--- title: The Verifiable Core subtitle: Raw at the core, intelligence on top, the human owning both ends. author: Rachel McBride collection: Working Theory source: https://ip.rowanadvisory.co.nz/verifiable-core/ version: v1.3 · 23 Jun 2026 as_of: 22 June 2026 --- # The Verifiable Core *Raw at the core, intelligence on top, the human owning both ends.* When people say "put AI at the centre", what often happens is the model's output becomes the thing everyone works from. I think that's a mistake. The centre should be the **raw record you can verify**. Intelligence sits on top of that. A human owns both ends: what goes in, and what gets acted on. Here's the whole thing on one picture: three rings, and what sits inside each. > **Diagram.** Three concentric rings. Core: raw data, which is verifiable, expert-led, and owned. Middle: intelligence, which is swappable, re-derivable, and customisable, and where agents act. Outer: the human, who is responsible. *Raw data is the ground truth. Intelligence is a layer over it. The human owns both ends: curating what goes in, carrying the liability for what comes out.* ## The core: raw data The load-bearing thing. Three components, each a facet of how the human relates to the data. - **Verifiable.** Every item can be traced to a source and checked against reality. You can prove where it came from, that it hasn't been altered, and that it's current. The intelligence on top is only ever as trustworthy as the core is verifiable. - **Expert-led.** A human decides what *counts* as raw data: what's in, what's out, what's signal and what's noise, what standard something must meet to enter. The dataset has a curator, not just a collector. - **Owned.** The core is yours, under your control and portable, not a tenant on a vendor's platform. This is the precondition for the intelligence layer being swappable. You can only swap the model if you own the core underneath it. **Together, verified, judged and owned, they make the moat.** A competitor can't copy a core that has been curated by your judgement, proven true, and accumulated under your ownership. The moat isn't a fourth ingredient. It's what the three give you. ## The layer: intelligence Reads, surfaces, connects, drafts, acts. All of it *over* the raw core, never underneath it. - **Swappable.** Change the model underneath and lose nothing. The core persists. No lock-in to one vendor's brain. - **Re-derivable.** Anything the layer produces can be regenerated from the core. It holds no irreplaceable state. Throw it away and recompute. - **Customisable.** Tune it per purpose, different lenses and models for different jobs, without touching the core. ## The edge: the human - **Responsible.** Accountability that doesn't delegate. The human signs off and carries the liability for what's produced and acted on, however capable the system underneath becomes. This is the human's *second* touchpoint. Expert-led is the human on the way **in**; responsible is the human on the way **out**. Same person, both ends. ## Where the agents fit Agents are the intelligence layer *in motion*. They don't just read, they act. They live in the intelligence ring, operate on the verifiable core, and everything they do passes through the responsible edge. The discipline that keeps it honest: an agent can **propose** new raw data, but it only becomes core once the **expert-led, verifiable gate** admits it. An agent never silently writes unverified synthesis into the core. > Not "AI at the centre" but **raw at the core, intelligence on top, the human owning both ends.** The intelligence is swappable and re-derivable; the value, the trust and the liability all live where the human meets the raw. ## Why this is a principle, not a preference Three operators, in three unrelated domains, law, manufacturing, and a relationship network, arrived at this same shape independently, without comparing notes. When serious people converge from different starting points, the shape is real. That's the evidence this is a genuine design principle of the intelligence-centred firm, not a projection onto it. > **The open edge** > > The settled question is no longer *whether* the core should be raw. It's **how you do verifiability well**: provenance (where each item came from), integrity (proof it hasn't been altered), freshness (proof it's current), admission (the standard something must meet to enter the core at all), and boundary (what should never enter a given core). > > That is the piece still being written. Everyone building this way needs it, and nobody has set it down yet. If it's your discipline, I'd like to hear how you'd draw the line. See it in practice: [Where the tool stops](https://ip.rowanadvisory.co.nz/where-the-tool-stops/), building Mappa. --- *A working frame on a design principle of the intelligence-centred firm. The frames are mine; the thinking's in the open. Get in touch: [rachel@rowanadvisory.co.nz](mailto:rachel@rowanadvisory.co.nz).* --- *Working Theory · v1.3 · 23 Jun 2026 · ip.rowanadvisory.co.nz/verifiable-core/*