Research Methods
RESEARCH METHODS
SRI constructs a research methodology that utilizes the speed and scope of AI while preserving source attribution, claim boundaries, human accountability, and falsifiability.
Core Cycle
Question → Evidence → Model → Computation → Prototype → Observation → Critique → Revision
It is managed not as a linear progression toward a conclusion, but as a cycle that can return to the question and model from any stage.
Human–AI Research
AI handles the exploration, structuring, hypothesis generation, counterexample search, computational assistance, and representation conversion of literature candidates. Humans bear responsibility for the research objective, ethics, critical assumptions, external publication, and social implications. AI output should not be treated as source attribution or proof.
Evidence & Claim Levels
- Observed Fact：Matters directly verifiable from sources or data
- Source Claim：Statements asserted by external materials
- SRI Concept：Concepts and frameworks defined by SRI
- Working Hypothesis：Candidate explanations under verification
- Computational Evidence：Calculation results with limited conditions and scope
- Countermodel / Limitation：Counterexamples, competing theories, or limits of application
- Revision：Changes based on evidence and critique
Computational Verification
We record the code, input, model, environment, random numbers, scope of application, and failure cases to the extent possible. We do not extend verification of a finite range into infinite general proofs; instead, we establish reproducibility and stopping conditions.
Research Provenance
Outputs are linked to source attribution, version, creation date, conversion history, models/tools used, and review status. Confidential information, personal data, and credentials are not included in public Provenance.
Questions and Evidence
Instead of deciding on a conclusion or product first and then gathering information, we define the scope of the question and the conditions under which the answer changes. Evidence retains the source URL, acquisition time, original text, version, author, publishing entity, and extraction method. Search results or AI summaries alone are not treated as primary sources.
Models and Computation
A model is a representation of selected parts of reality. Variables, assumptions, boundary conditions, and exclusions are explicitly stated, and multiple models are not integrated prematurely. In computation, we record reproducible inputs, environments, errors, and failure cases, and stop claims that cannot be derived from the results at the Validation Gate.
Prototypes and Observation
A prototype is not designed to prove the correctness of a theory, but rather an instrument for observing constraints that emerge during implementation. We collect real-world feedback, including user experience, operational load, failures, safety concerns, and legal conditions, and return this to Theory and System.
Critique and Revision
Internal reviews seek the strongest counter-explanations, missing sources, irreproducible sections, and conflicts of interest. When subjected to external critique, we do not silently rewrite the original output; instead, we distinguish between corrections, retractions, and new editions. The ability to revise is viewed not as a weakness but as a requirement of research.
Additional Guidelines for Using AI
- Record the model used, prompt purpose, and primary inputs in the work history.
- Verify plausible citations, numbers, and proper nouns against original sources.
- Do not complete verification solely based on self-evaluation by the same model.
- Do not pass confidential or personal information to models unnecessarily.
- Humans must confirm claim boundaries, rights, and safety before publication.
Advancing Claim Levels
Moving from Working Hypothesis to Computational Evidence requires reproducible conditions. Advancing to Proof-grade requires not just finite computation, but the closure of logical dependencies and the explicit statement of unproven assumptions. Externally Verified refers to a state where an independent third party has examined both the methodology and the results, rather than internal reconfirmation by SRI.
Cross-Disciplinary Considerations
Even when the same terms are used in mathematics, physics, AI, and social systems, their definitions and standards of evidence differ. When transferring concepts like “observation,” “choice,” “energy,” or “quantum” between fields, we distinguish between metaphor, formal correspondence, and empirical identity. The existence of a common diagram does not lead to the conclusion that there is an identical causal mechanism.
Lighthouse Method
The Lighthouse Method is a research and decision-making approach that does not quickly close onto a single conclusion within a complex information environment, but rather maintains multiple Map, evidence, counterarguments, and undetermined areas in parallel. By neither conveniently rewriting past facts nor confining the future to a single optimal solution, it leaves room for humans to re-observe, re-select, and assume responsibility.
SRI applies this method to research, AI governance, and organizational decision-making. The public Charter and current implementation can be viewed at the independent site SEERAVERSE Lighthouse ↗.
Pre-Publication Review
Published materials are checked for factual accuracy, source attribution, logic, rights, privacy, safety, and potential reader misunderstanding. Research articles examine Claim Boundaries and Limitations; practical reports check the target date and verification methods; and translations verify the retention of proper nouns, numbers, and links.
Revising the Methodology Itself
The Method is not a fixed set of rules. It is updated in response to incidents, reproducibility failures, external critiques, and technological changes. When making changes, we record the effective start date, affected outputs, and the necessity for re-verification, ensuring that current standards are not retroactively applied to past results.