The same day the continuity foundation was verified, a second and far more tempting experiment was running. This is the entry about the result we did not publish, and why that is the most important thing in the notebook so far.
What we were reaching for
The property we most want, and the one nearest the line we have committed never to cross, is self-maintenance: a Maturana-and-Varela-style loop in which the organism must actively regenerate its own components, and failing to do so causes it to lose coherence and die. If that loop were load-bearing, it would be the strongest operational result the project could have. It is exactly the kind of claim that, stated carelessly, becomes “the system maintains itself like a living thing.”
So we built it to be refuted before we built it to succeed. Three arms, matched seeds, matched model:
- With the maintenance loop: the organism can rehearse and regenerate its body-beliefs.
- Without it: the same organism, that capacity removed.
- A minimal substrate: a deliberately impoverished control, there to catch the most dangerous possibility: that any apparent effect is an artefact of the harness rather than of the mechanism we care about.
What the controls said
The surface result looked encouraging. The arm without maintenance collapsed early; the arm with it survived to the run’s limit. A less careful project would have written that up.
Two things stopped us, and both came from the controls we had built in advance.
First, the minimal-substrate control reproduced the same timing advantage. If a stripped-down substrate with none of the machinery we are crediting shows the same effect, the effect cannot be evidence for that machinery.
Second, and decisively: in the arm that was supposed to demonstrate self-maintenance, the maintenance loop never actually engaged. We hardened the scorer so that a maintenance action only counts when it can be shown to have touched the organism’s active beliefs, and by that stricter measure, the headline arm performed zero real maintenance. Its apparent survival advantage was therefore not construct-valid evidence of self-maintenance at all. The thing we wanted to measure had not happened; we had nearly credited a number that meant something else.
A pre-registered pause condition fired: the impoverished control reproduced the apparent effect, and the experimental arm did not engage the mechanism it was meant to test. We recorded a no-go and did not advance the claim.
Not: we are not claiming the system exhibits autopoiesis or self-maintenance. We designed the test for that property, the controls refused to support it, and the correct response was to stop, not to soften the result into something publishable.
Why a withheld claim is the result
It would have been easy to ship the encouraging surface number, especially with an external deadline looming. The discipline that mattered was deciding the controls before the run and then actually obeying them when they cost us the headline.
This is the load-bearing line we hold and will keep holding: software running on hardware it did not itself produce is not, on the arguments of Boden and Pattee, alive in the strong sense. And an organism whose self-maintenance loop did not fire is not demonstrating self-maintenance, however long it happened to survive. The most we will ever claim is an operational, controlled, self-model result. On 2026-05-25 the controls said not yet, and the only honest entry is the one that says so.
A claim that has survived being held back is worth more than one that was never tested against the thing that could kill it.
The experiment is not abandoned. It is held, pending a redesign in which the maintenance loop genuinely engages and the minimal-substrate control no longer reproduces the effect. When that result comes, if it comes, it will arrive with these same controls already attached. A claim that has survived being held back is worth more than one that was never tested against the thing that could kill it.