README.md

# 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:

https://github.com/Tech-Tweakers/xct