Skip to content

Algorithm inventory ​

The runner registry contains 13 methods. The resource and selection columns describe their FlashRSI implementations. Each guide documents configuration and any differences from upstream.

MethodEditable resourceFeedback and selectionWorker
OPROInstructionRank accumulated candidates by train score and return the best.Text proposal
GEPAInstructionReflect on trajectories and use per-task scores to guide candidate search.Text proposal
EvoPromptInstructionEvolve a population with GA or DE and select using dev scores.Text proposal
ACEPrompt / playbookBuild and revise a playbook from feedback and reflections.Text proposal
SkillOptSkillsCombine fast and slow updates and adopt strict dev gains.Text proposal
Evo-HarnessNative directory skillsCompile training failures into topic and general skills and retain valid updates.File editing
AHEHarness resourcesDiagnose train traces, evaluate each valid edit, and retain it for the next step.File editing
ADASHarness resourcesArchive measured designs and select the best dev candidate.File editing
DarwinXHarness resourcesScreen and confirm branches, then evaluate complementary recombinations.File editing
Meta-HarnessHarness resourcesUse historical sources, traces, and scores to propose edits and select on dev.File editing
ModularRSIDeclared modulesEvolve module branches and integrate them through review and execution gates.File editing
RRSIHarness resourcesDiagnose recurring failures and admit revisions using score and cost gates.File editing
ArborHarness resourcesExplore a hypothesis tree and synthesize findings from measured branches.File editing

Text proposals use worker.ask() and can come from an LLM Worker or an Agent Worker. File editing requires an Agent Worker with edit() support. Each Target must expose the resources and runtime capabilities required by the method. See algorithm configurations and harness capabilities.

Released under the Apache License 2.0.