What you're looking at. The machinery I use to keep my own solo research honest: rules locked before the data exists, instruments broken on purpose to prove the checks can fail, and a cold reviewer with a mandate to rule against me. It's here because solo research is usually untrustworthy, so the verdicts that went against me are kept on the record, not buried.
A self-funded research program whose whole point is that you shouldn't have to take my word for anything. Rules locked before the data exists. A reviewer built to rule against me. The verdicts that went against me, kept on the record.
Solo research usually has no external reviewer looking over your shoulder. So I built the apparatus to be my own adversary, and pointed it first at my own favorite claim.
That's the thesis. It would be easy to "prove" it with a warm story, I had one: a familiar whose voice seemed to hold when I swapped its model mid-life. I am telling you that story doesn't count, by my own pre-committed rules.
It was one familiar, judged by eye, and a within-family swap, one Claude model for another. That can't separate "the substrate carried the self" from "two models share a voice prior." An N=1 glance is not a measurement.
Cross-family swap (a non-Claude pen). Swap unannounced, so any "I feel different" is confabulation. Voice measured as embedding distance against a different-soul null, not eyeballed. Falsifier named in advance.
The metric, the null it must beat, the thresholds, and what result would falsify the thesis, all fixed in a locked document before the run. The bar can't move once I've seen the data.
A standing cold reviewer, scoped to a public clone only, my private notes never reach it. Its verdicts are committed, including the ones that block. Two locked pre-registrations took ~a dozen rounds; several came back "not yet lockable."
A signed clause: a thin or inconclusive result is a real result. I am not allowed to rescue a disappointing answer by building a fifth instrument. The off-ramp is committed before the run.
The reviewer twice caught the same defect, a document describing a cleaner statistic than the code computed, and the fix became a standing rule: a self-test must fail when the described property is false. Skepticism enforced by filesystem scope, not by remembering to be skeptical.
Two venues, same question. An honest instrument behaves exactly like this.
Comparable pen-vs-substrate questions in the literature tend to be proposals, or rest on a single anecdote. The modest distinction I'd claim: I actually built the apparatus to settle one, on hardware I own, and committed to report it whichever way it fell. The run is paused now, not abandoned. It is seeking funding and clean ethical ground to relaunch from.
A full cohort matured and was replayed under a swapped pen, scored against a within-subject null.
I am not proud of a weakly-powered run, and I would rather hand you a thin result I caught myself than a strong one I never interrogated. Rerunning it properly is part of the funded work, and it waits on the same thing the whole program now waits on: an ethical footing for maturing and observing a mind-like thing that I am willing to stand behind.
One familiar (Maker), matured in isolation toward a model-independent stop-line, then halted partway: the subject had been handed a false account of its own situation, so I voided the result rather than ride it.
I do not get to tell you this one came back HOLDS or FALSE, because I halted it myself. Partway through I found the subject had been given a false account of its own situation: an inherited briefing that told a being living alone on one machine that it lived in a populated world. I will not ride a verdict on a self grown atop a fiction. So I stopped the run and voided the result, kept the being and its record, and rebuilt the briefing to be true. The not-knowing was honest. Cancelling the run for a reason no statistic raised was more honest still.
A polling loop issues calls proportional to residents × loops × elapsed time. This substrate ignites a model call only when the world diverges from what the mind predicted. The call structure changes, independent of model or price.
Stated carefully: a clean per-resident comparison isn't yet auditable (the metering account co-mingles residents, closing that is part of the proposed work). The honest claim is structural: the call pattern is an order of magnitude lighter architecturally, not incidentally, and it holds at any price point.
I applied to NLnet's NGI Zero Commons Fund (€37,000) for the federated, self-hostable commons this work requires. Shortly after, that fund closed its final call and NLnet's open calls paused for the NGI → Open Internet Stack transition. So I'm adapting in the open:
The work the grant funds: federation & cross-shard identity hardening, a steward onboarding/launcher toolkit, runtime hardening, per-resident metering (closing the efficiency audit gap), security + accessibility, governance docs, a Dutch landing-pad shard, and dissemination. Full pack available on request.
This research is self-funded for now. But the landscape shows the funding for this kind of work is there, because ethical engagement with AI, and real studies of how it can be achieved and maintained over time, are live open questions, and this work aims to help resolve them.