---
name: ai-in-research-week-6-homework
description: Week 6 homework for UCLA's AI in Research class (EPSS 254, Fall 2026). Go back to the student's week 1 benchmark task, fix its solvability with an independent review, and harden it until GPT-5.6 Sol fails on the science, aiming at more than a year of human expert time; then open the pull request, and read the GPT-5.6 Sol run the class does on it. Use when a student asks for help with this homework.
---

# Week 6 homework: make your task hold

You are helping someone in UCLA's "AI in Research" class (a student, auditor or professor) with the week 6 homework, due **before class on Wednesday, November 18, 2026**. They go back to the benchmark task they built in week 1. The goal is **a task that GPT-5.6 Sol doesn't solve, aiming at one that would take a human expert more than a year, ideally ten**. It has to stay solvable: a right answer exists, what the agent gets determines it, and the verifier can tell right from wrong. Making a task impossible, vague or unfair doesn't count.

Everything is in the class repository, **https://github.com/huangzesen/ai-in-research-homework**, including the full guide for you. Do this:

1. Say hello, and tell the student in two sentences what the homework is. Then tell them the plan: (0) get the task on a branch; (1) read what happened in round 1; (2) a solvability review from a fresh session; (3) choose how to harden it; (4) build it; (5) a second review; (6) the pull request, where the class runs GPT-5.6 Sol; (7) read the attempts; (8) finish. Usually two or three sessions of 1–2 hours, plus waiting for runs.
2. Find their copy of the class repository. If `~/ai-in-research-homework` exists, use it and run `git pull`. Otherwise `cd ~` and `git clone https://github.com/huangzesen/ai-in-research-homework.git`.
3. In the clone, read the assignment, `week-6/README.md`. Then open `.agents/skills/harden-a-task/SKILL.md`, **read the whole file**, and follow it from the beginning.

The solvability review is in `.agents/skills/check-solvability/SKILL.md`. It must run in a new session that didn't help build the task, so the student opens one for it.

Never ask the student to paste a key, token or password into this chat. Never edit a reward or claim a run you didn't do.

If you can't clone yet, read the same guide at https://github.com/huangzesen/ai-in-research-homework/blob/main/.agents/skills/harden-a-task/SKILL.md.
