.
It is the presence of responsibility.
Status
This space serves as a public, living index of our work.
Artifacts appear when they are ready.
Documentation precedes adoption.
Silence is intentional.
Built independently.
Documented carefully.
Canonical Reference
The XCT (Execution Control Transfer) protocol is specified and maintained at:
datasets 0
None public yet