steeropathy: Skill for Claude Code

Build a new steeropathy experiment

steeropathy is a lab for agents that talk through model internals, never text: one agent's state is read off its activations (or its J-space — the words forming in its layers that never become tokens) and either measured or pushed into another agent's forward pass. Every experiment is a probe, not a product — the goal is a reproducible existence-proof you can play with, run honestly with controls, and write up including the failures.

Read the closest existing bench before writing a new one — they are short and each is a set of answers to the same four questions:

• steeropathy/transmit.py — the thesis in ~15 lines: read a mood off one agent, push it into another. The direction/injection/before-after core.
• steeropathy/ecosystem.py — Eco: a population, per-round step(), the post/get/_state_of plumbing everything else subclasses.
• steeropathy/resonance.py — Reso(Eco): ledgers, transfer/conservation, the J-space channel, an unsteered decision turn.
• steeropathy/unsaid.py — Unsaid: the pure J-space channel (_flicker), no steering at all — only reading.
• steeropathy/warmer.py, steeropathy/zombie.py — a hot-and-cold game and a refusal-outbreak game built on the above.

The four choices that define an experiment

1. Whose internals you read — one agent, a pair, a room of N.
2. What crosses — a mood/concept vector, a scalar on one axis, the J-space word cloud, a similarity, or nothing (read-only).
3. Who decides — nobody (mechanical rules), or agents choosing via an unsteered tool call (steering breaks JSON long before prose).
4. What you measure — and against what control.

The plumbing to reuse (don't reinvent)

• Subclass Eco (or Unsaid) for self.post(path, body) / self.get(path) against brainscope. Write a custom __init__ (like Reso) — don't call super().__init__ if you don't want its mood seed.
• Build a direction from the model's own contrast, not a baked vector: capture_mood(url, texts) returns mean(texts) − neutral, unit-normed, at a good layer. For a custom axis, mean-pool /capture over two prompt sets and subtract. Register it with POST /directions {name, vector} and steer with {"name", "strength", "layer_from", "layer_to"} in the chat body.
• Read J-space from the generation's trace: tag each call with metadata:{demo, case, variant}, then match it in GET /traces, fetch GET /traces/{id}, and walk trace["jlens"] for {t, p} tokens. Filter with a dictionary + stopwords (see unsaid.STOP/WORDS), or a task lexicon (see zombie.REFUSE_WORDS).
• Enable the J-space channel once with POST /jlens {"on": true}; if it fails, the server has no lens — degrade, don't crash.