An independent research practice focused on deployed AI behavior, evidence integrity, and evaluation methods.
Menu
About SENCI
Independent by structure. Adversarial by method.
Independent research / 2026
SENCI Group is the independent research practice of Matthew L. Yates, developing evidence-disciplined methods for examining deployed AI behavior.
Mission
Preserve what happened. Reconstruct the behavior. Make the claim answer to the record.
SENCI Group develops and adapts forensic methods for examining how deployed AI systems behave under real-world and controlled conditions. The work compares system claims against source evidence and documents omission, distortion, reliability failure, and change under correction.
Observation is separated from functional classification, causal explanation, and subjective or ontological speculation. That boundary is not timidity. It is what makes a sharp claim difficult to dismiss.
Structure
What SENCI is—and what it does not pretend to be.
A home for publications, methods development, case analysis, and prospective controlled studies.
A university lab, staffed institute, corporation, accredited forensic discipline, or formal standards body.
That a DOI proves peer review, a hash proves truth, selected cases prove prevalence, or output language proves motive.
Researcher
Matthew L. Yates
Independent researcher · SENCI Group
Matthew’s work develops transcript-grounded and horizon-sensitive methods for reconstructing system behavior across conversations, source disputes, tool use, agent trajectories, and cross-system incidents.
The program emphasizes evidence preservation, source–output fidelity, sequence-sensitive analysis, human-controlled review, and explicit limits on causal and ontological claims. Six open preprints form the current public record; a matched Version 2 methods paper and long-horizon field paper are forthcoming.
Research stance
Neither institutional deference nor metaphysical theater.
Direct observation matters
Provider documentation supplies context. It does not erase behavior preserved in the product surface.
Alternatives must be tested
A compelling pattern still has to survive ordinary explanations, counterexamples, and symmetric standards.
Strong language needs clean claims
Words such as deception, intent, memory, and preservation must be operationalized at the claim level.
Contradiction stays in the file
Failed replications, revisions, and reviewer disagreement are evidence—not debris to sweep from the narrative.
Public commitments
Credibility is built out of inspectable habits.
Status labels remain literal and date-bound. Preprint, submitted, under review, accepted, and published are not synonyms.
Current repository links and latest known versions are favored over stale indexes, aggregators, or inherited metadata.
Public derivatives are separated from restricted originals. Private records are not implied to be open datasets.
Material errors should be corrected visibly, versioned, and preserved rather than silently rewritten into a cleaner past.
Contact
Research correspondence
For methodological critique, research discussion, evidence questions, or peer-review correspondence.