Working Theory · Rachel McBride

← Working frames · Rowan Advisory ↗

The Verifiable Core

A checkable record. AI working over it. A person responsible for what happens next.

Suppose an AI tells you that someone in your network could open a door. Is that a recorded fact, or the model's guess? Who checks it? Who decides whether to act?

The Verifiable Core keeps those answers visible.

The record sits at the centre. AI can read it, connect it, draft from it and suggest changes. A person decides what enters the record and what happens because of it.

THE RECORD CheckableChosenOwned a person decideswhat belongs AI LAYER finds patternsprepares options draftsacts, in bounds provisional A PERSON owns both ends what enters the record, and what happens next accountable reads proposes only what the person confirms enters the record
The record is the only thing that persists. AI reads it and proposes; nothing it proposes becomes a fact until a person confirms it. That same person decides what happens next, which is what owning both ends means.

The record

It has three properties:

The AI layer

AI works over the record. It can find patterns, prepare options, draft and carry out bounded tasks. Its outputs remain separate from the facts underneath them.

You can change the model or rebuild an answer without losing the record. A machine's suggestion only becomes part of the record after a person checks it.

The person

A person owns both ends. They decide what the system may use, and they remain responsible for the decision or action that follows.

This matters more as agents begin to act. An agent may propose a new fact or a next step. It cannot quietly turn its own guess into the source everyone else relies on.

Keep the record checkable. Let AI work over it. Keep a person responsible for both the input and the action.

What still needs work

The open problem is keeping a record verifiable over time: where each item came from, whether it changed, whether it is still current, and what should never enter it.

Mappa is where I am testing those questions in an organisation. My own AI setup tests the same design at the scale of one person.

Last true: 18 August 2026

v2.0 · Rewritten around a concrete example and three plain responsibilities: a checkable record, an AI layer and a person accountable for the result. v1.0–1.3 · First public statement and refinements.