AI Agents Propose, Code Verifies, Humans Decide

AI agents can propose, deterministic code can verify, and humans can decide. The research team described by a Grafy Research Lab source presented an AI-assisted pipeline for ingesting heterogeneous scientific tables into relational databases, according to the source posting.

Public scientific datasets rarely arrive ready for analysis. According to the source, they often arrive as messy folders containing wide matrices, long tables, spreadsheets and per-cohort files. Their keys and encodings differ, and every reanalysis has to reconnect them first. A coding agent can automate that work, but the source warns that agent mistakes can look plausible and that without an audit trail a mistake capable of corrupting data can go uncaught. The source framed this as a problem for research data. The source material comes from a feed entry and source page, and it describes a research effort. The central claim is procedural: let models propose, code verify, humans decide. The source did not describe results or benchmarks in the retrieved material.

What problem was the source addressing?

The source described a problem around data ingestion for scientific research. According to the source, public scientific datasets rarely arrive analysis-ready. They arrive as messy folders of wide matrices, long tables, spreadsheets and per-cohort files whose keys and encodings differ. Every reanalysis must reconnect them first. The source said a coding agent can automate this, but its mistakes can look plausible, and without an audit trail a mistake that corrupts data can go uncaught. The work described an AI-assisted ingestion pipeline as a response. The source's feed summary described public scientific datasets rarely arriving analysis-ready. They arrived as messy folders of wide matrices, long tables, spreadsheets, and per-cohort files whose keys and encodings differ. Every reanalysis had to reconnect them first. The source said a coding agent can automate this, but mistakes can look plausible, and without an audit trail, a mistake that corrupts data can go uncaught. That combination of messiness and plausible error is the problem the source addresses.

How does the propose, verify, decide pipeline work?

The source said the pipeline follows a rule: LLM agents propose, deterministic code verifies, and humans decide. Under that rule, agents generate or propose ingestion steps, code checks them, and people retain final authority. The source also said narrow agents see compact, citable evidence about the files and do not see the data itself. The source did not provide the full implementation details in the retrieved material. The design therefore rests on a human checkpoint after automated proposal and verification. The source said the pipeline is built on one rule. LLM agents propose. Deterministic code verifies. Humans decide. The source said narrow agents see compact, citable evidence about the files, never the data. The source did not detail the return format or the exact deterministic checks. The human decision step is what the source presents as the control point. The source did not say how often humans review or what training was used.

What did the source say about the research?

According to the source, the work is about verified agentic ingestion of heterogeneous scientific tables into relational databases. The source listed the paper as having seven pages and four figures, and it reported seven comments. The source named Bharatha Rankothge, Suzanne Hodgkins, Virginia Rich and Michael Kirby. The source listed the publication date as 2026-11-10 and placed the work in databases and artificial intelligence subject areas. The source did not provide benchmark results in the retrieved material, so the news should not claim performance numbers. The source attributed the work to Grafy Research Lab. The source listed authors. The source gave a title. The source gave subject terms. The source gave publication date. The source gave comments count, pages, figures. The source did not give results. The source did not give quotes. The source did not give deployment details. The source did not give evaluation metrics.

How could this fit into AI-assisted research workflows?

For a workflow built around AI coding and research, the closest Grafy surface is Code, the AI coding studio. A developer could prototype deterministic checks around an agent's proposals. The source does not say Grafy has a dedicated scientific table ingestion product, so the practical lesson may be to keep human review and code-based verification in the loop. Reader is another listed surface for studying papers and source material. The Grafy News Desk would treat the pipeline as a design pattern rather than a finished product announcement. The connection to Grafy is genuine but limited. The listed Code app is an AI coding studio. The listed Reader app is for papers. The source is a research paper about ingestion. The practical takeaway is to keep deterministic verification and human oversight. Grafy users could use Code to build checks. They could use Reader to read the paper. The Grafy News Desk would not claim Grafy already ships this exact pipeline. Instead, it would frame the source as a design pattern for AI-assisted work.

Frequently asked questions

Does the verified agentic ingestion method teach smaller AI models?

The source does not say that. It describes a pipeline in which language model agents propose and deterministic code verifies, while humans decide. The narrower agents receive compact, citable evidence rather than the underlying data. Because the source does not describe teaching smaller models, the answer to whether this method teaches them is not supported by the material.

When do humans decide in the proposed pipeline?

According to the source, humans decide after the proposal and verification steps. The pipeline rule is that agents propose, deterministic code verifies, and humans decide. The source does not specify the exact timing or interface for that final human choice, but the ordering is central to the described method.

What does the source say about AI agents and scientific table ingestion?

The source says public scientific datasets rarely arrive analysis-ready. They often come as messy folders of wide matrices, long tables, spreadsheets and per-cohort files with differing keys and encodings. A coding agent can automate reconnection, but the source warns that mistakes can look plausible and without an audit trail corrupted data may go uncaught.

---

Source: Grafy Research Lab — Agents Propose, Code Verifies, Humans Decide: Verified Agentic Ingestion of Heterogeneous Scientific Tables into Relational Databases