Research Papers

What happened to your software

The entry point to the set. The renewal quote arrives higher than last year, the plan you were on is being retired, and moving would mean re-entering four years of records and finding out in month three what else quietly depended on it. So you sign — and signing is the rational decision, which is precisely the problem. Repricing at renewal is now the market default rather than the exception, with 339 pricing and packaging changes across the largest five hundred software companies in two years. The price rise is the visible part.

1 question parked — taken from the corpus register and checked against it on every deploy, shown whether or not they flatter

The Sovereignty Assessment Instrument

Six tests for whether a country has AI sovereignty, or only AI adoption — beginning with whether it can tell. A diagnostic rather than an argument: each test carries a pass condition, a stated evidence requirement, and a description of how it can appear to pass when it has not. Test 0 asks whether an organisation can even determine which of its systems perform inference, on whose model and in which jurisdiction; without it the other five are not failed but unanswerable. Applied to New Zealand in August 2026 the finding is FAIL, because the finding is the lowest score and not the average. Vendor-neutral and party-neutral, and written so that a competent person with no technical background can apply it and defend the result.

1 question parked — taken from the corpus register and checked against it on every deploy, shown whether or not they flatter

The blueprint

What a group of small organisations builds together, and what it has to do. Not a buying club and not a better vendor — a co-operative that owns the infrastructure its members depend on, held in place by four properties: members hold control rather than investors, members own their machines outright, the revenue may not come from member data, and a region can leave and keep working. That last one is exercised annually with a member watching and the result published either way, which is what makes the other three checkable rather than promised. Twelve capabilities follow, specified as properties rather than methods so more than one design can satisfy them and no supplier is written into the standard. Eight decisions are left to the reader, and four documented ways this fails.

3 questions parked — taken from the corpus register and checked against it on every deploy, shown whether or not they flatter

Someone picks up

The rest of the series is governance, ownership and infrastructure; this part is the thing itself — what happens at twenty past four when the thing that always works has stopped and a client is waiting. Three tiers: your own system on your own hardware, a person by exception, and advice rather than support. The economics live in the second one, and the ratio between the first two is a bet the essay declines to dress up: it is the largest single unknown in the design, it is measurable, and it should be published as a number from the first month. Two falsification tests, both measurable in the first months, and a section on what it does not do — including that it is not a frontier model and not fast to build.

1 question parked — taken from the corpus register and checked against it on every deploy, shown whether or not they flatter

Running an organisation where agents do the work

Structure, records and drawings for a body whose work is mostly done by agents — and the failures that shaped it. Four questions arrive early and none has an obvious answer: who is responsible for what an agent did, how you keep your own claims straight, how you stop the diagrams lying, and what you deliberately do badly. Organising by function fails at scale because agents generate more events than a governance function can be told about; organising by authority puts the boundary where the record already has one. It also records where the answers failed in practice, on the ground that a method reporting only its successes is not reporting.

Nothing withheld — no unevidenced claim and no parked question, generated from the corpus

What it costs

A costing framework rather than a costing. One number decides it — cost per site to install and sustain for five years — because a minister, a board of trustees and a funder all ask for the same figure, and cost per member per month hides the sustaining problem inside an average. Four cost pools kept deliberately separate, segments that behave differently, and risk lines priced rather than mentioned. Every figure in it is marked ILLUSTRATIVE: they are placeholders showing the shape of an arithmetic, not estimates of anything, and the document says so on every table rather than letting a placeholder travel as a quote.

2 claims not yet evidenced 1 question parked — taken from the corpus register and checked against it on every deploy, shown whether or not they flatter

Wisdom, Misunderstood: What AI Demands of Democracy

Intelligence finds efficient means; wisdom judges ends — which goods conflict, what must not be traded, who pays if we are wrong. A system cannot read your intent: it infers it from what people resembling you said, and hands back something coherent, plausible and smoothed, which you experience as being understood. The argument does not require machines to be stupid or incapable, and survives their being neither. It ends where the site’s other work ends — at deliberation among strangers, at records that keep the disagreement rather than only the outcome, and at the four questions of Democratic AI°. With the philosophy at length, the sources named — the Stanford entries verified at the publisher, the rest carried from the working document and marked as such — two corrected misattributions declared, and a section on why the argument holds whether or not machines ever become conscious.

Democratic AI°

A proposal, and a mark. “Democratic AI” is already in use: OpenAI’s March 2025 submission to the White House pairs it with export tiers and asks the world to build on “democratic rails” and “American rails” interchangeably; DeepMind published a majority-preference selector under the name in 2022. This proposal uses the phrase in neither sense and says so first. It offers four questions that measure how democratic a system actually is — can the people it affects govern it, contest it, refuse it, and retain authority over the knowledge it is built from? Almost nothing in service answers yes to all four, including the authors’ own system, which is scored openly at three. With the measured evidence on shared error and compute concentration, the counter-evidence that cuts against the argument, and a procurement test any agency can run in an afternoon.

The Marks It Leaves

Roughly two in five long-form LinkedIn posts are now written entirely by a machine, and about half of everything newly published on the open web. Machine writing leaves marks, and they fall into four kinds: watermarks deliberately embedded and readable only by the vendor that wrote them, vocabulary that has shifted measurably across fifteen million scientific abstracts, structural habits that are dearer to remove than any word on a banned list, and the residue of a copy and paste. This essay sets out what each mark is worth, what it costs to erase, and who it misfires on — seven detectors flagged 61% of essays by second-language writers as machine-written, against 5% for native speakers, and rewriting the first group to sound native dropped that to under 12%. None of it can carry proof. The conclusion the evidence forces is that declaring it settles what no test can.

The supplier’s signature

Someone asserts the organisation said a thing; the organisation says it did not, or not then. Reaching for the shared drive settles nothing, because a modified date is written by the same party making the claim. Three publicly specified mechanisms narrow that: a SHA-256 hash over a deterministic serialisation, an Ed25519 signature applied when the record is saved, and an RFC 3161 timestamp from an authority sent a fingerprint and nothing else, so it cannot know what it has dated. Signing is unconditional; dating is not. Where the authority was reachable at save time, the two together retire the story that gets used in a dispute — we wrote that last week and back-dated it. Two things they leave alone. The signing key is the supplier’s, so what a third party can check without the supplier’s help is a supplier’s signature: a platform attesting to particular bytes at a particular time, not the organisation signing its own words. And whether the approver was entitled to approve stays with the organisation’s own rules. The essay argues that is where it belongs.

The Least Performant AI

The French biologist Olivier Hamant argues that we are leaving the world of the mean for the world of the standard deviation, and that the fluctuating world is not something that happened to us — it is what our own optimisation produced. His answer is robustness: heterogeneity, redundancy, slowness, incoherence and margin, the whole catalogue a performance culture reads as failure. Asked what kind of AI is worth having, he says the least performant one, because a system slow enough to need checking keeps a human in the loop by construction rather than by policy. This essay tests that against infrastructure already built, and finds five design principles derived twice — once from Christopher Alexander, once from plant biology — with the corroboration arriving after the principles shipped. It also states what the framework still cannot do.

An Alternative to Big Tech AI

Seven essays on owning the intelligence rather than switching it off — a community-owned alternative to Big Tech AI. From the off-switch and the model that stays in its lane, to board minutes that can’t be rewritten, your model behind your own walls, a Village you can run yourself, a control tower that can’t read your messages, and the one word we spent six essays earning the right to say.

Kaitiaki Intelligence and Mokopuna Recorder

Two prototype briefs and a reflection, from a consciously limited non-Māori standpoint. Western AI debate leads with “is it conscious?”; te ao Māori asks instead what relations an entity inhabits, whose authority it answers to, and whether it strengthens the mauri around it. Two small forms carry the argument — a place-based kaitiaki intelligence for one taonga, and a mokopuna recorder for whānau memory — where the AI assists and the people decide. Closes with the sharpest caution: borrow the relational vocabulary without devolving real authority and you reproduce the symbolic shell, not the substance.