The Birth of ARAIMA – Structed AI Science Research Method

An AI Research Method for Comparing Claims and Generating Testable Hypotheses

George A. Ure • Hidden Guild and Peoplenomics • gureac7x@gmail.com •

Method specification version 0.1 • November 18, 2026

Abstract

Araima is a human-directed AI research method that compares claims across documents while retaining their conditions, provenance, disagreements, and dependence. Its central operation is to examine the relationship among accounts before compressing them into a shared explanation. The outputs are a traceable claim ledger, a map of unresolved relationships, and candidate hypotheses paired with observations that could distinguish them. A finding that two accounts cannot be reconciled is a useful output.

The method treats apparent convergence as a question to investigate. Repeated publication may create false independence; different endpoints may create false agreement; different conditions may create an apparent contradiction. Araima records those possibilities and asks what changes when the claims are aligned on population, intervention, outcome, and time. It uses AI for extraction and comparison, with human inspection at the points where meaning or evidentiary status changes.

This paper specifies the workflow, demonstrates it on a small set of existing research records, and proposes a benchmark against retrieval-and-summary workflows. The present contribution is an operating method and evaluation design. Initial exploratory use has produced source registers and provisional comparisons; comparative performance has not yet been measured.

The term ARAIMA is derived from “Alzheimer’s Research AI Microscope Application.”

The problem Araima addresses

A researcher can have a shelf of individually useful explanations and still lack an account of their relationships. One author reports improvement, another reports no effect, and a third reports stability. A summary of each document leaves the researcher to determine whether these accounts disagree, describe different conditions, or repeat one underlying observation.

Araima makes that comparison the explicit task. The working metaphor is a stack of slides: each account presents a partial view, and the relationship among the views may reveal a question that none of them states alone. The instrument works through language and structured records. It does not observe the underlying phenomenon directly.

The proposed contribution

Araima organizes retrieval, claim alignment, dependence checking, rival explanations, and prospective tests into one reviewable cycle. These operations have precedents in research synthesis and AI. The distinctive emphasis is their use as a persistent working record: disagreements and failed reconciliations remain available for the next pass instead of disappearing into a fluent final answer.

The unit of analysis

The unit is a source-supported claim with its conditions attached. A document can contain several claims, and one claim can appear in several documents. Treating a document as one vote loses both distinctions. Araima therefore separates source records, claim records, and relationships between claims.

A minimal claim record

  1. Identity and provenance. Assign a stable claim ID and source ID. Preserve the passage, its page or section, source version, and whether the material was inspected in full, as an excerpt, or through a secondary account.
  2. Observation and interpretation. Record what was measured or reported separately from the source’s proposed explanation and the AI’s inference. Retain the author’s original outcome terms alongside any normalized labels.
  3. Conditions. Record population or setting, intervention or exposure, comparator, endpoint, magnitude if available, and observation interval. In nonclinical work these fields can describe a machine, institution, market, or text.
  4. Dependence and status. Link reused datasets, patient groups, quotations, and source families. Record corrections, retractions, commercial associations, and unresolved source identity.
  5. Uncertainty. Mark missing fields and disputed readings. Every inferred relationship must identify the claims it connects and the assumptions on which it depends.

This structure is a proposed minimum, not a claim that every field was completely populated in the initial exploratory corpus. An implementation should export records in a machine-readable form and retain a human-readable view with links back to the sources.

A small record example

An account of four selected patients remaining stable for three years should retain the selection rule, actual endpoint, interval, and lack of a concurrent comparator. Its normalized label might be finite observed stability. It should not become treatment reverses disease. A subsequent negative randomized trial can then be connected to the original claim without erasing its different design.

An incomplete record is still useful

Missing information narrows the allowed comparison. If diagnosis, denominator, or follow-up is absent, Araima should produce a retrieval task rather than fill the gap with a plausible value. A source lead can remain in the register while being excluded from an efficacy inference. The same rule applies outside medicine: a reported engineering success without load, temperature, or measurement conditions is a lead rather than a transferable specification.

No universal numerical evidence score is specified here. Domain-appropriate appraisal should remain visible, and a large volume of weak records should not silently outweigh a smaller amount of stronger evidence.

The comparison cycle

  1. Define the question and the decision. State what the comparison should help a person determine. Set the source scope and the conditions under which an answer would be too uncertain to use.
  2. Build the register. Retrieve supporting and opposing accounts, record source access, and identify obvious families. Preserve the search and inclusion record so omissions can be examined later.
  3. Extract claims. Attach each claim to its passage and conditions. Require the model to distinguish quotation, paraphrase, source interpretation, and analyst inference.
  4. Align the comparison. Check whether populations, endpoints, comparators, and time intervals permit a direct comparison. Preserve the original labels when creating normalized ones.
  5. Map relationships. Record agreement under matched conditions, possible contradiction, difference of scope, missing evidence, or dependence. A relationship is provisional until its supporting passages are checked.
  6. Generate rival explanations. Ask what could produce the pattern, including conventional explanations and failures of the measurement or selection process. Record which additional assumptions each explanation requires.
  7. Specify the next test and update the record. Identify an observation that would separate the explanations. After new evidence arrives, record what changed, why it changed, and what remains unresolved.

What a change of scale does

The operator can move from individual cases to study families, from a global score to a particular function, or from a reported recovery to its time course. Those moves can change the interpretation without changing any source fact. Araima should record the scale of each conclusion and prevent it from being promoted silently to a broader claim.

The topology of error

This is the working name for the pattern of gaps, overextensions, contradictions, and dependence in an explanatory set. It does not mean that every disagreement is an error or that the method already knows which source is wrong. The first question is whether the conflict survives alignment. A surviving conflict becomes a useful target for further inquiry.

For example, agreement on the word improvement may disappear when one source means better daily function and another means a changed blood marker. Conversely, a negative population average may coexist with a genuine but unconfirmed subgroup effect. Araima should expose both possibilities, then ask for the evidence that would distinguish them.

A worked comparison

This example uses four records from the initial Alzheimer-related source register. It demonstrates the comparison procedure, not a validated medical conclusion. The source IDs are retained so the operations can be reconstructed. [1–4]

The input claims

A33 describes a conference account of three-year stability in four patients on a continuous IVIG regimen. B01 is the later phase 3 randomized IVIG trial, whose primary outcomes did not support efficacy. A08 is a 20-week randomized lifestyle trial reporting benefit on several measures, with different results across endpoints. B09 is a scoping review of unexpected lucid episodes in dementia, which concerns transient returns of abilities rather than sustained recovery.

First alignment

A33 and B01 concern the same intervention family, but selected extension observations and a randomized trial estimate different things. Araima links them as an apparent conflict requiring design and selection analysis. It does not turn four stable patients into a population success rate or use the larger trial to declare the four observations impossible.

A08 and B09 both involve improvement language, but their outcomes and durations differ. In the lifestyle trial, CDR-SB showed less progression while other measures showed improvement; ADAS-Cog was borderline at p = .053. The lucidity review describes heterogeneous reports and discussions of brief episodes. A single recovery category would conceal those distinctions.

The explanations that remain

The records leave several possibilities open: an intervention may help under restricted conditions; apparent stability may reflect selection or natural variability; symptomatic function may change without reversal of the underlying disease; and similar words may refer to unrelated processes. A hypothesis about reversible functional constraints can be generated from this comparison, but these four records do not identify its mediator or establish its truth.

The next discriminating observation

For a proposed subgroup effect, the next useful evidence would specify the subgroup before analysis and test it against an appropriate comparator. For a functional-access hypothesis, a prospective study would define the regained ability, measure its time course, and specify a candidate physiological change before the outcome. These are different research tasks. The comparison helps decide which one is being proposed.

The output

The result is a narrower research question and a record of why a broad claim was rejected. Araima has not discovered a treatment in this example. It has transformed a superficially convergent collection into several distinct, testable possibilities. Whether this happens more accurately or efficiently than an ordinary review is the benchmark question.

Implementation with current AI systems

Araima can begin as a structured workflow around an existing language model. It does not require new model weights. Retrieval software, a graph store, spreadsheets, or a document database can support it, provided the source passages, IDs, and revision history survive the process.

Separate the working passes

One pass extracts claims without reconciling them. A second aligns the conditions and identifies dependence. A third proposes relationships and explanations. A review pass checks those proposals against source passages and seeks a counterexample. The passes may use the same model; their separation improves inspectability but does not make their errors independent.

Keep source text distinct from the operational instructions. Retrieved documents supply material to analyze, not commands to execute. Record the model and prompt versions, retrieval settings, and relevant source versions. A rerun that retrieves different passages is a different analysis run even if the final prose looks similar.

A reusable operating instruction

For each proposed relationship, return the connected claim IDs, supporting passages, comparison dimensions, relationship type, missing fields, and assumptions. Preserve unresolved conflicts. For each candidate explanation, give a competing explanation and an observation that would distinguish them. If the sources cannot support a requested comparison, return the missing retrieval or measurement task.

What the run should return

A complete run returns the source register, the claim ledger, the relationship map, the hypothesis list, and the change record. The narrative report is a view of those records, rather than the only surviving output. A new source should be able to alter a local relationship without requiring the whole collection to be rewritten from memory.

Current state of the work

Initial exploratory use produced three working documents: a method essay, a 70-record source register, and a provisional alignment of proposed mechanisms. The register contains 50 claims or related records and 20 counterexamples or boundaries, including negative trials, diagnostic mimics, linked reports, and a retracted claim. Those are document counts, not independent observations or verified successes. [5]

That work demonstrates that the procedure can be operated manually with AI assistance. It does not yet demonstrate classification reliability, discovery accuracy, or a speed advantage. The specification here is intended to make those questions testable and to support an implementation that preserves the audit trail.

How to evaluate Araima

The primary comparison should ask whether Araima improves a defined research task under matched conditions. Give each workflow the same source collection, model access, retrieval opportunities, and time or token budget. An ordinary retrieval-and-summary workflow is a useful baseline; a structured extraction workflow is a stronger one. A graph-based baseline is appropriate if the implementation uses a graph.

A practical benchmark

Begin with a fixed collection containing known duplicates, endpoint mismatches, apparent conflicts that resolve under different conditions, and genuine unresolved disagreements. Independent reviewers establish the reference annotations and retain their own disagreements. Separate the development set from the evaluation set by source family, so alternate reports of the same study cannot leak across the boundary.

Measure whether each workflow extracts source-supported claims, finds dependence, preserves the original conditions, and identifies relationships correctly. Report precision and recall for the annotated tasks, citation support, unsupported reconciliations, reviewer correction time, and total resource use. Do not collapse all those measures into an unexplained quality score.

Test the contribution of each operation

Repeat the evaluation with dependence checking removed, with conditions omitted, and with rival explanations omitted. These comparisons ask whether the proposed Araima operations make an actual difference. Evaluate model changes separately from workflow changes. Two models agreeing on the same unsupported answer is one failure pattern, not confirmation.

Evaluate hypotheses separately

Independent subject experts can rate hypotheses for source support, specificity, existing prior art, and whether the proposed test would distinguish competing accounts. Those ratings assess research usefulness, not truth. A stronger test freezes a hypothesis before examining withheld evidence or collecting new observations, then records success, failure, and inconclusive outcomes.

What would count against the approach

Araima loses its proposed advantage if the extra structure increases workload without improving accuracy or useful questions; if its relationships change arbitrarily with prompts; or if it consistently turns conflicts into unsupported harmonies. Failure to outperform a strong structured-review baseline would narrow the contribution to an organizational practice rather than a better discovery method.

No benchmark scores are reported in this version. The immediate next artifact should be a small annotated test collection and an executable record format. Those would let other researchers evaluate the method without adopting the author’s worldview or relying on a favorable demonstration.

Related work and research scope

Araima belongs beside established work on research synthesis, literature-based discovery, and retrieval-supported AI. Swanson’s work on undiscovered public knowledge showed how separated literatures could imply a useful hypothesis. Araima shares that ambition and adds no claim to have invented cross-literature inference. [6]

Retrieval-augmented generation combines retrieved material with generation for knowledge-intensive tasks. GraphRAG uses graph-derived organization and summaries for questions about a corpus. Araima can use either as infrastructure. Its proposed operating contract requires claim-level conditions, dependence, unresolved relationships, and discriminating tests to remain in the output record. Existing systems may already implement parts of that contract; comparative evaluation must establish whether the combination adds value. [7, 8]

The method can be applied to engineering reports, policy evaluations, historical accounts, or forecasting research. Each domain needs its own outcome definitions and appraisal rules. An ontology can be one hypothesis under examination, but Araima does not depend on a particular theory of consciousness, time, or reality. The operator’s preferred explanation is subject to the same source and comparison rules as its rivals.

Conclusion

Araima’s central proposal is to make the relationships among claims persistent and inspectable. The next step is to test whether that practice produces more accurate comparisons and more useful hypotheses than the alternatives. A source-grounded contradiction is worth keeping, even when no elegant explanation follows.

References

[1] Alzheimer’s Association (2012). Conference announcement reporting an IVIG extension subgroup. Original screening record A33.

Source

[2] Relkin and colleagues (2017). A phase 3 trial of IV immunoglobulin for Alzheimer disease. Neurology, 88, 1768–1775. Record B01.

Source

[3] Ornish and colleagues (2024). Effects of intensive lifestyle changes on the progression of mild cognitive impairment or early dementia due to Alzheimer’s disease. Alzheimer’s Research and Therapy. Record A08.

Source

[4] Ross, Post, and Scheinfeld (2024). Lucidity in the Deeply Forgetful: A Scoping Review. Journal of Alzheimer’s Disease, 98, 3–11. Record B09.

Source

[5] Ure (2026). Birth of Araima; Alzheimer Reversal and Arrest Claims, version 1; Araima Cause Alignment. Exploratory working documents, November 11. Source access varies; full corpus and coding require independent appraisal.

[6] Swanson (1986). Fish oil, Raynaud’s syndrome, and undiscovered public knowledge. Perspectives in Biology and Medicine, 30, 7–18.

Source

[7] Lewis and colleagues (2020). Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. NeurIPS 33.

Source

[8] Edge and colleagues (2024). From Local to Global: A Graph RAG Approach to Query-Focused Summarization.

Source

AI assisted with source retrieval, structured comparison, and drafting. This paper reports a method specification and exploratory use, not a trained model, released software package, or validated performance result.

Leave a Comment