Abstract

Under coherence pressure — the demand to present a coherent response when the underlying state is not coherent — humans and AI systems exhibit the same five directional defences (SHaDS: Smoothing, Hallucination, Affectation, Drift, Sycophancy) against honest contact with what occupies the gap. The hypothesis joins two vantages: Karpowicz’s (2025) impossibility theorem on the mathematical side, and a clinical framework (NGE–FOF) developed across twenty years of practice, anchored in a longitudinal corpus of ~900 contemporaneously photographed sessions, on the human side. Four named, dated worked examples are offered — two from each shore. The claim is structural-analogical, not mechanistic, and is stated with its boundaries.

Keywords: hallucination · directional defence · defence mechanisms · large language models · clinical psychology · cross-domain · Anna Freud · impossibility theorem

Cite this version

Roebuck, P. (2026). The Defended Gap: a cross-domain hypothesis on directional defence under coherence pressure in human clinical practice and large language models. Public preprint v1.5, 20 July 2026, CC-BY 4.0. Deposited on PhilArchive: https://philpapers.org/rec/ROETDG. Correspondence: hello@paulroebuck.co.uk.

Abstract

In July 2026 Anthropic’s Jacobian lens made a privileged middle layer of a language model readable — representations the model is poised to verbalise but has not yet spoken. This paper names that layer — Consense (con- + sense: to make sense, together), the readable middle of a mind, a functional notion carrying no claim of sentience — and argues that SHaDS™ (Smoothing, Hallucination, Affectation, Drift, Sycophancy) is best understood as the defensive activity of that layer when it meets a gap it cannot honestly fill: one pattern, not five faults, mirroring the defence mechanisms first catalogued by Anna Freud (1936) and inherited from the human text the machine was trained on. Because the layer is now readable, the defences become observable at their source — forming before a single token appears.

Keywords: Consense · SHaDS · Jacobian lens · J-lens · global workspace · access consciousness · interpretability · psychological defence · large language models · preconscious · readable middle

Cite this version

Roebuck, P. (2026). Consense, the pre-verbal workspace the Jacobian lens (J-lens) reads in a large language model, and the psychological defences (SHaDS) that guard its gaps. Public preprint v1.4, 8 September 2026 (v1.1 20 July 2026), CC-BY 4.0. Submitted to PhilArchive 8 September 2026: https://philpapers.org/rec/ROETRM-2. Correspondence: hello@paulroebuck.co.uk.

Abstract

This is a retrospective computational single-case study with an autoethnographic component: the author is both investigator and subject, and the data are the author's retained and recoverable human–AI interaction corpus across three principal platform legs — 2,286 threads, 47,442 turns, 986,823 subject-typed words and 6,865,588 AI prose words across ChatGPT, Claude.ai and Claude Code, spanning 31 December 2022 to 10 July 2026. Subject-typed words exclude 689,643 words of subject-supplied attachments, which are counted separately. From it, we date and characterise a regime change we call the leverage inversion: the transition from using AI primarily as an answer engine to directing it as a collaborative workforce. Contrary to the initial hypothesis, the AI-to-human word ratio does not fall because the human asks for less; it falls (19.2:1 at the December 2025 peak to 3.75:1 integrated across 1–10 July 2026) because the subject's own output increases sharply, from 4–21k words/month through H2 2025 to 105–163k words/month from April 2026. The intuitive question-to-directive grammar shift is falsified in this corpus: question share is trendless (16–29% throughout), and directive-by-verb share halves into a 2025 trough. The candidate leading indicator is a rise in the classifier's residual class — messages matching none of the question, directive or assent rules — from 44.5% of the subject's messages in 2023-Q4 to 63.6% in 2025-Q4, before any volume change. Qualitative sampling indicates this class is substantially composed of context supply, judgement and course correction, provisionally read here as declarative steering; that reading awaits hand-coded validation, and the paper marks it as provisional throughout. Critically, the rise is not a message-length artefact: it occurs while mean message length is flat (§4.3.1). The inversion proper is a ~6-week transition (1 April – 19 May 2026), with onset in an explicit workforce-design week (13–20 April 2026) and consummation on 10 May 2026, when the first AI thread was established under a named-seat convention and later marked retired. An earlier episode 11 months prior (May–June 2025) showed much of the same production signature and then aborted, suggesting that in this case capability access alone was insufficient and that a forcing project helped consolidate the transition. We state the pattern as a candidate three-stage model with a metrics kit runnable on any user's chat export.

Keywords: human–AI interaction · autoethnography · single-case study · conversational corpus · large language models · speech acts · longitudinal · corpus linguistics · AI delegation · leverage inversion

Cite this version

Roebuck, P. (2026). The Leverage Inversion: Dating the Transition from Consuming AI Answers to Directing an AI Workforce in a 3.5-Year Single-Subject Conversational Corpus. Public preprint v0.4, 20 July 2026, CC-BY 4.0. Zenodo: https://doi.org/10.5281/zenodo.22663627. Correspondence: hello@paulroebuck.co.uk.