Experiment explorer
Inspect how a method actually searches and selects a version. The archive covers 12 methods × 2 target harnesses on SearchQA, HotpotQA, Terminal-Bench, and AppWorld. Each method has its own selection view; a single dev curve would hide important differences between these runs.
Available experiments
| Method | SearchQA | HotpotQA | Selection view |
|---|---|---|---|
| SkillOpt | Explore | Explore | Fast / slow skill updates |
| OPRO | Explore | Explore | Optimization history |
| EvoPrompt | Explore | Explore | Population arena |
| GEPA | Explore | Explore | Coverage frontier |
| ACE | Explore | Explore | Playbook ledger |
| AHE | Explore | Explore | Repair audit |
| ADAS | Explore | Explore | Design archive |
| Meta-Harness | Explore | Explore | Experiment notebook |
| DarwinX | Explore | Explore | Branch admission |
| ModularRSI | Explore | Explore | Branches and integration |
| RRSI | Explore | Explore | Round tournament |
| Evo-Harness | Explore | Explore | Skill production |
Every experiment includes MiniAgent and Claude Code. The pages preserve each historical run's recorded policy; they do not substitute the latest algorithm implementation or reselect historical candidates.
Open View diff on a candidate, branch, or checkpoint to inspect its file changes without leaving the selection view. The comparison includes added/removed lines, before/after text, and side-by-side views; you can also choose another recorded version as the comparison base.
The explorer does not display test results. Training scores, dev scores, per-task coverage, admission gates, and final selection are labelled separately. Runtime progress logs and algorithm event records are distinct from complete model conversation transcripts.
To add an experiment, see maintaining the explorer.
Terminal-Bench
| Experiment | Selection view |
|---|---|
| OPRO × Terminal-Bench | Optimization history |
| EvoPrompt × Terminal-Bench | Population arena |
| GEPA × Terminal-Bench | Coverage frontier |
| ACE × Terminal-Bench | Playbook ledger |
| AHE × Terminal-Bench | Repair audit |
| ADAS × Terminal-Bench | Design archive |
| Meta-Harness × Terminal-Bench | Experiment notebook |
| DarwinX × Terminal-Bench | Branch admission |
| ModularRSI × Terminal-Bench | Branches and integration |
| RRSI × Terminal-Bench | Round tournament |
| Evo-Harness × Terminal-Bench | Skill production |
| SkillOpt × Terminal-Bench | Fast / slow skill updates |
Open a method to switch between MiniAgent and Claude Code and inspect its recorded decisions, selected artifact, and file changes. Test results are omitted.
AppWorld
| Experiment | Selection view |
|---|---|
| OPRO × AppWorld | Optimization history |
| EvoPrompt × AppWorld | Population arena |
| GEPA × AppWorld | Coverage frontier |
| ACE × AppWorld | Playbook ledger |
| AHE × AppWorld | Repair audit |
| ADAS × AppWorld | Design archive |
| Meta-Harness × AppWorld | Experiment notebook |
| DarwinX × AppWorld | Branch admission |
| ModularRSI × AppWorld | Branches and integration |
| RRSI × AppWorld | Round tournament |
| Evo-Harness × AppWorld | Skill production |
| SkillOpt × AppWorld | Fast / slow skill updates |
Each page includes MiniAgent and Claude Code, recorded evolution steps, and preserved Snapshot file comparisons. Test results are omitted.