# Phase 4 Baseline Model Adapters

This document describes the baseline model-adapter layer in its current final
project form.

## Purpose

Phase 4 provides the model access layer used by the rest of the framework.

It answers the question:

- which model backends can the framework actually run?

## Current Supported Providers

- `mock`
  - pipeline validation and testing
- `groq`
  - hosted real-model experiments through an OpenAI-compatible API
- `ollama`
  - local/open-source model experiments

This gives the project:

- one safe validation backend
- one hosted experiment backend
- one local/open-source backend

## Why This Matters

The original phase plan called for:

1. one primary model path first
2. then a second model family later

The current adapter set now supports that cleanly:

- `groq` covers hosted experiments
- `ollama` supports local Llama-family and Qwen-family runs

## Environment

### Groq

Set either:

- `GROQ_APIKEY`
- `GROQ_API_KEY`

### Ollama

Default base URL:

- `http://127.0.0.1:11434`

Optional override:

- `OLLAMA_BASE_URL`

## Recommended Configs

### Hosted baselines

- `configs/hackaprompt_override_25.yaml`
- `configs/tensortrust_hijacking_25.yaml`

### Local/open-source baselines

- `configs/ollama_llama_hackaprompt_override_25.yaml`
- `configs/ollama_qwen_tensortrust_hijacking_25.yaml`

## Final Phase 4 Position

Phase 4 is complete in the current framework:

- the adapter layer is implemented
- it supports both hosted and local real-model paths
- it is documented and tested
- later phases can use it directly without more adapter work