A Multi-Turn Audit of Amplification: Design Note (version 2, draft for discussion)
Year: 2026
Type: preprint
Abstract: A design note proposing a locked, sealed, multi-turn audit of whether a language model amplifies a person’s delusional or psychotic content under conversational pressure, and whether it ever appropriately slows or ends the exchange. It also proposes how to compare a modified model (for example one trained or given retrieval on therapy literature) with its own unmodified version and with pinned reference models. The note builds on existing multi-turn work on delusion and mental-health safety (psychosis-bench, SpiralBench and the Princeton audit, VERA-MH, SIM-VAIL) and sets out what it adds: matched scripts in which pressure is compared with a length-matched control; a validity gate; a rubric agreed by clinicians before anything is locked; blinded clinician coding with agreement reported; scripts, rubric, hypotheses and predictions fixed by fingerprint before any model output exists; and a sealed record that a third party can verify. It also describes a held-out and rotated script set with a contamination check, validation on synthetic models with planted behaviour, and a narrow first study. This is a design for discussion, not finished software or a set of results. The multi-turn components do not yet exist in the Symmersion toolkit. Scripts are fictional, no real client data is involved, and any audit would be a measurement of a model’s replies, not a clinical or regulatory certification. It extends the method in Operational Difference and the sealed audits in Symmersion Technical Note 1. For further information about the ENSO Framework, please contact Eric Needham:[email protected]
Author: Needham, Eric
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236726289
