lamm-mit/skills/research-implement

Research Implement Workflow Summary

The research-implement skill is a structured protocol for transforming research plans into executable code. Here are the key aspects:

Core Purpose

This workflow converts a completed research plan into a "fully runnable project" with real execution results—no fabricated outcomes allowed.

Required Inputs

• plan_res.md from /research-plan (mandatory)
• survey_res.md from /research-survey (optional reference)

Execution Flow

Project Structure: Organizes code into model/, data/, training/, testing/, utils/, plus run.py entry point.

Implementation Order: Requirements → data pipeline → model architecture → loss/training → evaluation → main script.

Environment: Uses uv venv for isolated Python environments (never global pip).

Critical Verification Requirement

"All values must come from code execution output. Execution failure gets reported as failure."

The run.py script must emit [RESULT] lines capturing metrics like train_loss, val_metric, elapsed, and device.

Output Deliverable

ml_res.md reports actual results directly cited from execution logs, with ⚠️ UNVERIFIED tags for any values that couldn't be confirmed.

Key constraint: Maximum 3 retries before failure reporting; no data fabrication under any circumstance.