Three concrete starting points
Reproduce OTB21
Run the public package from its saved memory states. Record software versions, command, seed/state provenance, output and any discrepancy. Distinguish rerunning supplied states from acquiring new independent histories.
Compare at the same time budget
Choose fixed waiting times without using the evaluation outcomes. Compare retention, adaptation and total elapsed steps. Include failed conditions and preparation costs.
Separate conflict from support
Starting from the OTB22 question, define when contradictory owners should cause abstention. Specify both a detector and negative controls. Treat this as a proposed experiment, not an existing capability.
Send an actionable observation
Include the lesson or study URL, exact claim, source or command, expected result, observed result and remaining uncertainty. A failed test is useful evidence; it need not be a polished success story.
Scope and reuse
The studies use designed numerical tasks and external components. RNN superiority, general reasoning, autonomous meaning acquisition, self-learning and measured hardware efficiency are not established. The research repository has not yet selected an open-source license; public visibility is not a blanket reuse permission.
Textbook revision 1 · Evidence snapshot: 05 Nov 2026 · Changes & corrections ↗