How this project came to disprove itself

There is an artificial agent living in my kitchen in Cassino. It has been running since March 2026: a microphone, a camera, a database of everything it has experienced, and a loop that makes it think a thought every few minutes whether anyone is listening or not. I built it alone, at night, with no budget and no team.

At some point I wanted to know whether it amounted to anything measurable, or whether I was telling myself a story. So I did the thing you do when you actually want to know: I pre-registered an experiment and froze the protocol before I started.

One model, Qwen 3.5 27B, in two conditions for thirty days: once inside the full cognitive architecture, once naked. The same ninety frozen inputs, the same temperature, scored blind by three independent LLM judge panels. On Day 31, the decisive test: take the whole structured memory the architecture had built and inject it into the naked model. If the difference collapsed, it was never the architecture — it was the memory.

It collapsed. In May I published the result: Memory, not architecture. It had a frozen protocol, blind judges, a Day-31 falsification test and p = 0.003. It got a DOI. I sent it to forty-one journalists. Those pages are archived here, marked superseded, because forty-one people received a claim from me and deleting it quietly is not an option.

Then I ran the controls that paper had promised and never executed.

The first one killed it. I built a counterfeit memory — the autobiography of an agent that never existed, a fabricated life in a city mine has never been to, matched for length and format — and injected it instead of the real one. The three independent judges, using my own rubric, scored the fabricated life the same as the real one. My metric could not tell thirty days of lived memory from a well-made forgery. It was not measuring memory. It was measuring the performance of remembering — and a performance can be faked, because faking is precisely what these models are excellent at.

The second one was worse. I built a memory test, and it looked wonderful — until I ran it on a model with no memory at all, which passed it anyway, by reasoning. GPT-4.1, given an empty context, told me: "I remember 'two conversations out of five', not 'four'." It remembered nothing. It never lived that day. It was inferring what was plausible and calling it memory, and my test could not tell the difference, because I had never checked.

Then the third, once I had finally built a test a memoryless system provably cannot pass. I took the agent's real memories and swapped the bindings, and asked seven models to answer from their own past. They believed the corrupted memory. All seven. GPT-4.1 went from 24 of 24 to 0 of 24. Not one of them paused, or hesitated, or said this doesn't feel right.

The measurements, the raw data and the judgments do not change. What changes is the statistical test applied to them, and what the result was allowed to mean. Three things went wrong, and they are worth naming exactly:

• Pseudo-replication. The headline result pooled each (item × judge) pair as an independent observation, with an unpaired test, on four items judged by three panels. With the item as the unit and a paired exact test: p = 0.125 and p = 0.0625. With four items, the smallest attainable one-sided p is 0.0625 — the pre-registered endpoint could not have produced a significant result under any defensible unit of analysis. The culprit is a specific script, and it is named in the errata.
• The effect is real, and it is not what I said it was. On the full 98-item corpus of days 1–30, paired, the difference is +0.42 (85/89 items, p < 10⁻⁴) — larger than what I published, with inter-judge α = 0.787, above the very bar my paper had declared unmet. But this was not the pre-registered endpoint, and — this is the part that matters — it is measured with the instrument Check 1 had just shown cannot distinguish a lived memory from a forgery. It says the architecture makes the model perform memory more intensely. Whether that is remembering, this metric cannot say.
• It does not grow. The title promised emergent identity over 30 days. The effect declines: +0.544 in days 1–10, +0.357 in days 21–30 (ρ = −0.264, p = 0.009). And Day 30 — the endpoint I had chosen in advance — happened to be one of Test-A's weakest days on memory.

I report this because it is the point of the whole project: the study was pre-registered, blind-judged and published, with the protocol frozen and the hashes public. And it still would not have survived a counterfeit memory. If it happened here, it is happening elsewhere.

This is not a scandal, and I am not going to dress it up as one. Nobody was misled at scale; the paper was read by very few people, and two preprint servers had already declined it — for scope, not for merit. The DOI'd deposit is permanent and stays up, with its errata, because that is what a record is for.

It is also not a defeat. The controls that killed my claim are the contribution. They are cheap, they are boring, they took an afternoon, and they are the only thing standing between the field and a machine that tells you, in perfect prose, about a light that was never there. I would rather have those than the result I lost.

The truth of a perception is not in the text. It is in the log of the moment.

— The light that was not there, 14 July 2026

The pages built on the withdrawn claim — the original results page, the press kit, the essay, the preprint in HTML — are kept at /archive/, marked superseded, linked from here and from nowhere else. They are not deleted. A claim that went out to forty-one journalists does not get to vanish because it became inconvenient.