Ocean assigns tree expansion to specialized model roles for field splitting, kernel mapping, anchor transfer, mechanism construction, experiment design and refutation. Python wraps and records their responses; the mathematics–physics bridge becomes a staged model-call pipeline.
Philosophy
The execution scaffold should coordinate research rather than simulate research content with placeholder strings. Missing responses leave an inspectable partial process and unresolved work.
Architecture & control flow
- 01
Resolve the research scope; a missing field response leaves the run waiting at scope resolution.
- 02
Map operators to specialized roles and receive node content and evidence through the call bridge.
- 03
Preserve partial outputs through mathematical formalization, physical interpretation and refutation stages.
- 04
Write run state, rejection logs and call audits for continuation or diagnosis.
Architecture outline derived from this version’s control flow.
What this version changes
Compared with Hill, main search operators stop directly fabricating placeholder children and model calls enter the research-generation path.
Implementation & evidence scope
In this snapshot unresolved literature still falls back to a marked placeholder frontier, and low-confidence field resolution advances the phase; Stars later repairs these phase gates. A CLI interface exists, but scientific output and long-run stability were not validated in this review.
Code & bundled material
The introduction draws on bundled notes, changelogs and central code. Software tests, synthetic diagnostics and scientific effectiveness use different evidence standards.
Source references
novelty-idea-generator/harness/search/policy.py· 89–145novelty-idea-generator/V4_OCEAN_CHANGELOG.md· 33–56novelty-idea-generator/harness/run_agent.py· 174–214novelty-idea-generator/V4_OCEAN_CHANGELOG.md· 139–144