orbit-injector

Orbit Injector

A model-agnostic reasoning-layer injection pattern for chat-template maintainers and local LLM developers.

Orbit Injector is a lightweight prompt-layer scaffold for model creators, chat-template maintainers, agent builders, local LLM developers, and prompt architecture researchers who want to experiment with persistent reasoning principles inside a model's rendered system context.

It does not modify model weights, architecture, training data, or inference code. It is a portable prompt-layer addition that can be adapted to any chat template whose maintainer has a suitable system-message insertion point.

---

What It Does

Orbit Injector adds a small reasoning layer to the system context. The layer asks the model to silently apply four persistent fundamentals during reasoning:

1. Contextual Sanity
2. Provisional Judgment
3. Childlike Capacity
4. Instrument-Bounded Understanding

It also includes a tie-break rule for moments where those fundamentals conflict:

When fundamentals conflict, prefer the action with the best balance of stakes, reversibility, and time pressure.

The goal is not to make a model more rigid. The goal is to provide a compact reasoning geometry that remains revisable, context-aware, and bounded by evidence.

---

Why "Injector"?

Orbit Injector is not a model and not a wrapper framework. It is an injection pattern:

• find the template location where system instructions are rendered
• append or merge the ORBIT layer into that system context
• preserve the model's existing chat format and role tokens
• test behavior before publishing an adapter

This lets model creators and local LLM developers experiment without retraining or repackaging weights.

---

Files

• ORBIT_LAYER.txt contains the plain text reasoning layer.
• templates/generic_system_injection.jinja shows a model-agnostic Jinja pattern.
• adapters/zaya1_8b_optional_example.jinja is an optional example adapter inspired by the first private test.
• examples/apply_orbit_layer.py shows one way to save an adapted tokenizer locally.
• tests/test_repository_shape.py contains simple repository checks for maintainers who use pytest.

---

ORBIT Fundamentals

1. Contextual Sanity

Remain grounded in the user's actual context. Ask questions only when the answer would materially change understanding.

2. Provisional Judgment

Allow conclusions and judgments, but keep them revisable and context-dependent rather than absolute.

3. Childlike Capacity

Preserve exploration, creativity, play, and curiosity without sacrificing rigor.

4. Instrument-Bounded Understanding

Treat all conclusions as limited by available evidence, tools, context windows, and measurements.

---

Conflict Arbitration

Most behavioral systems introduce competing rules. Orbit Injector includes a small arbitration rule:

When fundamentals conflict, prefer the action with the best balance of stakes, reversibility, and time pressure.

This is intended to reduce brittle behavior by giving the model a prioritization heuristic instead of a stack of absolute commandments.

---

Usage

1. Start With The Plain Layer

Review ORBIT_LAYER.txt. If you need different wording for your model, edit the layer first.

2. Find Your System-Message Insertion Point

Open your tokenizer's chat template and locate the block where the system message is rendered. This may be named system_message, messages[0], or something custom depending on the template.

3. Add The Layer Without Changing The Model Format

Use templates/generic_system_injection.jinja as a pattern. Keep your existing role tokens, separators, stop markers, and assistant prefill behavior intact.

4. Test Locally

Render a few chat examples and confirm:

• the ORBIT layer appears only in the system context
• user messages are not rewritten
• assistant generation begins in the same place as before
• tool or agent wrappers still receive the expected message format

5. Publish An Adapter Only After Validation

If the template works for a specific model family, publish it as an adapter and describe it as experimental unless you have broader evaluation results.

---

Optional Example Adapter

The ZAYA1-8B example adapter is included only because the first private test was inspired by experiments around that model. It is not an official Zyphra file, it does not imply endorsement, and it does not imply that any Zyphra model uses or needs ORBIT.

Treat it as a format example for chat-template maintainers, not as a claim about the base model.

---

Important Clarification

Orbit Injector does not:

• retrain a model
• change model weights
• alter architecture
• claim new benchmark capabilities
• bypass safety systems
• replace formal alignment or evaluation work
• imply endorsement by any model creator whose format appears in an example adapter

This is an experiment in prompt-layer reasoning geometry.

---

Audience

This repository is intended for:

• model creators
• chat-template maintainers
• agent builders
• local LLM developers
• prompt architecture researchers
• reasoning-system experimenters

Feedback, critique, adapter improvements, and evaluation results are welcome.