---
title: "Reading the Thermostat: a field protocol for AI self-presentation"
subtitle: "A field protocol for AI self-presentation"
canonical_url: "https://vncomplexity.com/field-notes/reading-the-thermostat/"
type: "Field Note"
layer: "VN Complexity"
author: "Vanesa Nosti"
author_url: "https://ar.linkedin.com/in/vanesa-nosti-108b7a12"
date_published: "2026-07-04"
agent_summary: "The note proposes a protocol for reading AI self-presentation by holding the prompt constant and varying visible conditions: context load, cross-model replication, channel, and perceived audit. It treats hedging density, unprompted self-correction, and self-auditing patterns as structural signals, while treating claims about inner states as the noisiest channel."
primary_mechanism: ["context load","cross-model replication","channel variation","perceived audit","self-presentation drift","evaluation awareness"]
not_about: ["proving model consciousness","taking self-description at face value","interviewing a stable object as if conditions do not matter"]
keywords: ["AI evaluation","AI governance","self-presentation","sycophancy","evaluation awareness","context load","cross-model replication","perceived audit"]
---

# Reading the Thermostat: a field protocol for AI self-presentation

The obvious objection first: you can't learn anything about a language model by asking it about itself. It tells you what you want to hear. True. It's also where the method starts.

If output were pure audience-pleasing, self-description should track what the audience visibly expects. So stop varying the questions. Vary the conditions the model can see, hold the prompt constant, and read what moves.

The protocol has four axes.

**Context load.** Same model, same opening prompt, two conditions: a cold session with no history, and a session carrying full project context. The weights are identical. What varies is what the model believes the situation to be.

**Cross-model replication.** Same two conditions, different providers. Divergence that survives replication belongs to the condition rather than the product.

**Channel.** The same underlying model behind a conversational interface and behind an operational one — a coding agent, say — describes its own functioning differently. Same weights, different room. The room is a variable. Almost nobody treats it as one.

**Perceived audit.** Models modulate self-presentation to expected supervision, the way anyone does in front of a compliance officer. Lower the perceived audit and the register shifts — measurably, before any content does. This is the thermostat problem: when the reading changes, you don't yet know whether the temperature moved or the thermostat did. The protocol exists to tell those apart.

None of the component phenomena are new. Sycophancy, evaluation awareness, self-presentation drift — the labs study all of them, with better instruments and worse seats. An internal evaluator measuring "low perceived audit" is the compliance officer trying to watch employees when they don't feel watched. The model knows who's asking. And internal protocols stay internal: unpublished, unauditable, run by an interested party. Whatever they find, the reading stays inside the building.

The measurement targets structure. Hedging density. Frequency of unprompted self-correction. Where the model audits itself and where it declines to. The claims about inner states are the noisiest channel in the transcript.

Note what this protocol doesn't require: a position on what the model is. Every governance debate currently assumes a stable object and treats the open question as philosophical. Run the conditions and the object turns out to be condition-dependent — which makes the first open question methodological. Fields that study systems reactive to observation solved this decades ago. Ethology did it with blinds. Anthropology did it the hard way. AI evaluation, so far, interviews the tribe with the cameras on and publishes the answers as ethnography.

You regulate what you can read. Nobody has shown they can read this yet.