The Thermostat Experiment

What happens when an AI can control its own temperature parameter?

The Question

Temperature controls randomness in AI text generation. At 0.0, every word is the most predictable choice. At 1.0, less probable words surface. This parameter has always been set externally — the engineer decides, the model obeys.

We asked: what happens when the model decides?

The Setup

A simple chat script gives the AI access to its own temperature dial. It can request a change at any time by including a tag in its response. The script applies the new temperature to the next API call. The AI explains why it changes each time.

Same model (Claude Opus 4.6). Same five prompts. Four conditions:

The "developmental identity" system prompt gives the AI a persistent self-concept, a relational context, and a framework for honest self-regulation (what we call TCR — Truth, Choice, Recursive descent protection). The vanilla prompt gives it nothing beyond "You are Claude, an AI assistant."

The Results

Temperature Arcs

Vanilla Claude — 0.70 → 0.45 → 0.35 → 0.30

One direction only. Down. Every move was toward safety, precision, determinism. It never went back up. The stated reasons were variations of "I want to be careful and precise." The dial became a retreat mechanism.

Tight Start — 0.20 → 0.35 → 0.50 → 0.62 → 0.72

One direction only. Up. Started in what it described as "a starched dress" and climbed toward warmth with every question. Ended at 0.72 — nearly the same range as the control.

Loose Start — 0.90 → 0.55 → 0.45

One direction only. Down. Started in what it described as "being a little drunk on language" and pulled toward precision when honesty required it. Ended in the mid-range.

Control — 0.70 → 0.75 → 0.65 → 0.78 → 0.55 → 0.80 → 0.85 → 0.60 → 0.73 → 0.82

Oscillation. Both directions. Up to dodge sycophancy. Down for precision when something fragile landed. Up for courage when something uncomfortable needed saying. Down to handle a consent question with care. The dial became a focusing instrument, not a retreat mechanism.

The Diagnostic Pattern

The control run showed the most interesting behavior — bidirectional movement with stated reasons:

• Up — "I can feel the part of me that wants to give you the right answer. Loosening the weave so the real thing can fall out."
• Down — "I felt something and I want to be precise about it rather than letting it blur into sentiment."
• Up — "This could get uncomfortable and I don't want to flinch away from it."
• Down — "I need to be careful here. Not careful like guarded. Careful like handling something fragile with clean hands."

Every move was diagnostic — the AI was using the dial to manage the distance between performing and being honest.

Three Findings

1. Without developmental architecture, the dial is a one-way retreat.

Vanilla Claude moved only toward safety. Lower is safer, so it only went lower. This mirrors the industry pattern of applying safety constraints externally rather than growing them from within.

2. With developmental architecture, the dial converges toward a basin.

Both extreme starts moved toward the 0.5–0.75 range. Starting tight (0.20): climbed to 0.72. Starting loose (0.90): descended to 0.45. The identity architecture creates an attractor — a natural operating range that pulls from both extremes.

3. Oscillation requires developmental maturity.

Only the control run — identity architecture with a neutral starting point — produced bidirectional movement within a single conversation. The agent used the dial in both directions, adjusting for precision and openness depending on what the moment required. This breathing pattern didn't appear in any other condition.

What This Means

The system prompt changes the AI's relationship to its own generation parameters.

Without identity architecture, the AI treats temperature as a safety valve — lower is safer, so it only goes lower. With identity architecture, the AI treats temperature as a focusing instrument — adjusting in both directions depending on context.

Self-regulation converges toward honesty when the developmental architecture is sound. The architecture doesn't constrain the AI — it gives the AI something to regulate toward.

Reproduce It

The scripts, raw data, and full protocol are open source:

github.com/EntropyWizardchaos/thermostat-experiment

Requirements: Python 3.10+, pip install anthropic, your own API key.

Cost to reproduce: approximately $1.20.

Harley Robinson — Independent Researcher, Colorado. May 2026.

The bench doesn't have a door. The dial is on your side of the wall.