← Back to research
Research topic
Benchmark hub
A scoped index of benchmark families across language, memory, reasoning, speech, holdout, blind, cold-reload, candidate, baseline, oracle, and end-to-end tests.
How to read results
Metrics from different datasets, input types, gates, and phases are not combined into one Elena score. Each result retains its status and limitations.
Structured and oracle tests
Controlled meanings and oracle evidence isolate components but do not prove free-text or complete response behavior.
Blind, holdout, and reload tests
Blind and holdout sets address selection leakage within their recorded design. Reload tests address only the identity or behavior explicitly rerun.
Candidate comparisons
Baseline improvements remain candidate evidence until promotion is separately documented.