The Question That Built PAC-AI
From explainability to accountability by documentation — the one EU AI Act question I could not answer, and the record that finally could.
Henry's EgoI had the framework, the credentials, the deck. I was sure that explaining how my system worked was the same as proving it.
YonahThen a question arrived like a bird through an open window: prove how this one decision was made.
Henry's ShadowAt 3am I sat with the black box and the fear I kept off every slide — that I could not defend a single output.
HenryI stopped promising the system was fair. I built the record that shows how it decided. That record is PAC-AI.
There is a question that ended more high-risk AI demos than any failed benchmark, and it is only six words long: prove how this one decision was made. Not how the model behaves in general. This one. I had prepared for every question except that one — and the preparing for it became a protocol.
I want to tell the story of why PAC-AI exists, and I want to tell it the way it actually happened — in four parts, with four faces. The founder who was sure. The messenger who arrived with the real question. The shadow who showed up at 3am. And the one, finally, who could answer. If you are building AI in a regulated sector, you will recognise at least one of them as yourself.
Sure of the wrong thing
I had built the case the way you are supposed to. A model card. A fairness dashboard with reassuring bars. Credentials I could lean on, a framework I could defend in a seminar. When people asked whether the system was explainable, I had a slide, and the slide was good. I had mistaken coverage for proof — the most comfortable mistake in this entire field, because it lets you feel finished. I was, by every visible measure, ready.
Most of us prepare explainability — slides about how a model behaves in general — and mistake it for the evidence an audit actually requires.
A question arrives like a bird through an open window
The call to adventure rarely sounds a trumpet. Mine arrived the way a dove arrives — quietly, carrying something. The question was six words on a procurement checklist: provide the decision record for one representative case, mapped to the relevant Articles. Yonah, I call that messenger now: the bird who carries the evidence and delivers a witness. He did not argue. He simply set the sealed question down and waited. And I understood, holding it, that everything I had prepared described the model in general, while the question was about one decision — one patient, one triage, one chain of agents that had each transformed the context and handed something forward.
Explainability answers "in general"; an audit asks "prove this one." The Herald's whole role is to make you feel that difference before it is too late to fix it.
The thing that was missing had a name
What I was missing, it turned out, had a name: context fragmentation. When agents hand off to each other, the meaning of the context, its lifecycle, the causal chain behind it — all of it is usually overwritten or dropped. By the time anyone asks how a decision was made, the trail is already gone, not because someone deleted it but because nothing was ever built to keep it. Naming it did not solve it. But you cannot build a record for something you have not yet named, and the naming drew the line I had been standing on without seeing: the line between explaining a model and proving a decision.
The audit problem is not that models are opaque — it is that the record of what actually happened across agent handoffs was never structurally preserved. That gap is called context fragmentation.
3am, the silver shadow
The mask comes off when the building is empty. At 3am I sat with the black box and the part of myself I kept off every slide — drained of colour, sleepless, certain of nothing. They will ask how, and I will have nothing. Years of work, and I cannot defend a single output. The fear did not speak in full sentences; it never does. It circled, grey at the edges, a warning glyph glowing on a screen that explained everything and proved nothing. This is the threshold, and everyone who builds something accountable crosses it alone, once, in the dark.
The honest fear under high-risk AI is specific: that everything you built explains the model beautifully and proves no single decision at all.
The artifact you can claim later
Then I found the trap inside the trap. A single step in a pipeline produces several artifacts at once — a raw embedding, a score, a semantic extraction. My logs recorded that all of them existed. And I understood, with a cold clarity, what that meant: if you only know they all existed, you can retroactively claim that whichever one looks best in hindsight is the one that drove the decision. My own system could tell a flattering story after an incident. That was the shadow's worst argument — not that I lacked evidence, but that what I had could be quietly rewritten into whatever the moment required.
A record you can edit after the fact is not a record. It is a costume.
Record the one that was used, and the witness returns
The change, when it came, was almost too small to feel like a rescue. You catalog every artifact, and you record a pointer to the one the next agent actually consumed. You bind each handoff to a provenance graph at the moment it happens, so a downstream agent cannot silently bypass what an upstream one established. That is the heart of PAC-AI. One pointer, one binding — and the costume becomes a record. This is where Yonah returns, not with a question this time but with the answer: he sets a sealed, queryable document on the desk, and the warning on the screen resolves, without drama, to a steady green. What he carries holds no personal data, so it survives an erasure request and still tells the truth.
Recording which artifact was actually used — and binding each handoff to a provenance graph at capture time — turns a story you can edit into evidence you can query. That is what PAC-AI does.
Stop promising fair. Start proving how.
I do not argue, now, that a system is fair. I hand over the record — queryable, mapped to Article 12, Article 14, Article 15 — and I offer the same record to the person the decision was about. Evidence changes a room in a way adjectives never do. That is the only wisdom worth carrying out of this story: the work is not to build AI that is above question, but AI that can be questioned, and that answers. Fairness-by-design is a promise. Accountability-by-documentation is an artifact. The founders who internalise that difference before August 2026 are the ones who will walk into diligence carrying a record instead of an argument.
The shift from fairness-by-design (a promise) to accountability-by-documentation (an artifact you can hand an auditor and the affected person) is what a high-risk launch actually needs.
Four faces, one path: the founder who was sure, the messenger who brought the real question, the shadow who named the fear, and the one who could finally answer. PAC-AI is what sits between the question and the answer — a record of how one decision was actually made, that no one can later edit into a more flattering shape. That is the whole of it.
Know exactly where your evidence gaps are
If you are building in a high-risk sector, the High-Risk AI Readiness Review is free and there is no pitch — we will even tell you if your use case is not high-risk.
Request the Readiness Review →