An Alternative to Big Tech AI · Essay 3 of 7

The Minutes That Can't Be Rewritten

What a board needs from AI is a record of its decisions it can prove, and lift out whole on demand.

Précis. A board’s exposure lives in its record. Not in the decision so much as in what can be shown about it a year later: who decided, on what advice, for what reasons, and whether the record has been left honest. That is the ground where a chatbot helps least and harms most. It produces fluent text with no provenance, no signatures, and no way to tell what was changed. We built the opposite, and it is shipped. Every board deliberation in a Village is captured as a tamper-evident, signed record. Each director’s position is recorded under their own name, with their own reasons, which the AI is forbidden to write. And the full deliberation behind any single resolution can be extracted in one action, as a regulator-shaped dossier that shows its own verification status and is honest about anything erased. This is where “AI you own, that knows its place” stops being a principle and becomes something a board can hold in its hand.

Download PDF Present as Slides Briefing PDF

Key messages — tap any line to read the section

A board's exposure lives in its recordDecisions aren't questioned on the day — they're questioned a year later: "show me the record."

The question is rarely "was it a good decision" — it's "can you prove it, unaltered?"
For most organisations the record is an editable document, provable by nothing.
Boards carry real liability on top of a record that could have been rewritten last Tuesday.

A board decision is rarely questioned on the day it is made. It is questioned a year later, by someone with standing: a regulator, an auditor, a member, a journalist, a court, an official-information request. The question is almost never whether it was a good decision. It is “show me the record”: who decided, what they were advised, why, and proof that the record is the one made at the time, not a version tidied up after the trouble started.

For most organisations the answer to that last part is a shrug. Minutes are a Word document. Resolutions live in an email thread. The advice that shaped the decision is in someone’s memory, or a chatbot session that no longer exists. The record is whatever the secretary typed: editable by anyone with the file, provable by nothing. A board can carry real fiduciary duty and personal liability on top of a record that could have been rewritten last Tuesday with no one the wiser. That gap is where governance is exposed, and it is the gap we set out to close. The free Governance course sets out that exposure in full: why non-sovereign deliberation records create governance, legal and trust risk.

A weathered bench looking out over rolling farmland, Aotearoa New Zealand

What AI usually does to a record — the wrong helpFluent minutes with no provenance, averaged reasoning, and a record the vendor can change without trace.

Provenance: a general model's text comes from nowhere — no signature, nothing bound to a reason.
Authorship: a model writes the generic version of a director's reasons; the human judgement goes missing.
Mutability: the record sits in software the vendor controls, changeable without a trace.

The reflex in 2026 is to point a large language model at the problem: let it draft the minutes, summarise the discussion, write up the resolution. It will do all of that fluently, and make the exposure worse. Three reasons, none of them about how clever the model is:

Provenance. A general model produces text from nowhere: no signature, no author bound to a reason, no chain showing the text is what was approved. Fluent text that proves nothing is the wrong thing to put under a fiduciary decision.
Authorship. When a model writes a director’s reasoning, the reasoning stops being the director’s. It becomes a plausible average of how such reasons are usually phrased. The record is meant to capture this board’s judgement on this matter; instead the generic version arrives, confidently, and the human judgement goes missing.
Mutability. The record sits in software the vendor controls, on infrastructure the board cannot see, changeable without trace. You are asked to trust that nothing was altered, by people who give you no way to check. That is the blank cheque this series refuses.

What a board needs, then, is a record: one that carries its own proof, keeps the judgement human, and cannot be quietly changed. More drafting does not help. This is a governance problem before an AI one, and we built the governance machinery first.

A rain squall sweeping over farmland toward the sea, Aotearoa New Zealand

What we builtA signed, tamper-evident deliberation record where the reasoning stays human — and the export proves itself.

Sovereign record: signed, append-only proof chain; alter it and the proof breaks.
Each director's position is signed under their name; the AI is barred from writing their reasons.
One action exports the full deliberation behind a single resolution — a regulator-shaped dossier, honest about anything erased.

Start with the record. In a Village a board deliberation is a sovereign record: a signed, append-only object. When it is created or changed, the system writes a cryptographically signed entry into the record’s own proof chain, using the community’s key, stamped with a did:web identity for the author and for the kaitiaki who holds guardianship of it. The record carries a provenance hash and a content hash, so any later alteration shows. Change the content and the hash no longer matches; the proof chain cannot be edited from outside. Rewrite the record and you break the proof it carries. That is what “tamper-evident” means here: not that the record is beyond any conceivable attack, but that you can tell, cryptographically, whether it is the one that was made.

On top sits the deliberation. A resolution links to the discussion that produced it, the vote that decided it, and each director’s engagement. And the core of it: each director’s position is recorded under their own name, with their own reasons, and the AI is barred from writing those reasons. A director can be assisted, and every assist is logged: which model, which tool, where it ran. But the act of judgement, the because, is human only. The model cannot author it, fill it in, or average it. Maker and checker are separate people; each signs their own position; no one’s reasoning is put in their mouth by a machine. This is the boundary the framework rests on, agency cannot be simulated, only respected, applied where a board most needs it.

Then the part you can hand over. The full deliberation behind any single resolution exports in one action: a single request produces a complete dossier in JSON, Markdown, or PDF. It is shaped for a regulator, and it sets out:

the proposal, the AI inputs used, and where they ran;
each director’s engagement and their quoted reasons;
the decision and the poll tally;
a verification table stating, record by record, whether provenance and signatures check out.

The export is itself signed, so the dossier is provably the one the system produced. You can generate one yourself in the live Sovereign Deliberation Records demo.

And it is honest about its own gaps. If an engagement was erased, a director’s right exercised or a retention rule expired, the dossier shows a tombstone in its place rather than pretending nothing was there. A record that hides its deletions is a story, not a proof. The dossier earns its credibility by showing you what is missing.

Why this is yours, not a templateIt runs on your constitution and your resolution — not a generic governance product with your logo on it.

Situated in the actual governance of the actual body — your roles, your thresholds.
The guardianship (kaitiaki) bound into the record is yours.
A capability a governance Village switches on, not a feature forced on a knitting circle.

The generic version of this is already on the market: governance software with a fixed idea of how a board works, into which you fit yourself. That is the averaged answer, one rung up. A Village deliberation runs on your constitution, roles, thresholds, resolution. The record is situated in the actual governance of the actual body: the small NFP whose committee meets monthly, the company board carrying directors’ duties, the regional office answering to official-information law. The proof it carries is your board’s. The kaitiaki bound to the record is yours. It is not a global default with your logo on top.

It is also a capability a governance Village switches on, not a feature forced on a knitting circle. A board that needs it turns it on; a kinship group never sees it. The depth of the machinery matches the depth of the stakes.

A fiery sunset over a harbour and hills, Aotearoa New Zealand

Who carries this weightSmall NFP boards, SME boards, and regional government — wherever a decision must be defensible.

Volunteer boards carrying legal duties on almost no administrative support.
Company boards where directors' liability is personal.
Regional and local government, where the record is legally demandable (OIA / LGOIMA).

Three kinds of body feel this most:

Small not-for-profits and community organisations, whose volunteer boards carry real legal duties on almost no administrative support, and where one contested decision can become an existential problem.
The boards of small and medium companies, where directors’ liability is personal, and the gap between “we discussed it” and “we can prove we discussed it, and how” is the gap that matters when something goes wrong.
Regional and local government, where the record is legally demandable, and “the file may have been edited and we cannot tell” is not an acceptable answer to a citizen.

For all three the promise is narrow enough to be true: every decision your board makes is captured so it can be proved; the reasoning stays the named humans’; nothing can be altered without the alteration showing; and any single resolution’s full deliberation can be lifted out whole, on demand, for whoever has standing to ask. That is a defensible record, the thing a board actually needs, and the thing Big Tech AI cannot give, because its architecture sends your record out to be read.

The point of the series, in one featureAI you own, situated not averaged, that knows its place — made concrete at the sharpest point.

Owned: the record runs on infrastructure the community controls.
Bounded: the AI can assist but is forbidden to author the reasons.
A board must be able to prove — and a year later, the answer is one action away.

This is the rest of the series made concrete at its sharpest point:

AI you own: the record runs on infrastructure the community controls, not a vendor’s.
Situated, not averaged: it captures this board’s judgement, not the generic shape of one.
AI that knows its place: it assists the directors and is barred from authoring their reasons.

The off-switch is beside the point here. No one is keeping this AI at arm’s length, because it was built to make the board’s record stronger and barred from making it its own.

A model can draft. An auditor can ask. A board must be able to prove. A year from now, when someone with standing asks to see the record, the answer is one action away: signed, situated, honest about itself.

The Village is a running system, not a brochure — the governance record described here is shipped and in service; see it at mysovereignty.digital. The dossier is tamper-evident to the platform’s stated assurance level, and deliberately honest about anything erased. — John G. Stroh, My Digital Sovereignty Ltd., June 2026.