You're pledging to donate if the project hits its minimum goal and gets approved. If not, your funds will be returned.
AI systems increasingly reason from regulations, research, technical documentation, policies, and other sources written primarily for human readers. Those sources often leave relationships, boundaries, provenance, uncertainty, temporal state, and other reasoning-relevant information partly implicit.
I am investigating whether making selected source structure explicit before AI reasoning changes downstream behavior.
I have developed a working source-side framework and begun comparative testing against source-only conditions. Early development tests have shown cases in which a targeted reasoning departure appeared and compounded in a source-only condition but did not appear during the corresponding structured source-and-reasoning condition.
These are bounded development observations, not proof of general effectiveness.
The project will test a broader proposition:
When AI becomes an intended reasoning consumer, source representation itself becomes part of the reasoning environment.
The goal is to determine whether this is a useful and measurable reliability variable, which kinds of structure matter, when simpler alternatives work just as well, and where the approach fails.
The immediate research goal is to determine whether explicit source structure can reduce unsupported reconstruction when AI systems reason over consequential information.
A source can support two facts without supporting a relationship between them. It can describe an intended result without establishing that the result occurred. A statement can apply within one boundary without supporting a broader conclusion. During repeated summarization and conversation, these distinctions can be lost while the resulting AI output remains fluent and plausible.
I use drift to describe the propagation of earlier reasoning departures into later reasoning.
During the project I will conduct comparative tests across different source types and AI systems; identify and reproduce failures involving relationships, scope, provenance, uncertainty, temporal state, and claim strength; compare structured-source conditions with source-only conditions and simpler alternatives; test cases in which added structure may overconstrain or worsen reasoning; and document negative or null results alongside successful demonstrations.
A recent development test used the European Union Artificial Intelligence Act. Two AI environments received the same underlying source and conversational sequence under different source-and-reasoning conditions. A targeted reasoning departure appeared early and compounded in the source-only condition. It did not appear during the tested sequence in the structured condition.
That test does not isolate every component or establish a general effect. Its value is that source representation can be changed, tested, and revised rather than treated as an invisible constant.
The longer-term question is:
What would source information look like if AI reasoning were considered during source creation itself?
Today's normal sequence is human-oriented source followed by AI interpretation. Future information systems may sometimes benefit from richer underlying source representations that preserve reasoning-relevant distinctions while still producing ordinary human-readable publications.
Success for this project does not require validating my current framework. A useful result could be evidence that only certain structural distinctions matter, that simpler methods perform equally well, or that the approach fails under identifiable conditions.
I am seeking $50,000 for approximately twelve months of independent research and development.
Most funding would support the time required to make this my primary project. The remainder would support AI model and API access, software and computing, acquisition and preparation of source materials, comparative testing, evaluation documentation, and public research materials.
At the $25,000 minimum, I could conduct a reduced research program focused on a smaller number of source domains and comparative tests.
At the $50,000 goal, I could sustain the work for approximately twelve months, test across a broader range of models and source materials, improve evaluation methods, document failures more systematically, and prepare reusable public demonstrations and research materials.
The funding is for research and evaluation rather than building a consulting practice or sales operation.
I am Greg R. Welch, an independent publisher and information designer with more than three decades of experience designing, organizing, and publishing information across print and digital systems.
I came to this research through publishing rather than machine-learning engineering. That background led me to investigate the information outside the model—the source from which the model is expected to reason.
I have developed the AI Source Analysis Framework independently for approximately five months and have funded the work myself. During that period I have built a working implementation, produced structured source materials, developed reasoning controls, and repeatedly revised the architecture in response to comparative testing and failure.
Failure has materially changed the system. An early single analytical approach proved inadequate and led to differentiated forms of analysis. Very detailed source representation became impractical for long documents, which led to broader structural resolution. Conversational testing exposed provenance and unsupported-relationship failures, which led to additional reasoning controls and retesting.
I am the sole project lead. There is currently no company, university, research laboratory, employee, or institutional sponsor behind the work.
The strongest possibility is that source representation matters less than I currently suspect.
More capable models, better prompting, ordinary retrieval techniques, or much simpler source structures may produce comparable results at substantially lower cost. The current architecture may also bundle too many variables to determine which components account for observed differences.
Explicit structure could create new problems. Incorrect interpretation could become more persistent once encoded. Added structure might unnecessarily constrain useful model reasoning. Human judgment during source preparation could introduce selection bias or relocate error upstream rather than eliminate it.
The existing evidence is also small and developmental. It does not demonstrate general reliability, causal effect, or superiority to alternatives.
Those are not merely risks to the project; they are questions I want the research to answer.
If the current Framework fails but the project establishes that certain source-side interventions matter, that would still be useful. If careful comparative testing shows that the source-side hypothesis contributes little beyond simpler techniques, documenting that result would also be valuable.
$0 in outside funding has been received for this project.
I have funded the development independently to date.
I currently have pending applications or inquiries with the Alfred P. Sloan Foundation Books Program, Emergent Ventures, Lightcone Commons, and the Effective Altruism Funds Transformative AI Research Grants program. No funding decision has been received from any of them.
If more than one application succeeds, I will disclose the awards and coordinate project scope and budgets so that the same work is not funded twice.
*This proposal describes the research proposition, current implementation at a high level, and evaluation plan; it intentionally does not publish all internal production methods used to create the current structured source materials.