Research

On AI Safety

We do not build foundation models. We use them, principally Anthropic Claude Sonnet at the Institution tier, with substantial system-prompt engineering and a verification layer that constrains the model to verified data. The principal AI safety question for a firm in our position is therefore not the alignment of frontier models, which is the work of the laboratories; it is the safe deployment of those models in a domain where their failure modes have specific consequences.

Two failure modes preoccupy us. The first is hallucination: the production by the model of a confidently-stated assertion not grounded in verified data. Our verification layer is the principal defence; the system prompt that instructs the model to flag what it does not know is the secondary. The second is regulatory drift: the temptation, given a sufficiently capable model, to allow it to slide into producing investment advice rather than analysis. Our system prompts include explicit refusal instructions; our internal review examines model output for signs of such drift.

We follow the published work of Anthropic, OpenAI, DeepMind, and the academic literature on language-model safety with seriousness. We are not contributors to that literature, but we are attentive consumers of it.

On the Limits of Quantitative Modelling

Our scenario modelling and risk computation rest on assumptions about the distribution of future returns, the persistence of correlations, and the representativeness of the historical record. Each of these assumptions can fail. The 2008 financial crisis, the 2020 pandemic-induced volatility, and the 2022 rate-cycle break each exposed the limits of models calibrated on prior regimes.

Our position is that this is not an argument against quantitative modelling but an argument for stating its limits explicitly. The methodology documentation sets out the assumptions on which each computation rests; the user is invited to weigh the output accordingly. A model whose limits are visible is more useful than a model whose confidence is artificially high.

On the AI-Generated Content Question

Substantial portions of Drusus Daily are produced by a language model. We have considered whether this should be disclosed to readers in a more prominent fashion than the methodology documentation already does, and have concluded that the structural answer is yes: the model-generated nature of the prose is declared on every edition. We do not, however, identify which specific sentences are model-generated and which are human-edited, because the line is in practice blurred (the editorial review revises model output materially) and because the binary disclosure would mislead more than it would inform.

The relevant question is not whether prose is generated by a model but whether the figures it contains are correct. The verification layer addresses this. The editorial review is responsible for the analytical judgement.

Research We Publish

Our long-form research appears in Drusus Insights. The pieces published there are positioned for the working investor and the analytical professional, not for the academic literature; we make no claim to be conducting peer-reviewed research. Where our methodology contributes to academic work conducted by users of the platform, the citation guidance is set out on the methodology page.