DSA: Evidence-Aware LLM-Agent Orchestration for Multi-Market Stock Research

AI summary

Overview

Research area: Artificial intelligence — specifically LLM-agent orchestration and evidence management applied to financial stock research.

Technical level: Advanced. The paper describes a systems-and-engineering framework, with vocabulary around model routing, output validation, parsers, contract testing, and profile-specific safeguards that assumes familiarity with LLM application architecture.

Scope: The paper presents DSA, a framework for organizing LLM agents, evidence, model routing, and risk controls into a multi-market stock-research workflow, and reports software contract-test evidence for its implementation rather than evidence about report quality or investment performance.

What This Paper Is About

Summarizing financial text with a large language model is not the same as running a usable stock-research operation. A real system has to gather scattered evidence from heterogeneous sources, be explicit about what data and model capabilities are not available, and govern how generated opinions are allowed to influence the final report.

DSA is the authors' answer to that engineering problem: an orchestration framework that sequences evidence acquisition, context construction, model-routed analysis, optional agent reasoning, and diagnostic report generation across multiple regional markets.

Key Contributions

1. An evidence-aware orchestration pipeline. DSA decomposes multi-market stock research into five organized stages: evidence acquisition, structured context construction, model-routed analysis, optional role and Strategy Skill reasoning, and report generation with selected context and diagnostics.

2. A two-profile design with shared services but distinct governance. A default report profile and an optional agentic profile reuse the same evidence and model-routing services, while each applies its own output validation and risk safeguards.

3. Explicit handling of disagreement and risk in the agentic profile. Core role outputs pass through role-specific parsers; Strategy Skill opinions go through an additional signal-eligibility partition before synthesis; disagreement is surfaced directly to the decision agent and followed by a conservative risk override.

4. A reference implementation plus a contract-test evidence base. The implementation covers six regional market paths, fifteen bundled Strategy Skills, hosted and local model routes, and multiple execution and delivery surfaces, and the authors report testing results from a frozen software snapshot.

Main Findings

• Contract tests passed at a frozen snapshot. A selected manifest of 1,457 portable offline backend contract tests passed at a frozen software snapshot.

• A subset was mapped to architecture-contract families. 596 of those cases were retrospectively mapped to six contract families the authors identify as central to the reported LLM-agent architecture.

• Two profiles share infrastructure but not safeguards. The default report profile and the optional agentic profile share evidence and model-routing services while maintaining profile-specific output validation and risk safeguards.

• The agentic profile adds parsing, eligibility, and override steps. Role-specific parsers handle core role outputs, Strategy Skill opinions face an additional signal-eligibility partition before synthesis, and disagreement is passed explicitly to the decision agent ahead of a conservative risk override.

• Breadth of the reference implementation. Six regional market paths, fifteen bundled Strategy Skills, both hosted and local model routes, and multiple execution and delivery surfaces are included.

• The claimed scope is explicitly limited. The abstract states that this evidence establishes implementation conformance for the tested software contracts — not superior report quality, forecasting accuracy, or investment returns. The abstract provides no quality, accuracy, or return figures.

Methodology in Plain English

The authors built a system rather than running a comparative study. Work flows through fixed stages: first collect evidence, then organize it into a structured context the models can use, then route analysis to appropriate models, optionally add role-based or Strategy Skill reasoning, and finally produce a report that carries the selected context and diagnostics alongside it.

Two operating modes sit on the same foundation. The default profile produces a report. The optional agentic profile lets specialized roles and Strategy Skills contribute opinions, but subjects those opinions to stricter handling: each role's output is parsed by a parser written for that role, Strategy Skill opinions must additionally pass a signal-eligibility partition before they can be synthesized, any disagreement is handed to the decision agent in explicit form rather than smoothed away, and a conservative risk override can intervene afterward.

To show the software behaves as specified, the authors ran a manifest of portable offline backend contract tests against a frozen snapshot of the code. They then went back and mapped a subset of those test cases to six contract families that represent the architectural commitments of the LLM-agent design. This is a conformance check on software contracts, and the authors are careful to say it does not measure whether reports are good, forecasts are accurate, or portfolios perform.

Why This Matters

Impact on research. Much LLM-for-finance work treats the model as the whole system. DSA treats the model as one routed component inside a governed pipeline, and makes the unglamorous parts — evidence assembly, unavailable data, output validation, disagreement handling — first-class design concerns. The retrospective mapping of tests to architecture contract families is also a small methodological contribution: a way to argue that an agent architecture is actually implemented as described.

Real-world applications (as implied by the framework's components):

• Multi-market equity research desks that need one workflow covering several regional market paths rather than separate tooling per market.
• Evidence-traceable research reports, where the selected context and diagnostics travel with the output so a reader can see what evidence produced a given opinion.
• Deployments with data or model restrictions, where hosted and local model routes must be interchangeable — for example when a firm cannot send certain inputs to an external provider.
• Human-in-the-loop advisory workflows, where disagreement between analytical roles should be surfaced to a decision-maker rather than averaged into a single confident-sounding conclusion.

Industry relevance. The dual-profile design — a plain report mode and a stricter agentic mode with signal eligibility and a conservative risk override — maps onto how financial institutions typically want to separate routine summarization from anything that can influence a decision. The emphasis on validation and safeguards, plus the authors' refusal to claim performance gains, speaks to a regulated environment where demonstrating that a system does what its specification says is a distinct and necessary form of evidence.

Future Directions

• Evaluating the outputs, not just the contracts. The abstract explicitly disclaims claims about report quality, forecasting accuracy, and investment returns. Whether DSA's evidence-aware pipeline and risk override actually improve any of these is the obvious open question and is not addressed.

• Broadening contract coverage. Only 596 of the 1,457 passing tests were mapped to the six architecture contract families. Extending that mapping, and defining contract families for areas not yet covered, would strengthen the conformance argument.

• Testing the agentic profile under adversarial or degraded conditions. The signal-eligibility partition, explicit disagreement handling, and conservative risk override are described as mechanisms; how they behave when evidence is missing, conflicting, or deliberately misleading is not reported.

• Scaling markets, skills, and routes. With six regional market paths and fifteen bundled Strategy Skills, questions remain about how the framework extends to additional markets, how new Strategy Skills are validated, and what the operational cost is of running both a default and an agentic profile over shared services.

Target Audience

AI engineers and applied researchers building LLM-agent systems for financial or other evidence-heavy domains, who will find the orchestration stages, dual-profile design, and contract-test approach directly reusable. Financial-technology architects and quantitative research platform owners evaluating how to structure agentic research tooling with validation and risk controls will also benefit. Readers looking for demonstrated accuracy, return, or report-quality improvements should note that the abstract does not offer them.

Authors’ abstract

Large language models can summarize financial information, but an operational stock-research system must first assemble heterogeneous evidence, expose unavailable data and model capabilities, and control how generated opinions affect a final report. We present DSA, an evidence-aware orchestration framework for multi-market stock research with large language model (LLM) agents. DSA organizes the workflow into evidence acquisition, structured context construction, model-routed analysis, optional role and Strategy Skill reasoning, and report generation with selected context and diagnostics. A default report profile and an optional agentic profile share evidence and model-routing services but use profile-specific output validation and risk safeguards. In the agentic profile, core role outputs are processed by role-specific parsers, whereas Strategy Skill opinions undergo an additional signal-eligibility partition before synthesis; disagreement is supplied explicitly to the decision agent, followed by a conservative risk override. The reference implementation includes six regional market paths, fifteen bundled Strategy Skills, hosted and local model routes, and multiple execution and delivery surfaces. At a frozen software snapshot, a selected manifest of 1,457 portable offline backend contract tests passed; 596 cases were retrospectively mapped to six contract families central to the reported LLM-agent architecture. This evidence establishes implementation conformance for the tested software contracts, not superior report quality, forecasting accuracy, or investment returns.