spb/wp13_uqo

1% ============================================================================2\section{Introduction}3\label{sec:introduction}4% ============================================================================56Large language models (LLMs) now produce fluent commentary on earnings, valuations, and7macroeconomic conditions, and a growing literature documents their ability to extract signal8from financial text \citep{lopezlira2026chatgpt,wu2023bloomberggpt,hansen2023fedspeak}. Fluency,9however, is not rigour. A model that \emph{describes} a value-at-risk calculation has not10\emph{computed} one; a model that \emph{recalls} a price from its training data has not11\emph{retrieved} the current quote; and a model that narrates a Monte Carlo experiment has not run12the simulation whose numbers a portfolio manager would act upon. Surveys of hallucination in13natural-language generation \citep{ji2023hallucination} and of LLM evaluation14\citep{chang2024survey} make the same point from the other side: generated numbers are not15trustworthy unless they are produced, or at least checked, outside the model. The central16engineering problem in applied financial AI is therefore \emph{grounding}: how to bind a17probabilistic text generator to live data, deterministic computation, and reproducible artifacts18without giving up the conversational interface that makes the technology accessible.1920This paper describes \textbf{VQuant}, a deployed web platform built around one answer to that21problem. VQuant accepts a natural-language request---``analyse NVDA's fundamentals, technicals,22and run a Monte Carlo simulation,'' for example---and returns a multi-section report with live23data, computed figures, and downloadable documents. The platform is a full-stack TypeScript24monorepo: a React front end, a modular Express back end, and an isolated Python25scientific-computing layer. Its organising principle is a division of labour: the language model26\emph{plans and synthesises}, while deterministic server code \emph{retrieves, schedules,27computes, and accounts}. Every number in a VQuant report is meant to originate from a market-data28endpoint, a closed-form formula, or Python code that the reader can inspect. The design belongs to29the family of program-aided and tool-augmented language models30\citep{gao2023pal,chen2023pot,schick2023toolformer,yao2023react,mialon2023augmented}; its31specific contribution is to work out, at the level of a production system, what that family's32principles require when the tool surface is large, the data are live, and the deliverable is a33document.3435\paragraph{What kind of paper this is.} This is a \emph{design-science} paper in the sense of36\citet{hevner2004design} and \citet{gregor2013positioning}: it presents an artifact, states the37problem it addresses, and specifies its construction precisely enough to be reproduced and later38evaluated. It is \emph{not} an empirical evaluation. We report no measurements of numerical39accuracy against ground truth, no hallucination rates, no latency or cost benchmarks, and no40ablations; where the first version of this paper used the words ``results'' and ``robustness'', the41present version uses ``design rationale'' and ``threats to validity''. Section~\ref{sec:evaluation}42instead specifies the evaluation we think the artifact warrants---benchmark, metrics, ablations,43and baselines---and marks it explicitly as not yet executed. We regard this separation as a44matter of research integrity: a system paper that reports qualitative impressions as results45invites exactly the kind of unverifiable claim that the system itself is designed to prevent.4647\paragraph{Audit against the code.} A second feature of this version is that every mechanism48described has been checked against the platform's source code at a fixed commit49(\texttt{vquant.git}, commit \texttt{e2a148a}, 2 October 2026, hereafter ``the audited50commit''). Where the implementation differs from the platform's own documentation, or from the51first version of this paper, we say so. Three such discrepancies are material and are reported in52the text: the tool catalogue contains 245 tools rather than the ``200+'' and ``213'' figures used in53earlier descriptions; the token-accounting correction uses a general-purpose byte-pair tokenizer54whose vocabulary differs from the provider's, so the ``exact'' count of the earlier description is55an approximation; and several pre-built numerical routines (Monte Carlo, GARCH, VaR, portfolio56optimisation) have server-side handlers but are absent from the tool catalogue, so the model cannot57invoke them. Appendix~\ref{app:audit} tabulates each claim of the paper with the file that58supports it.5960\paragraph{Terminology.} Three terms recur and are used in a fixed sense. An \emph{agentic} loop is61a control structure in which the model's output can request actions (tool calls) whose results are62fed back to the model, repeatedly, until it produces a final answer63\citep{yao2023react,wang2024agents}; the loop is agentic in this operational sense and no claim64about autonomy is implied. \emph{Tool use} (or function calling) is the provider-level protocol by65which a model emits a structured call against a declared JSON schema and receives a structured66result \citep{anthropic2026tooluse,schick2023toolformer}. A \emph{reasoning-capable} model is one67that can allocate an adaptive budget of private intermediate tokens to deliberation before and68between its visible outputs \citep{wei2022chain}; the platform surfaces a summary of this69deliberation to the user but does not depend on its content.7071\paragraph{Contributions.} Within these bounds the paper makes four contributions.72\textbf{(1)~A specified orchestration loop.} Section~\ref{sec:methodology} gives the streaming73transport, the recursive tool-calling procedure, the batching throttle, and the prompt-level74routing that let a model navigate a catalogue of 245 tools without an explicit planner.75\textbf{(2)~A specified context-management procedure.} We state, as an algorithm with its actual76constants, the three-stage mechanism---per-result truncation at execution time, a re-send cap, and77adaptive summarisation under a token budget---that keeps long sessions within the model's context78window, and we relate it to the known degradation of long-context attention79\citep{liu2024lost}. \textbf{(3)~A single-call computation convention and its substrate.} We argue80that consolidating data acquisition, estimation, and visualisation into one sandboxed execution81removes model round trips, and we specify the inter-process sandbox that executes model-authored82code. \textbf{(4)~A document pipeline.} We describe how one Markdown artifact becomes PDF, DOCX,83Beamer slides, and spreadsheets with live formulas. A fifth, methodological, contribution is the84evaluation protocol of Section~\ref{subsec:protocol}.8586The classical quantitative methods the platform automates---option pricing, conditional-volatility87modelling, simulation, and risk metrics---are standard and are recorded compactly in88Appendix~\ref{app:numerical}. The remainder of the paper is organised as follows.89Section~\ref{sec:literature} situates the artifact in the literatures on tool-augmented LLMs, LLM90agents and their evaluation, and LLMs in finance and economics. Section~\ref{sec:architecture}91describes the architecture, data layer, and model layer. Section~\ref{sec:methodology} develops92the orchestration mechanisms. Section~\ref{sec:capabilities} describes the execution substrate and93the document pipeline and walks through one request as an illustration.94Section~\ref{sec:evaluation} gives the design rationale, threats to validity, reproducibility95information, and the proposed evaluation protocol. Section~\ref{sec:conclusion} concludes.96