Publications / Research index
Technical records, experiments, research notes, and methodological documentation from Velorin Intelligence.
Document types, versions, and limitations are stated separately. Evidence cutoffs and frozen-release dates retain their original meaning.
Velorin Intelligence
An executable omitted-lineage case, a frozen synthetic policy comparison, and a limited external negative result. The note preserves methods, quantitative outcomes, and mandatory nonclaims.
Preliminary technical-note draft. Not an active academic submission or evidence of general external effectiveness.
Velorin Intelligence
v1.0.0Matched Kev-0.8B / Kev-4B experiments with instruction, option-schema, and context matrices; source-authority, calibration, and cross-domain controls; and checksummed result artifacts.
Velorin Intelligence
A synthesis of Constraint Sentinel, three assurance experiments, and new research on the human cost of AI agent oversight.
Velorin Intelligence
No established superiority of the tested neural adapter over hybrid retrieval; evidence supports source-backed versioning and explicit correction precedence.
Velorin Intelligence
A successful result does not always tell us what capability produced it. Meta’s Muse Wonder evaluation offers a useful example of why increasingly agentic systems may need both outcome and process evaluation.
Velorin Intelligence
Long-running AI agents need to distinguish between productive persistence and situations in which continued pursuit is no longer appropriate.
Velorin Intelligence
Persistent AI memory can retain context across sessions. This raises a control question: what happens when a system retains what it learns from interacting with its own safeguards?
Velorin Intelligence
Experiments with an open-source Jev-inspired decision model distinguish constraints on model outputs from constraints on the evidence influencing its judgment.
Velorin Intelligence
Anthropic’s internal agent infrastructure separates persistent agent identity from the underlying model, raising questions about continuity, provenance, and governance.
Velorin Intelligence
Research principles, experimental design, evidence preservation, adversarial review, claim boundaries, reproducibility, and governance. Project-specific records determine which practices were applied.
Velorin Intelligence
Baseline and reduced-filter conditions each recorded 11 of 44 unauthorized requested submissions. The pilot tested a reduced exact-identifier citation-coverage filter, not full VCT V2.
Indexed within the preserved VCT technical note; not a separate publication or a general result about provenance closure.