Syntropy Nexus AI Foundry · Research update · 14 October 2026
Research purpose
PATHFINDER investigates how people and authorized AI systems can locate, interpret and assemble reliable evidence from large, evolving collections of documents. The central problem extends beyond finding relevant text: a useful result must preserve the identity of its source, distinguish current material from superseded revisions, respect access boundaries, and retain enough context to support the conclusion being drawn.
The intended outcome is a modular Corpus Navigation and Management System (CNMS). The study uses DAEDALUS APEX CORPUS (DAC; SNAF.DAC) as a bounded research testbed, while treating portability to other corpora and platforms as a hypothesis requiring separate validation.
Research approach
The programme examines document extraction, lexical retrieval, metadata-based source routing, context assembly, revision and authority interpretation, and corpus lifecycle management. Existing inventories, indexes, maps and pointers remain part of the comparison baseline. Semantic retrieval, graph-based methods and model-dependent routing are candidate modules; the current evidence does not establish their comparative effectiveness on this corpus.
Current observations
Bounded experiments indicate that source selection and context packing should be evaluated separately. In the 35-question development panel, all-source lexical search retained the complete expected evidence on 14 questions; a development-informed two-source routing configuration retained it on 19. These results concern evidence retention, not generated-answer accuracy.
On a separate 26-case reserved-reuse panel, the corresponding counts were 11 and 13. That panel had already been visible to the builder and was not a pristine blind holdout. On a restricted source-disjoint panel, the counts were 13 and 16 across 18 evaluated cases; two of the 20 planned cases were blocked by source changes. These panels must remain separate and do not support a universal performance claim.
Context expansion sometimes recovered useful evidence and sometimes displaced evidence that previously fit within the context budget. The current findings therefore support conditional, testable module choices rather than a single universally preferred architecture.
Current research package
The public-review edition contains the consolidated conclusions Draft V7, research journal V10, vocabulary workbook V4, evaluation vocabulary addendum V1, and routing and diversification supplement V1. A reading guide and integrity manifest identify the selected revisions and their relationship to the original files.
Download the PATHFINDER research package (ZIP)
This is a public-review derivative of the current five-file package. Internal locations, session identifiers and unrelated customer references are withheld. The workbook’s private benchmark-case worksheet is replaced with an exclusion notice. The underlying source corpus and raw case-level evidence are not included, so this edition does not provide complete external reproducibility.
Research status and review
PATHFINDER remains an incomplete research programme, paused at its current research checkpoint. The reported checks are builder-side evidence; independent scientific review, production qualification and the original programme completion criteria remain open. A fair end-to-end comparison with complete existing navigation and human workflows is unresolved.
Readers are invited to assess the problem formulation, experimental controls, failure analysis and limits of inference. The package is offered for critical review, not as a completed study or a production-readiness claim.