UQO Working Papers · WP13 · v1.1
VQuant: An Agentic Large Language Model Architecture for Automated Quantitative Financial Analysis and Document Generation
Reasoning-Augmented Tool Orchestration, In-Loop Numerical Computation, and Multi-Format Reporting
- Year
- 2026
- First release
- October 2026
- Latest revision
- October 2026
- Publication status
- Working paper
The model plans and writes; auditable code computes every number
VQuant exposes 245 tools, two market-data providers (249 functions over 233 endpoints), three web-research services and a sandbox with 29 scientific packages through a recursive, batched, streaming tool-calling loop; the paper specifies the architecture and reports no accuracy, latency or cost measurements.
Abstract
Large language models write fluently about finance, but fluency does not make a number correct, current, or reproducible. This paper is a design-science description of VQuant, a deployed web platform that couples a reasoning-capable LLM with a deterministic execution substrate so that a natural-language request ("analyse NVDA's fundamentals and run a Monte Carlo simulation") is turned into a report in which every figure is produced by live data retrieval or auditable code rather than by the model.
Four mechanisms are documented as implemented and audited line by line against the platform's source code (vquant.git, commit e2a148a): (i) an agentic orchestration loop that exposes 245 schema-defined tools, wrapping two market-data providers (249 typed functions over 233 endpoint paths), three web-research services, and a Python sandbox pre-loading 29 scientific packages, through a recursive, batched, streaming tool-calling protocol; (ii) a three-stage context-management procedure (per-result truncation at execution, a re-send cap, and adaptive summarisation under a token budget) that keeps long sessions within the model's window; (iii) a single-call computation convention that consolidates data acquisition, estimation, and figure rendering into one sandboxed execution; and (iv) a document pipeline that renders one Markdown artifact into PDF, DOCX, LaTeX Beamer, and spreadsheets with live formulas.
The paper reports no empirical evaluation: it contains no measurements of numerical accuracy, hallucination rates, latency, or cost, and makes no comparative claims. Its contribution is a reproducible specification of the architecture, an account of where the implementation departs from its own documentation, and a pre-registered evaluation protocol (benchmark, metrics, ablations, and baselines) that has not yet been executed.
Key findings
245
schema-defined tools exposed to the model, 17 families
249 / 233
typed market-data functions over distinct endpoint paths
29
scientific packages pre-loaded by the Python sandbox
3 stages
context management: truncation, re-send cap, summarisation
4
output formats from one Markdown artifact
0
empirical measurements reported; evaluation protocol pending
Scientific contributions
- An implementation-level specification of an agentic loop
- A recursive, batched (5 calls, 500 ms spacing), streaming tool-calling protocol over 245 tools with tool-use/tool-result pairing and prompt-level routing by request complexity, documented with its actual constants (Algorithm 1).
- Context management with real constants
- Per-result truncation at 3,000/15,000 characters, a 5,000-character re-send cap and adaptive summarisation at a 10^5-token budget keeping the last 5 messages; the local token count is a BPE approximation, not an exact tokenizer, as the audit established.
- A single-call computation convention
- Data acquisition, estimation and figure rendering are consolidated into one execution in a process-isolated Python child with a 90-second timeout and a 150-dpi figure sweep, so that no numeric value originates from the model.
- One artifact, four documents
- A single Markdown artifact is rendered to PDF (Puppeteer and KaTeX), DOCX (Pandoc), LaTeX Beamer (a second model call and two pdflatex passes) and XLSX with live formulas.
- An audit that corrects the system's own description
- Line-by-line comparison with commit e2a148a corrected the tool count, the tokenizer claim, the reachability of pre-built numerical routines (only the option pricer is in the catalogue), the multi-provider model layer, and removed three features (SSO, daily caps, bring-your-own-keys) absent from any committed version; all unmeasured latency claims were withdrawn.
- A pre-registered evaluation protocol
- Task suite, metrics including an unsupported-number rate, ablations of the single-call convention and summarisation, baselines and controls are specified in Section 6.6 and have not been executed; any claim of superiority waits on them.
Methodology
Methodology (4)
Design-science framing
Following Hevner et al. (2004) and Gregor and Hevner (2013), the paper contributes a situated implementation and a demonstration; the evaluation step is specified but explicitly not performed.
Mechanism specification
Transport, recursive loop (Algorithm 1), batching, routing, three-stage context management (Algorithm 2), token-accounting correction and the single-call computation convention are stated at implementation level with their constants.
Claim-by-claim code audit
Every architectural statement is mapped to a source file and line in an appendix table generated from a read-only clone of the platform repository; statements contradicted by the code were corrected or removed.
Proposed evaluation protocol
A benchmark of financial-analysis tasks, metrics (numerical accuracy, unsupported-number rate, latency, cost), ablations and baselines against simpler agents, written as a pre-registration for future execution.
Research data
- The VQuant source code (no empirical data)
- The paper describes a software system rather than a dataset: the TypeScript/Python platform maintained in the spbgit repository vquant.git at commit e2a148a (2 October 2026), deployed on the MacLustr cluster. Audit facts are extracted by docs/verify/audit_vquant_code.py into code_facts.json and tool_families.json.
Reproducibility
- docs/verify/audit_vquant_code.py regenerates paper/tables/tools.tex and the code facts quoted in the text from a clone of vquant.git at commit e2a148a
- No empirical pipeline, data or measurements: the paper has no src/, scripts/ or data/ directories by design
- Full LaTeX source of the 41-page paper with TikZ architecture diagrams compiled inline; pre-revision v1.0 PDF, review report and changelog in docs/
- Appendix E maps every claim to a source file; the evaluation protocol of Section 6.6 is specified for future execution
Research artifacts
Version history
| Version | Date | Description | File |
|---|---|---|---|
| v1.1current | October 2026 | Current manuscript as published on www.spboucher.ai (revised edition after the 2026-10-17 scientific review for papers 1–17). | PDF · 41 p. |