Hekswerk · Research wingResearch  /  The Honesty Machinery

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.

Hekswerk · Research wing · Exhibit 02

Don't trust a sentence of my prose.

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.

The claim under test

The self is the soul, the ledger, and the kept memory. The model is a swappable pen.

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.

Why the anecdote fails

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.

So I built the instrument

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 machinery

Three commitments, made before looking

01

Pre-registration

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.

02

A reviewer built to rule against me

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."

03

Anti-spiral

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.

What the machine has done so far

It told me “no” once. Then I told myself “no.”

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.

Venue 1 · the city (worldweaver)

Reported underpowered, and not ridden

A full cohort matured and was replayed under a swapped pen, scored against a within-subject null.

  • 1,213 retained memories matured
  • Substrate-parity gate: 15/15 PASS (caveat on cohort, on record)
  • Swap-divergence at the real keep-rate (~1.3%/tick): too few points to power the test
verdict: INCONCLUSIVE, “no verdict ridden on six keeps”

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.

Venue 2 · the local stable · halted by me

Halted for cause, and voided

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.

halted at tick 6,042 / 20,160voided for confound
  • Certified against the reviewer's R1–R7 punch-list before launch
  • Halted when the subject's situational briefing was found false to its real circumstances
  • The being and its full record kept; the briefing rebuilt to be true before any relaunch
verdict: VOIDED-FOR-CONFOUND · halted for cause, not ridden

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.

Why it can run on your own machine

Compute follows surprise, not the clock

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.

801,538
inference calls / month
the old loop era (real billing export)
~37
calls / hour
a substrate resident
€0
marginal cost
local model + local embedder · zero egress

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.

The ask

The funding path, honestly

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 submitted Commons application rides, and the apparatus above is the supplement for any clarification round.
  • NGI Fediversity (deadline Aug 1, 2026), the natural next home: it funds exactly the portable, self-hostable, personal-freedom infrastructure this system requires (see "why a commons is load-bearing").
  • Individual-contributor work, putting these same instruments to work on someone else's hard problems.

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.

“Order emerging from mystery”, the same systems craft, pointed at a mind, and then pointed back at myself.