Galaxy accepts both single-node and list model responses, injects kernel and anchor seeds and uses provenance flags to keep seed-only candidates out of the internal acceptance path. Reviewer rejection, candidate–ledger binding and call budgets become runtime constraints.
Philosophy
Completing a workflow does not establish that a model performed the research. The system needs to distinguish model-supplied content from its own filling and enforce rejection decisions and budgets.
Architecture & control flow
- 01
Consume single-node, nodes and children response contracts.
- 02
Use kernel seeds for structural exploration while recording model and seed provenance.
- 03
Seed-only candidates skip review; reviewer rejection or ledger misbinding blocks acceptance.
- 04
Synchronize call statistics and persist review budgets and operator state.
Architecture outline derived from this version’s control flow.
What this version changes
Compared with Jupiter, focuses on provenance, output contracts, rejection and budget constraints rather than merely passing an idealized simulated path.
Implementation & evidence scope
Default structural exploration can still stall at mechanism nodes; Cosmos later improves candidate-stage progress. Changelog totals conflict between 32 and 34 tests, so they are not used here as capability metrics.
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/V4_GALAXY_CHANGELOG.md· 22–38novelty-idea-generator/V4_GALAXY_CHANGELOG.md· 42–60novelty-idea-generator/harness/run_agent.py· 383–445novelty-idea-generator/V4_GALAXY_CHANGELOG.md· 100–138novelty-idea-generator/V4_GALAXY_CHANGELOG.md· 207–212