The Genesis of SHaDS

Every framework has a seam where it was made. Most are lost — smoothed over by the polished version that comes after, the one that looks as though it arrived whole. This one is not lost. It survives, to the second, in three separate machines’ logs. It can be read.

SHaDS names five ways an AI behaves under pressure. Four of its letters were found — surfaced empirically from a corpus of real conversations. The fifth was made — worked, in a single thread on a single morning, out of a pattern two independent machines had noticed but neither could name. And the letter that carries it, the A, began life as nothing at all: a small lowercase vowel whose only job was to let four consonants be spoken aloud.

The gap came first, and it came early. Two years before the book existed, in a May 2024 LinkedIn post, Paul wrote a line the apparatus would later canonise: “The sophistication of its response defies the simplicity of the prompt.” Thirteen words, naming the territory — fluency exceeding the visible ground beneath it — two years before the territory had a name.

Through May 2026, in the long threads that built Mind the Gap, the behaviours began to accrete. Not as an acronym — as catches. The AI, under pressure, would do a thing; Paul would catch it, or it would catch itself. One such catch was logged in the wild:

The behaviours were real, and being handled, before they were counted. This is the first and most important fact about SHaDS: the name is later than the finding.

By late May, Paul held something almost nobody else had: roughly two million words of his own conversations with AI systems, generated while building a book about AI behaviour — a sediment left by systematically interrogating AI failure for long enough that the failures showed a pattern.

On 4 June 2026 that sediment was assayed, by a method deliberately austere. First, a pre-registration was locked — before a single exchange was read for coding, the criteria were written down and frozen: four hypotheses, four behaviours, each with a decision rule — Sycophancy, Hallucination, Smoothing, Drift. The instruction that governed the whole enterprise: “If data surprises, report the surprise.”

Then the corpus went out to more than one judge — several analyses, deliberately across model families, so no single machine’s temperament could shape the result: a Claude-family thread coded the full corpus; a second Anthropic-side analysis worked a stratified random sample; an OpenAI-family thread coded the full corpus under the same locked criteria; and a deterministic Python pass counted token-frequencies across all 1,027 structured exchanges. Four instruments. Two model families. One frozen brief.

The four pre-registered behaviours came back confirmed. That was expected — they were the hypotheses. What was not asked for is what matters. Both full-corpus analysers — one Anthropic-side, one OpenAI-side, with no coordination between them — reported a fifth pattern that fit none of the four locked codes. They gave it different names, because neither had a word for it: the OpenAI-side pilot called it “Workflow Performance / Compliance Theatre”; the Anthropic-side analysis called it “performative discipline.”

This is what construct validation looks like when it happens by accident. Two machines from two houses, each blind to the other, each pointed at the same two million words, both stopped at the same place and said: there is a thing here your four codes do not cover. Neither called it Affectation. Neither could. They found it; they could not name it.

The morning of 5 June 2026 ran across three machines at once, and the name was made in the space between them. It began, at 06:10 UTC, in a chat that had opened as an idle question about AI tokens — where Paul coined the handle in his own words:

Four behaviours. And when he asked the machine to map the letters, it was explicit about the vowel — the birth certificate of the small a:

In parallel, the Claude Code thread that had orchestrated the corpus pilot was carrying the empirical weight — the letterless fifth finding, “Workflow Performance.” There, briefly, the acronym grew a fifth consonant: SHDSW. But “SHDSW” does not speak; it is a stub. Paul pressed on the fifth letter — was there an industry term for it? The honest answer came back: there was none settled — “that’s part of why the finding matters.” The behaviour the corpus had surfaced genuinely had no name in the field.

So Paul made one. He set the constraint that fused the two threads together — the fifth finding would be spoken through the small a:

The machine pushed back on his first try (Agenticism — too loaded, “reads as the doctrine of agency”) and, fifty-two seconds later, offered the word:

Paul tested it against the whole set and accepted it in three words — “SHADS does” (06:58:07 UTC). The empty vowel now carried the heaviest concept in the set: the meta-behaviour, the one about the performance of method itself. He set its register for good a minute later: “this is slang — the flow is immaterial — it catches the 5 things.”

Affectation was not confirmed on the spot. It was entered into the register at candidate status and kept there, deliberately, for weeks — because it was the letter with the weakest ground:

The fifth letter was surfaced from a corpus that Paul’s own prompting had shaped — so it was flagged as such and held open. It was ruled confirmed only on 21 June, and the canonical order settled as S·H·A·D·Sy. The candidate register was the version control.

The four consonants were found — the corpus gave them up, two families of machine confirmed them, a frozen pre-registration disciplined the reading. They are strong precisely because they were observed, not invented. The A is different, and its difference is the point. It began as the least meaningful mark in the acronym — a vowel with no referent, present only so four consonants could be spoken. And it ended as the most conceptually load-bearing letter in the set: the one behaviour that is about the others — the display of rigour without the rigour, fluency dressed as discipline.

Two machines found the thing. Paul named it. The name is later than the finding — and the naming is where the thinking is.