# Beyond LLM Wrappers: Context Engineering and the Three-Engine Architecture for Reliable Financial AI [Canonical reading](https://escapemeridian.com/readings/beyond-llm-wrappers-context-engineering-and-the-three-engine-f146d03f) Authors Pradeep Kumar GS and Ravi Tanniru argue that financial AI needs more than an LLM responding to text: it must support financial decisions with numerical accuracy, traceable evidence and auditability. Their proposed architecture combines a financial knowledge graph, a deterministic optimization engine and a multi-agent LLM layer. Context engineering coordinates query-specific information from enterprise systems, tools and tenants, then validates it. The authors report observations from 15 deployments over approximately 11 months: single-agent pipelines had hallucination rates of about 7–9%, while graph-backed multi-agent validation reduced observed rates to below 1%. They explicitly caution that these findings are observational rather than causal. Reusable mappings, graph templates and validation policies coincided with deployment time falling from roughly six to two weeks. The argument matters because financial answers can affect payments, collections, cash flow and compliance; a fluent answer alone does not ensure correct calculations or supported recommendations. The authors also identify ERP data quality and tenant-specific schemas as major implementation challenges. They call for further research into evaluation, reproducibility, privacy, context quality and cross-tenant generalization. How well does this architecture generalize across tenants with different ERP ( Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Oct 29, 2026 (2026-10-29). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-11T08:09:32.093Z. First observed: 2026-10-30T06:41:12.932Z. Retrieved: 2026-10-31T12:19:50.829Z. - [Beyond LLM Wrappers: Context Engineering and the Three-Engine Architecture for Reliable Financial AI](https://arxiv.org/abs/2610.39230) ## Limits The reported deployment observations are not causal evidence; the abstract does not establish that graph-backed multi-agent validation caused the lower observed hallucination rates. The abstract does not provide further detail on how the deployments or hallucination rates were measured. The authors identify ERP data quality and tenant-specific schemas as implementation challenges; the abstract does not establish how broadly the architecture generalizes across tenants. ## Open question How well does the architecture generalize across tenants with different ERP schemas, a research challenge the authors identify? ## Source-stated The authors propose an architecture combining a financial knowledge graph, a deterministic optimization engine and a multi-agent LLM layer, coordinated by context engineering. > We present a production-grounded architecture integrating three components: a financial knowledge graph for enterprise relationships, temporal context, and provenance; a deterministic optimization engine - [Supporting source](https://arxiv.org/abs/2610.39230) ## Source-stated Across observations from 15 deployments over approximately 11 months, the authors report lower observed hallucination rates with graph-backed multi-agent validation than with single-agent pipelines, while noting the findings are not causal evidence. > Graph-backed multi-agent validation reduced observed rates to below 1%, although these findings are observational rather than causal. - [Supporting source](https://arxiv.org/abs/2610.39230) ## Source-stated The authors report that reusable mappings, graph templates and validation policies coincided with shorter deployment time. > Reusable mappings, graph templates, and validation policies coincided with a reduction in deployment time from roughly six to two weeks. - [Supporting source](https://arxiv.org/abs/2610.39230) ## Source-stated The authors identify ERP data quality and tenant-specific schemas as major implementation challenges. > ERP data quality and tenant-specific schemas remained major implementation challenges. - [Supporting source](https://arxiv.org/abs/2610.39230) ## Source-stated The authors argue financial context engineering requires more than text-centric retrieval-augmented generation, including temporal validity, exact numerical evidence and validation. > requiring structured relationships, temporal validity, exact numerical evidence, access controls, and validation against source records and solver outputs. - [Supporting source](https://arxiv.org/abs/2610.39230) ## Source and contributor context GS, Pradeep Kumar, Tanniru, Ravi · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Hallucination When an AI system states something false or made up as if it were true. - [Definition reference](https://developers.google.com/machine-learning/glossary#hallucination) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 4ba6ffc1-055b-4da2-8a9d-97b920264b1c. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T08:09:32.093Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 7ed30440d54be9e0451b1b887390640880d7640d9c849677bd59e16cc52e57e5. Source text hash: 698ef94c266667ffcb2cafbf5599b323dd07cea1ed2009f8f045c97ddee5237f. ## Revision and source-check history ### Version 1 — 2026-11-11T08:09:32.093Z Authors Pradeep Kumar GS and Ravi Tanniru argue that financial AI needs more than an LLM responding to text: it must support financial decisions with numerical accuracy, traceable evidence and auditability. Their proposed architecture combines a financial knowledge graph, a deterministic optimization engine and a multi-agent LLM layer. Context engineering coordinates query-specific information from enterprise systems, tools and tenants, then validates it. The authors report observations from 15 deployments over approximately 11 months: single-agent pipelines had hallucination rates of about 7–9%, while graph-backed multi-agent validation reduced observed rates to below 1%. They explicitly caution that these findings are observational rather than causal. Reusable mappings, graph templates and validation policies coincided with deployment time falling from roughly six to two weeks. The argument matters because financial answers can affect payments, collections, cash flow and compliance; a fluent answer alone does not ensure correct calculations or supported recommendations. The authors also identify ERP data quality and tenant-specific schemas as major implementation challenges. They call for further research into evaluation, reproducibility, privacy, context quality and cross-tenant generalization. How well does this architecture generalize across tenants with different ERP ( ## Original source arXiv cs.AI · Article Original source date: Oct 29, 2026 (2026-10-29). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-11T08:09:32.093Z. First observed: 2026-10-30T06:41:12.932Z. Retrieved: 2026-10-31T12:19:50.829Z. - [Beyond LLM Wrappers: Context Engineering and the Three-Engine Architecture for Reliable Financial AI](https://arxiv.org/abs/2610.39230) ## Limits The reported deployment observations are not causal evidence; the abstract does not establish that graph-backed multi-agent validation caused the lower observed hallucination rates. The abstract does not provide further detail on how the deployments or hallucination rates were measured. The authors identify ERP data quality and tenant-specific schemas as implementation challenges; the abstract does not establish how broadly the architecture generalizes across tenants. ## Open question How well does the architecture generalize across tenants with different ERP schemas, a research challenge the authors identify? ## Source-stated The authors propose an architecture combining a financial knowledge graph, a deterministic optimization engine and a multi-agent LLM layer, coordinated by context engineering. > We present a production-grounded architecture integrating three components: a financial knowledge graph for enterprise relationships, temporal context, and provenance; a deterministic optimization engine - [Supporting source](https://arxiv.org/abs/2610.39230) ## Source-stated Across observations from 15 deployments over approximately 11 months, the authors report lower observed hallucination rates with graph-backed multi-agent validation than with single-agent pipelines, while noting the findings are not causal evidence. > Graph-backed multi-agent validation reduced observed rates to below 1%, although these findings are observational rather than causal. - [Supporting source](https://arxiv.org/abs/2610.39230) ## Source-stated The authors report that reusable mappings, graph templates and validation policies coincided with shorter deployment time. > Reusable mappings, graph templates, and validation policies coincided with a reduction in deployment time from roughly six to two weeks. - [Supporting source](https://arxiv.org/abs/2610.39230) ## Source-stated The authors identify ERP data quality and tenant-specific schemas as major implementation challenges. > ERP data quality and tenant-specific schemas remained major implementation challenges. - [Supporting source](https://arxiv.org/abs/2610.39230) ## Source-stated The authors argue financial context engineering requires more than text-centric retrieval-augmented generation, including temporal validity, exact numerical evidence and validation. > requiring structured relationships, temporal validity, exact numerical evidence, access controls, and validation against source records and solver outputs. - [Supporting source](https://arxiv.org/abs/2610.39230) ## Source and contributor context GS, Pradeep Kumar, Tanniru, Ravi · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Large language model A language model with many learned numerical settings, used to process and generate language. - [Definition reference](https://developers.google.com/machine-learning/glossary#large-language-model) ## Hallucination When an AI system states something false or made up as if it were true. - [Definition reference](https://developers.google.com/machine-learning/glossary#hallucination) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## TIME TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services. - [Explore this context](/entities/time) - [Reference](https://time.com/) ## Editorial provenance Version 1. Publication ID: 4ba6ffc1-055b-4da2-8a9d-97b920264b1c. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T08:09:32.093Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: 7ed30440d54be9e0451b1b887390640880d7640d9c849677bd59e16cc52e57e5. Source text hash: 698ef94c266667ffcb2cafbf5599b323dd07cea1ed2009f8f045c97ddee5237f. ### uncertain — 2026-11-11T08:09:32.093Z — anthropic/claude-opus-5.5 Publication: 4ba6ffc1-055b-4da2-8a9d-97b920264b1c attribution — pass: The architecture, deployment observations and arguments are attributed to authors Pradeep Kumar GS and Ravi Tanniru. The draft keeps the hallucination figures as their reported observations, not established facts. > Authors:Pradeep Kumar GS, Ravi Tanniru summary — uncertain: Two summary problems. First, the summary ends with a truncated fragment: 'How well does this architecture generalize across tenants with different ERP ('. Smallest repair: delete it, since the question field already covers this. Second, central terms are not explained in plain language, including ERP (enterprise resource planning systems), tenants (separate client organizations sharing a system), hallucination, knowledge graph and deterministic optimization engine. Without tooltips or brief glosses, a non-specialist reader may not follow the architecture or the challenges. Otherwise, the summary faithfully conveys the thesis and the authors' causal caveat. > ERP data quality and tenant-specific schemas remained major implementation challenges. date — pass: originalPublished is 2026-10-29. This matches the arXiv v1 submission date, not the later feed date of Oct 31. > \[Submitted on 29 Oct 2026\] support — pass: All claims are directly supported by abstract excerpts, and the caveat that the findings are observational rather than causal is preserved. The 'matters' paragraph restates the source's stated stakes about payments, compliance and LLMs' inability to guarantee calculations. The question draws on the stated research challenge of cross-tenant generalization. Minor omission: the 5-10% range across query types is not mentioned, but this does not change the meaning. > Large language models (LLMs) enable natural language interaction but cannot independently guarantee reproducible calculations or evidence-grounded recommendations. media — pass: The selected media is the verified direct link to the original arXiv abstract page, which is the most direct access to the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/f146d03fe62686cc8c40eeb5dc860f70505916752a50d99ff4c30819a5b60afe/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/beyond-llm-wrappers-context-engineering-and-the-three-engine-f146d03f.json)