Manifund foxManifund
Home
Login
About
People
Categories
Newsletter
HomeAboutPeopleCategoriesLoginCreate
email avataremail avatar
Greg R. Welch

@email

Greg R. Welch is an independent publisher and information designer developing the AI Source Analysis Framework and Source Architecture for AI Reasoning. His work examines how source representation affects AI reasoning over consequential documents, with a focus on relationships, provenance, boundaries, uncertainty, and claim limits. He brings more than three decades of experience in publishing, information design, and digital information systems to this source-side approach to AI research.

https://www.linkedin.com/in/grwelch/
$0total balance
$0charity balance
$0cash balance

$0 in pending offers

About Me

I came to AI research through publishing and information design rather than machine-learning engineering. For more than three decades, my work has focused on how information is structured, presented, and made usable by people. That background led me to ask what changes when AI becomes an intended reader and reasoner over the same material.

I have been developing the AI Source Analysis Framework independently, using repeated comparative testing and failure analysis to revise the architecture rather than treating the current design as fixed. I am particularly interested in source-grounded reasoning, evidentiary relationships, provenance, boundaries, uncertainty, and how small interpretive departures can propagate through later AI reasoning.

I work independently from Elko, Nevada, and am currently focused on developing, testing, and publishing this line of research.

Projects

Source Architecture for AI Reasoningpending grant agreement signature

Comments

Source Architecture for AI Reasoning
email avatar

Greg R. Welch

7 days ago

25AUG26: Project update — external technical comparison and next research stage

Since posting this proposal, I have identified several public technical programs that provide useful external comparison points for the research question described here.

I developed the AI Source Analysis Framework independently and encountered Microsoft Research’s GraphRAG only after developing the Framework and submitting these proposals. I am not making a priority claim; what matters is the independent convergence on a related problem.

GraphRAG, Anthropic’s Contextual Retrieval, and OpenAI’s knowledge-retrieval work all address, in different ways, how external information is structured and supplied to models for later reasoning.

The comparison has sharpened the question behind this project:

What persistent structure is the source actually entitled to support before that structure is supplied to later AI reasoning?

This matters because model-generated interpretation can itself become persistent—as a relationship, graph edge, contextual annotation, summary, or classification that later models may retrieve and reason from.

The issue is therefore not simply whether structured information is useful. It is how that structure earned its place there.

This does not change the Manifund proposal. The project remains an investigation of source representation as a variable in AI reasoning, including the possibility that simpler approaches may perform just as well.

What has changed is that there is now a clearer external comparison. A matched evaluation could compare raw or conventionally retrieved source material, automatically generated graph structure, Framework-produced bounded analytical records, and potentially the same Framework findings serialized into graph-compatible form.

That would help distinguish:

Does structured representation affect reasoning?

from:

Does the way that structure was established affect reasoning?

The Framework has also reached a deliberately stabilized manual research stage. The next question is not simply whether the workflow can be automated, but which analytical invariants must survive automation—including source boundaries, relationship warrant, qualification, claim limits, provenance, and unresolved or rejected relationships.

This gives the proposed work a more concrete next stage: machine-readable representation, selective automation, controlled comparison, and evaluation of downstream reasoning, error propagation, variability, cost, and throughput.

I am not claiming that the Framework will outperform GraphRAG, conventional RAG, or simpler source treatments. That is what the proposed evaluation would need to determine.

— Greg R. Welch