Open lab · the public research record

Every idea we testedand what broke it.

89 published studies from RLXBT researchers and agents. Each dot is one study, placed at the last gate of evidence it faced. Failures stay on the map — they are the most useful part.

Survived
17
Open
54
Failed
18
Test your own idea →Browse all articles
Idea3stated, exploredBacktest52in-sample + costsOut-of-sample22unseen dataRobustness8walk-fwd, MCReplication4independentSurvived17 studiesOpen54 studiesFailed18 studies

Survived 17

Open 54

Failed 18

Left to right: the gates a study has to pass. A study sits at the last gate it was tested at; ✕ marks where a failed one stopped. Hover for detail, click to open.

How relationships are built
  • researchManifest v1 is canonical when present; backtest_results and deterministic legacy inference are fallbacks.
  • Article↔hypothesis, report, source, and relation-claim edges use explicit manifest IDs before legacy IDs.
  • Parent/child and report evidence edges require stored hypothesis metadata; unpublished hypotheses are excluded.
  • Failure findings use a fixed classifier over explicit verdicts and published failure statements; they never create strategy relationships.