Compounding Relational Semantic Engine (CRSE) / Research program
When accumulated reasoning loses its foundations
If an AI research system builds on earlier conclusions, what should happen when evidence supporting one of them is withdrawn?
Some downstream conclusions may need revision; others may remain independently justified. A stored conclusion may also be valid in one context without being applicable in another.
Open experimental question: Can a system reliably distinguish these cases when semantic dependencies must be identified rather than supplied—and does it improve on established alternatives?
Current evidence: V0–V5 report bounded experimental outcomes. In V4, CRSE matched conventional explicit dependency tracking; no advantage was established. V5 distinguished unresolved evidential status from out-of-scope application in one native-kernel test. Automatic semantic-dependency discovery and independent replication remain unestablished.
What has been observed?
The short account below is not a claim of general semantic understanding, novelty or production readiness. Follow each link for method, observed output, replay and limitations.
Hypothesis and competing approaches
Explicitly representing evidence, dependencies, authority and applicability conditions may improve the reliability of cumulative machine reasoning. Established dependency graphs, truth-maintenance systems, provenance tracking and rule engines are relevant alternatives—not straw-man comparisons. The present evidence does not establish a unique CRSE advantage.
How could this hypothesis fail?
A simpler method performing as well on matched tasks, unreliable extraction of relationships from language, incorrect affected-set detection, inability to preserve independent conclusions, or computational cost outweighing any improvement would count against the approach. These outcomes should be reported even if negative.
Evidence available for inspection
Each experiment is documented with its conditions, observed outcomes, limitations and an explanatory replay. V0–V3 historical result summaries are available; original console excerpts are withheld pending source-byte reconciliation. V4 is represented by a structured comparison, not an original terminal transcript. V5 is described using its verified result summary; selected original console lines are withheld pending publication clearance. The graphics are teaching reconstructions, not live executions. Public reproducibility materials are not yet available.
Inspect the experimental record → · New to the research question? Read the visual introduction →
Correspondence
Research inquiries and adversarial evaluation
We welcome technically substantive criticism, proposals for stronger tests, research collaboration, and exploratory funding or pilot inquiries. No general-purpose production service is being offered.
If your browser does not open a mail program, copy the address and use your preferred email service.