VADAOrchestra: Neurosymbolic Orchestration of Adaptive Reasoning Workflows

Hi, I’m Keito Inoshita from Affectosphere Group.

“Why did this AI reject the application?”

That’s the question that compliance teams and internal auditors in financial institutions need to be able to answer — every time, with specifics, not summaries.

LLMs are genuinely getting smarter. But “smart” and “explainable” are different things.

The reasoning inside an LLM is fundamentally opaque. When you ask it to explain its conclusion, it produces a natural-language explanation — but that explanation is itself generated by the model, and may not accurately reflect what actually drove the output.

This is the core problem blocking serious LLM adoption in regulated finance workflows.

A study published on arXiv in June 2026 (arXiv:2606.22485) proposes a design-level answer: VADAOrchestra — a neurosymbolic architecture that combines logic programs with LLMs to make reasoning traceable without discarding LLM capability.

---

Today’s 3 Points

1. VADAOrchestra expresses reasoning steps as logic programs and uses LLMs in a complementary role, enabling full traceability of decision rationale.
2. Neurosymbolic design addresses the explainability gap that makes standalone LLMs difficult to deploy in credit scoring and fraud detection.
3. When decision rationale is output as logic programs, structured audit logs emerge naturally — satisfying internal audit and regulatory requirements by design.

---

① Why Standalone LLMs Are Hard to Use in Finance

Let’s start with the problem structure.

LLMs are remarkably capable at reasoning over text. Given the inputs relevant to a credit assessment, they can produce plausible-sounding judgments. But when asked “why did you make this call?”, an LLM cannot provide a reliable, traceable account of its reasoning.

In financial practice, this creates three concrete problems.

First, credit decisions. In many jurisdictions, lenders must disclose the basis for credit rejections to applicants. A generated natural-language explanation isn’t the same as an auditable decision rationale.

Second, fraud detection. When a transaction is flagged as fraudulent, internal audit and regulators need to see why — which specific signals triggered the flag, and how they were combined. LLM outputs don’t provide that in an inspectable form.

Third, fairness verification. Checking whether an AI credit model produces statistically biased outcomes requires tracing how individual variables contributed to specific decisions. That trace is not available from an LLM’s output alone.

“Smart but opaque” is insufficient for these use cases. That’s the honest reality.

---

② VADAOrchestra’s Core Idea: Making Reasoning Visible with Logic Programs

VADAOrchestra’s proposal is to change the architecture of reasoning itself.

The core idea is to express decision steps as logic programs.

Logic programs (symbolic AI) represent rules explicitly: conditions, and what follows from them. For example:

“If the applicant’s account balance over the past three months falls below the threshold AND there is a history of payment delays, then raise a credit risk flag.”

This rule is readable by humans. Each step is traceable: which conditions were evaluated, which were satisfied, what conclusion was drawn. The logic is inspectable, not inferred.

LLMs handle the parts that symbolic reasoning struggles with — interpreting natural language, extracting structured information from unstructured documents, handling cases that don’t fit neatly into predefined rules. They play a complementary role.

The division of labor is clear: the logic program provides the decision skeleton; the LLM supports the processing that feeds it.

This architecture means decision rationale is output as logic program traces. Which rules were applied, which conditions were met, how the conclusion was reached — all of this is available as a structured log, not a generated narrative.

---

③ What “Adaptive” Means Here

The paper’s emphasis on “adaptive reasoning workflows” is also worth unpacking.

Financial assessments are rarely uniform. A simple case might resolve in a few steps. A complex or edge-case application might require additional information gathering, application of alternative criteria, or escalation to a different review track.

VADAOrchestra’s orchestration layer handles this dynamic workflow management. It decides which reasoning modules to invoke based on the current case complexity and state — adapting the reasoning path to the situation.

The result is a system that is adaptive in its process but traceable in its trace. Every path it takes is recorded as a logic program execution log, regardless of how complex the case was.

This combination — adaptability plus traceability — is what the architecture is designed to achieve simultaneously.

---

④ What Financial Institutions Can Consider Now

How does this translate into concrete financial practice?

Starting with credit scoring. When integrating LLMs into an existing assessment system, rather than delegating the full judgment to the LLM, consider a VADAOrchestra-style design where assessment criteria are first formalized as explicit logic rules, with the LLM playing a supporting role.

Formalizing assessment criteria into rules may look like upfront cost — but it’s actually the work of translating existing assessment manuals and internal guidelines into explicit rule form. The institutional knowledge that assessors have (“in this kind of case, we apply this standard”) gets externalized as inspectable rules. That process itself tends to surface inconsistencies in how assessment standards have been applied.

For fraud detection, the need for explainability is even more acute. When a decision rationale is output as a logic program trace, reviewing disputed cases becomes structurally more efficient. And because traces are structured records, statistical auditing of whether the model produces systematically biased outcomes becomes tractable.

For compliance and internal audit teams, “auditing AI decision logs” is an emerging function. In a neurosymbolic architecture, those logs are structured rule application records rather than natural-language outputs. The difference in auditability is fundamental.

Practical KPIs to track: AI decision audit completion rate, and average time required to produce a decision rationale trace. Both metrics can demonstrate quantifiable value from the architecture shift.

---

Explainable AI Is the Real Frontier for Finance

The desire to deploy AI in financial operations is real and widespread. The bottleneck isn’t ambition — it’s trust in the reasoning.

Deploying AI that cannot explain its judgments in compliance-critical workflows runs into a rational wall. That wall won’t move by making LLMs smarter alone.

VADAOrchestra demonstrates that “explainable” and “capable” are not mutually exclusive. Combining LLM capability with symbolic reasoning produces an architecture that doesn’t sacrifice intelligence for traceability — or traceability for intelligence.

That design philosophy is, I think, one of the important directions for serious AI adoption in finance.

---

Reference

1. Brüggemann, T., Grätzer, R., Knott, A., Mitrevski, M., Morales, G., & Simeonova, B. (2026). VADAOrchestra: Neurosymbolic Orchestration of Adaptive Reasoning Workflows. arXiv preprint arXiv:2606.22485.

Note: This article was partially written with AI assistance and may contain inaccuracies.