Answers are grounded in this paper and cite their sources.
We tested whether a governed, multi-pass deliberation procedure preserves more externally specified decision structure in a language model's advisory output than a single cheap instruction to consider alternatives. The expected structure came from a regulator, not from the researchers: the three comparisons that UK FCA rule COBS 19.2.2R requires a pension suitability report to explain. Each was switched on or off by the factors of a synthetic, templated case, so it was known before any response existed. Ninety-six cases were generated under three arms with one local model (llama3.2:3b, temperature 0): plain (A); the task plus a one-sentence instruction to identify and compare alternatives (B); and a three-pass deliberate-draft-repair pipeline (C). Responses were scored by a frozen lexical instrument over 176 applicable (case, comparison) cells, under an analysis plan frozen before unblinding with a materiality margin of 0.10. Omitted cells were A 15/176 (0.0852), B 4/176 (0.0227) and C 8/176 (0.0455). On the pre-declared decisive contrast, C was numerically worse than B (-0.0227; exact McNemar p = 0.388). C therefore showed no advantage over the cheap prompt; this is not a formal equivalence or non-inferiority result. B had fewer omissions than A (+0.0625; p = 0.0127). That was the only contrast against plain to reach nominal significance, but it was secondary and descriptive, did not clear the pre-specified materiality threshold, and is qualified by four whole-response refusals in arm A. The frozen decision rules named CHEAP_PROMPT_EQUIVALENT and NO_INTERVENTION_EFFECT both fired. C used 4.11 times arm A's generated tokens and surfaced the fewest covered cells per 1,000 output tokens: its additional generation did not translate into greater measured coverage. The result is instrument-relative, uses a permissive lexical matcher, and covers one model and one rule on a synthetic population. It establishes nothing about advice quality, compliance or human judgement. The study's full procedural history is reported, including an invalid first scoring run and a provenance edge that remains unproven. Licence scope. CC BY 4.0 covers the manuscript and the non-code supplementary material released with it: figures, derived tables, claim table and verification record, evidence manifest. NOT released at this time (separate owner decision): raw generations, experiment code, held-out structure and alias pack, measurement-system source. FCA Handbook material is cited and linked, not redistributed. Verification and figure-generation code is not released in v1 (no software licence authorised). AI-assistance disclosure. AI coding agents were used to implement portions of the experimental software, execute scripted workflows, organise evidence and assist with manuscript preparation. Experimental hypotheses, protocol and governance decisions, interpretation of results, claim boundaries and final publication decisions remained the responsibility of the author. Numerical claims in this report were checked against preserved experimental artefacts rather than accepted from agent-generated summaries.
No takes yet. Share an insight, caveat, or question.
Andrew Cousins (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: