Only what matches shipped code on main counts as default.
Research is where claims get earned.
This surface is the public research snapshot for Vraxion: what the project is trying to prove now, how claims are judged, and which recent turns still matter.
It is not the full archive. The deep chronology, reversals, and retained residue stay in the wiki's Research Process & Archive.
The public story is governed by one contract.
The same result can be interesting, reproducible, or canonical. Those are not the same thing.
Reproducible result that matters, but has not yet been promoted into the canonical path.
Active direction that may be promising, but should not be described as a default.
Three gates still separate momentum from stronger proof.
The live research problem is no longer "can the line do anything at all?" It is whether the strongest signals survive harder discipline.
Language robustness
Single strong runs do not count for much if the band collapses under reruns or multi-window evaluation.
Seed stability
The landscape is rugged enough that single-seed stories overstate certainty.
Context-dependent task learning
The architecture claim gets stronger only if task memory survives outside toy or fragile settings.
Sparse circuit evolution
Empty-start networks build targeted circuits with 40x fewer edges than prefilled. Smooth fitness + jackpot reached 24.6% peak.
Latest-first turns that still shape the current public line.
This is the short rail. The full chronology stays in the archive.
Hyperparameter exhaustion clarified that the remaining Rust problem looks more like a search-regime limit than a missing tuning knob.
Rust v5 beta became a real evolution substrate with snapshots, mutation API, persistence, CSR acceleration, and richer seed evidence.
Tentacle I/O, SDR input, compact channel results, and architecture compressions clarified the current public line.
Canon consolidation narrowed the public story around English, evidence discipline, and cleaner architecture boundaries.
The website frames the research. The wiki keeps the canon record.
Use the full archive for protocol, chronology, reversals, retained carry-over findings, and cross-surface context.