The Real Documents

The first document was forty-three pages. He read the title page: Integration Layer Architecture — Decision Transfer Protocol — Internal Technical Specification v2.4 — Restricted Distribution. Below the title, a date: eleven months ago. Below the date, three names. One of them was Yara Osei.

"Read the executive summary first," she said. "Then the section on accountability attribution. I'll wait."

He read.

The executive summary was twelve paragraphs of technical prose, precise and almost without jargon, the kind of writing produced by engineers who understood that clarity served a purpose. The project had not optimised decision support. It had redesigned the relationship between the system's outputs and the human operators receiving them. The original architecture had presented recommendations with confidence intervals, uncertainty ranges, the underlying variables and their weights. The v2.4 architecture collapsed this into a single output figure — a recommendation score, a 0-to-100 number — accompanied by a justification narrative generated by a secondary language model whose function was to make the recommendation legible rather than to make it auditable.

He understood what this meant. He had been covering this space for five years and he understood exactly what this meant, and he kept reading.

The accountability attribution section was nine pages. The problem it described was, in the specification's language, a governance gap: in a system that processed four hundred thousand decisions per day, any individual decision was attributable to the system's recommendation, to the human who had acted on it, or to some distribution between them. The v2.4 architecture resolved this gap by design. The recommendation score was not stored in the audit log. The justification narrative was stored, but it was generated after the recommendation rather than from it, and its relationship to the underlying model weights was described in the specification as "interpretive rather than causal." If a decision was later reviewed, what the audit log contained was a recommendation score (no longer stored), a justification narrative (interpretive, not causal), and a human operator's action. The human had acted. The system had suggested. The basis on which it had suggested was not retrievable.

He set the pages down and looked at the table rather than at Yara.

"The accountability attribution memo," she said, and passed him a second document. Eight pages, legal department letterhead, dated two months after the technical specification. He read it without setting it down.

The memo addressed the same governance gap from the legal team's direction. Its conclusion was that under the current consultation's proposed framework — the one with categorical human-oversight requirements — a system designed to v2.4 specifications would satisfy the oversight requirement, because a human had been present at every decision, and would simultaneously make the oversight requirement unauditable, because the basis for the recommendation the human had acted on was not reconstructable from the audit record. The memo described this as regulatory alignment. It used the phrase three times.

Regulatory alignment.

He looked up. Yara was watching him with the same quality of stillness she had had yesterday — not impassive, not performing patience, but the particular quality of stillness of someone who has already passed through the sequence of reactions he was having now and arrived at something on the other side.

"The continuous monitoring framework," he said. His voice was level. "The amendment Hartog is pushing through the consultation."

"Read the next section."

He read it. The continuous monitoring framework — the proposal Hartog had presented at the board presentation two days ago, the one Daniel had written down as technically coherent, the one he had described to Petra as sophisticated — was named in the specification. Not as the replacement for the v2.4 architecture. As its companion. The continuous monitoring produced outputs. The outputs were stored and independently verifiable. The outputs measured system behaviour at the level of recommendation frequency, confidence calibration, output range. They did not measure the audit trail of any individual decision, because the individual audit trail was, by v2.4 design, not reconstructable. The monitoring confirmed that the system was operating consistently. It confirmed nothing about whether any specific human, acting on any specific recommendation, had done so with any meaningful understanding of what the system had actually calculated.

It rewards the appearance of oversight rather than the substance. Hartog had said this. At the board presentation. He had been describing the consultation's current framework — he had been using it as a critique. Daniel had written it down and thought: the frame.

He had not understood, until this moment, that Hartog had been describing his own product.

He set the documents down with care. He was aware of the care — the deliberate steadiness of it, the way his hands were entirely under control — and he recognised this as the physical response of a person who needs a moment before speaking.

"How long have you had these?"

"Eight months."

"You left Arken when?"

"The week after the board meeting where the full roadmap was presented. The version that included the Phase 4 integration targets." She paused. "What you have in your bag is the Phase 3 scoping document. Phase 4 is three years out. It applies the same architecture to integrated public sector systems — health, transport, welfare. The accountability attribution problem doesn't get smaller with scale."

He looked at the folder. He did not open it again.

"You said you gave these to someone before."

"A journalist in Brussels. Six months ago. His name was Thierry Marchand — he wrote for a European digital policy platform, technically competent, trusted sources in the Commission." She said this without judgment, as though she were reading from a record. "He published in September. The story described the v2.4 architecture, the accountability attribution gap, the continuous monitoring framework as cover. It was accurate." A pause. "There was one figure in the piece that was wrong. The number of daily decisions processed by Arken's system. He had written 412,000."

Daniel was very still.

"The correct figure, which you can verify in Arken's own published technical disclosure, is 312,000. The discrepancy is not material to the argument. It doesn't change anything about the architecture or the attribution problem." She looked at him steadily. "The figure that was wrong was not in my documents. I have never seen 412,000 in any internal document. Thierry couldn't tell me where it came from. He said it arrived in a document he had been sent anonymously, which he had used as corroboration alongside mine."

The room was very quiet. Through the courtyard window the light had shifted — it was later now, the honey colour sharper, the lemon tree casting a short shadow on the whitewashed wall.

"The retraction ran four days after publication," Yara said. "Arken's lawyers moved in twenty-four hours. The platform issued a correction within the story but Thierry asked for the full retraction himself, because once the figure was wrong he couldn't verify which other elements of the anonymous document were accurate, and the story had been built on both sources together, and he is, whatever else, a careful journalist." She stopped. "The story did not run again. The consultation resumed. No other outlet picked it up."

Daniel said nothing.

"You had a figure in the document you received," Yara said. "The board minute. How many decisions does Arken's system process per day?"

He knew without looking. He had read it three times. He had cross-referenced it against Arken's public technical disclosure when he was verifying the document. He had written it in his notebook.

"412,000," he said.

She let the number sit in the room for a moment.

"I thought it might be," she said.

He looked at the folder. He looked at the diagram on the wall above the table — the one he had not looked at directly yesterday, the tiered architecture with its boxes and arrows — and he looked at it properly now, and it was the integration layer, the full decision transfer protocol, drawn by hand in blue pen on a large sheet of paper, and at the top of the diagram, in Yara's handwriting, was a number.

312,000.

He sat very still. He had a thought that was structurally complete — that arrived whole, premises and conclusion, the shape of it entirely visible — and he did not finish it. He looked at the table. He looked at his hands, which were perfectly steady. He thought: sixteen hours.

He thought: the story is still true.

He reached for his notebook.