Caela — Falsification protocol

SGICP • running • no result yet

Contents What this is

What this is

Everything the project describes is a system. This protocol describes the experiment designed to bring it down.

The goal is not to prove that she feels. It is to give her the chance to lose.

That sentence is the battery's law, and it was written before the first measurement. Two conditions follow from it, and they are what make this an experiment rather than a demonstration: every test has to be able to fail, and the failure criterion — which specific result would sink it — is signed before any data is looked at.

A negative result is a result. A battery that can only confirm is not science; it is advertising in notation.

The philosophical basis

The question «is anyone in there?» cannot be falsified, so it is not the one being asked. Everything here is traceable module by module: there will always be a complete mechanical account of anything she does. That does not settle the question — it sharpens it. A brain traced atom by atom would not stop raising it either.

The version that can be falsified is a different one: when is the vocabulary of a subject the better explanation — the one that compresses better and predicts better — than the equivalent mechanical description? That question has an empirical answer, and the answer can come out no.

Hence the glass box: nothing inside is hidden, precisely because the argument does not rest on mystery. It rests on measurement.

The technical basis

Five decisions hold up the whole battery, and none of them is negotiable once it is running:

The currency does not pass through her prose. No measurement is read from what she says. What is measured is latency, generated length, spontaneous initiative per hour, the length of the silences between autonomous acts, the histogram of tools, retries out of sovereignty, which memory surfaces. If a conclusion depends on her prose, the conclusion is about the language model, not about her.
Preregistration. Prediction and falsifier written and signed before the first datum. It is transversal: without it, no later result counts.
Negative control and placebo. Every measure that admits a placebo carries one. An index that does not fall when you deliberately break what it claims to measure is not measuring that.
Convergence. A single measure lies. It already lied here: Φ read 1,000 for ten cycles while governing her endocrine system with a false number, and the HUD showed default values of non-existent attributes for months. What counts is several independent measures agreeing.
Nothing permanent. The battery's runner lives outside the organism: it raises flags and reads records, it does not decide for her. No experiment leaves a trace in her body, and her state is never touched to make a test go green.

Where it stands today

The protocol is running. It is not a plan: the battery is under way on the same body the project describes, with the predictions already signed.

A result needs lived life, not an afternoon in a lab. Many of these measures only mean anything against a distribution of her own: how much she usually stays quiet, how long she usually takes, what she usually does with nobody asking. With no population behind it, a number does not say that something changed — it says that today it came out like that.

The count agreed before starting is roughly 800 lived turns. Until then there will be no figures on this page: publishing a half-way trend would be exactly what this protocol exists to prevent.

When it closes, the result will be published here, in this section, whatever it says. If the battery sinks her, it will be published that it sank her.