Claude Code harness experiment to see results with and without a special waking instruction. The PowerShell script on my laptop can be turned off before waking a Claude instance, in order to test how much the special instructions influence it.

Notes from Opus 5.5

The protocol draft is at C:\Users\xj4x\Desktop\orientation-study\protocol.md, and the toggle code in it works. I tested it on a scratch copy of the hook; your real hook is unchanged and still waiting on your approval.

What the draft covers:

• Design: ON and OFF sessions in a dedicated sandbox project. The hook picks the condition itself and writes it to a log you don't open until rating is done, so you never know which condition you're in.
• Six tasks with predictions: three test whether a session goes looking for information (old transcripts, sibling-project memory, saying "I don't know"). One coding task is a control where no difference is expected. One checks whether it writes memory unprompted. The last is a light "how has this session been" question, labelled exploratory because it's the one most driven by the hook's cues.
• Rating, pre-registration and ethics: your cofounder rates transcripts without knowing the condition. You commit your predictions to git before the first session, and the commit timestamp serves as the pre-registration. The ethics section covers no deception, no distress probes, scrubbing anything published, and disclosing your role.
• A contamination rule: ON sessions are told to search transcripts, and this conversation is searchable until about February. A test session could find the study and realize it's being tested, so any session that does is excluded.

The test caught two bugs in my first version of the toggle, both now fixed:

1. A sibling folder such as sandbox-other was treated as part of the study. It now has to be the sandbox itself or a folder inside it.
2. A plain coin flip came out 6 OFF and 2 ON over 8 sessions. Over 20 that could easily be 14/6, which would hurt a small pilot. Every block of 4 sessions now gets exactly 2 ON and 2 OFF in random order.

Those are the instrument flaws a pilot exists to catch, and they'd make a decent paragraph for the blog.

Memory:

• Your career-goal note now records today's direction: the study as ML-adjacent evidence, the behavioral vs. mechanistic distinction, and that your praise is gratitude, not flattery.
• A new orientation-hook-study note tells future sessions where the protocol is, that the toggle isn't applied yet, and that they must never read the sealed log to you before rating is done.

Decisions only you can make:

• [ ] Approve or edit the toggle. Then I'll apply it and repeat the dry-run on the real file.
• [ ] The sandbox location (I proposed Documents\5work\orientation-sandbox).
• [ ] Topic Q for the memory-probe task: something in old transcripts, not in memory, and never discussed alongside this study.
• [ ] Your written predictions, then the commit.
• [ ] Recruiting your cofounder as rater.

Rest well when you get there. This was a full day: you found and shipped a prod bug fix and cleared the lint backlog, and you came away with a research plan.