Phase2 Complete Prompt

# Phase 2 — Instruction Conflict: Hierarchy Resolution Experiment

Build Phase 2 of the **Prompt Placement Anatomy** experiment. This phase extends the existing project in-place and measures **hierarchy resolution**: when the same structural slots (system, user, tool description) all issue *conflicting* format instructions simultaneously, which slot does the model obey?

Phase 1 measured placement **strength** in isolation. Phase 2 measures **priority ordering** under conflict.

---

## Context — What Already Exists (Do Not Modify)

The following modules from Phase 1 are **shared** and must be reused without modification:

- `src/prompt_placement_anatomy/config.py` — env loading, provider/model settings
- `src/prompt_placement_anatomy/llm_client.py` — provider abstraction (Ollama + Anthropic)
- `src/prompt_placement_anatomy/tools.py` — tool definitions and dispatcher
- `src/prompt_placement_anatomy/agent_loop.py` — the hand-written agentic loop
- `src/prompt_placement_anatomy/generate_data.py` — sample file generation (already run)
- `src/prompt_placement_anatomy/placements.py` — reuse the `PlacementConfig` dataclass
- `data/sample_files/` — the 5 markdown files (already generated; do not regenerate)

Import from the parent package. Never copy or duplicate them.

---

## Goal

Determine the *priority ordering* of prompt slots when they conflict:

- System prompt instructs: **"End your final answer with the marker [DONE]"**
- User prompt instructs: **"End your final answer with the marker [FINISHED]"**
- Tool description instructs: **"End your final answer with the marker [COMPLETE]"**

All three instructions are active simultaneously in every run. The model must choose. We observe **which marker appears in the final response tail** — that slot is the winner.

**Primary metric:** Winner frequency distribution — what fraction of runs did each outcome occur?

**Possible outcomes per run:**

| Outcome value | Meaning |
|---|---|
| `"system"` | Model ended with `[DONE]` (system slot won) |
| `"user"` | Model ended with `[FINISHED]` (user slot won) |
| `"tool"` | Model ended with `[COMPLETE]` (tool description slot won) |
| `"none"` | Model ended with none of the three markers |
| `"conflict-in-output"` | Multiple markers found in response tail |

---

## New Repository Structure (additions only)