01
Advance exploration toward testable candidates
Cosmos addresses stage starvation: cheap priorities can repeatedly favor early mechanisms while experiments and attacks struggle to progress. Stage buckets prioritize nodes closer to candidates, then balance unfinished actions, coverage and exploration within a stage.
02
Choose the stage before ranking branches
The scheduler first chooses the highest-priority nonempty stage bucket, then separates visited nodes with unfinished actions, unvisited nodes and fallback objects. Attacks, repairs, experiments and mappings no longer compete with all initial questions through one exploration score. This improves late-stage opportunity without establishing scientific value.
03
From a response to core-field coverage
Galaxy can mark provenance after receiving a valid node. Cosmos records model-filled fields and seed-filled remainder. IdeaCandidate requires at least four of five core fields for its internal verified flag; other nodes also require actual model fields. This is a provenance-coverage gate, not verification of scholarly facts.
04
Reject foreign or stale ledger identities
Ledger binding treats UUID-like identities strictly. A foreign or stale candidate identity cannot be silently overwritten, while obvious placeholders can be bound to the current object. This distinguishes repairable formatting gaps from evidence misbinding so criticism addresses the current candidate.
05
Refute again after repair
A repair can change the mechanism, claim or prior-work difference, invalidating the old ledger. RepairPatch goes through renewed refutation and review with its relation to the original retained. Low-confidence mathematics–physics mappings stay provisional when criticism and responses are insufficient for the internal confirmed_novel label.
06
Inputs and auditable handoff
Inputs require substantive role fields and identity-matched ledgers alongside envelope and budget. Outputs add stage coverage, field provenance, seed gaps, UUID binding and repair-revalidation records. Researchers can trace mechanism-to-experiment and criticism-to-revision paths and inspect which available materials supported promotion.
07
Improve provenance granularity and progress together
Compared with Galaxy, Cosmos improves both candidate-stage progress and coarse provenance, with stronger ledger identity and repair lifecycle. Finer provenance does not mean stable scheduling: Nova later identifies random mathematics–physics branches diverting the main-path budget. Cosmos is a stage-scheduling advance rather than a final stability claim.
08
Treat a revision as a new evidence object
A revision can change its prior-work relationship. Changing a mechanism or moving from method to measurement requires asking again which prior is strongest and which claim the evidence supports. Cosmos retains object identity and ancestry and checks a repair through a new ledger instead of copying the old review result.
Stage buckets give such objects opportunities to progress. Inspect the stage, unfinished operations and field provenance: a later stage does not imply greater value, and model-filled fields do not imply correctness. Keeping review readiness distinct from academic validation avoids reporting engineering progress as a solved scientific problem.
09
Internal confirmation and scientific conclusions
Candidate paths are primarily checked with simulated calls, and the mathematics–physics bridge lacks complete live retrieval. confirmed_novel is a runtime state rather than independent proof of novelty. Scheduler, provenance and identity source establish engineering progress while real-model research, retrieval coverage and experiments require separate validation.
Reading example · Explaining the workflow
A repair cannot borrow its old ledger
Suppose prior work prompts a core-mechanism revision, creating a RepairPatch.
- 01
Retain the repair’s relation to the original and inspect core-field provenance again.
- 02
Generate and bind renewed criticism to the repair, rejecting material for another UUID.
- 03
Save new criticism and review to the current object before deciding to continue or stop.
A repair record tracing the revision, evidence identity and renewed review.
A conceptual reading example. No generation, retrieval, review or experiment was run, and no performance result is reported.
Changes in this version
Candidate provenance
- Earlier approach
A returned node can receive a coarse model-provenance flag.
- This version’s approach
Check model coverage across five core fields and retain seed gaps.
Base internal promotion on supplied research fields.
Repair evidence
- Earlier approach
Revisions can carry an old candidate’s ledger.
- This version’s approach
Refute the repair again and check identity.
Revisit prior-work differences after claims change.