academic-code-replicatorlisted

## 1. The Core Principle Academic code replication is an **archaeology problem**, not a software engineering problem. The fundamental tension that drives every decision in this skill is: **You must not change what the experiment *computes*, but you often must change how it *runs*.** This skill is grounded in real replication sessions: R code OOM-crashing at 78GB, TensorFlow 2.3 refusing to install on ARM64, nested loops taking 95 million iterations, C compilers silently missing. The protocols here are not theoretical best practices — they are patterns distilled from those failures. ## 2. When To Use / When NOT To Use **Use when**: User is working with existing code from a published paper and wants to run it and verify results. **Do NOT use when**: User wants to RE-IMPLEMENT a paper from scratch (different goal — recommend `ml-engineer` or domain skill). User has their own original codebase they want to debug (use standard debugging, not this skill). User wants to understand a paper theoretically without running code. ## 3. Phase Overview | Phase | Name | Goal | Reference Doc | |-------|------|------|---------------| | 0 | Context Detection | Profile user's environment, route correctly | `references/context-detection.md` | | 1 | Ingestion & Comprehension | Read-only code audit before touching anything | `references/phase-1-ingestion.md` | | 2 | Environment Construction | Create isolated, working dependency environment | `references/phase-2-environment.md` | | 3 | Code Ada