01
Ask who supplied the work
Galaxy examines Jupiter’s mixed generation path: did a model supply a complete candidate, or did system seeds fill the gaps? Model-provenance flags, response-contract compatibility and a seed-only gate make provenance a review prerequisite, while rejection decisions and call budgets gain direct runtime effect.
02
Accept varied outputs without erasing provenance
Response extraction accepts single nodes, nodes and children so useful content is not lost through different wrapping. Returned fields override seeds and audit records subagent_verified; absent node responses retain the seed path. This coarse provenance flag is later refined by Cosmos into core-field coverage.
03
The role of kernel seeds
Structural exploration can inject kernel or anchor seeds so mechanism nodes have required fields and can expand. These interpretable system materials support tree and progress checks, not proof that a researcher or model produced a new mechanism. Seed-only candidates remain an unfinished frontier outside review.
04
Bind candidates, ledgers and vetoes
A ledger must point to its candidate to avoid reusing evidence from another idea. Review verdicts such as reject, repair or borderline_repair cannot be overridden by a high mean score. Alongside numbers, role decisions become gate inputs that explain why a candidate cannot yet enter the accepted list.
05
Make the budget operational
Before review, the call budget and five-role panel determine how many candidates can be evaluated; overflow remains on the frontier. Phase-end synchronization records bridge statistics, while parent visits and tried operations enter tree events. Resumed budgets and scheduling therefore reflect more of the actual previous work.
06
Inputs, state and handoff artifacts
Inputs require a direction, envelope, role responses and an explicit budget. Outputs include provenance, ledger links, verdicts, call statistics and tree events in addition to candidate text. Readers can distinguish runnable system-filled structures from model-content candidates eligible for review, while still checking the literature and mechanism.
07
From lifecycle to provenance constraints
Relative to Jupiter, Galaxy addresses response wrapping, seed-only promotion, reviewer vetoes, ledger misbinding and budget synchronization. These sharpen execution semantics. Default exploration can still stall at mechanism nodes rather than reach experiments, attacks and candidates, making stage scheduling a central Cosmos task.
08
Read provenance together with budget
The review budget is more than a number printed in configuration. Five-role evaluation determines how many candidates the call allowance can cover. Provenance first filters seed-only and refuted objects so scarce review resources do not go to structures without research content. Unreviewed candidates can be distinguished by missing provenance, missing ledger or insufficient budget.
Phase-end bridge statistics, parent-node events and per-role verdict artifacts can be compared to see what a call produced, whether review resources were consumed and whether candidate state changed. They give resource use and decisions a shared context. Some model fields can still be seed-filled, motivating the later finer field-coverage audit.
09
Limits of provenance flags
Here subagent_verified mainly indicates that a model returned nodes; it does not establish that every key field came from the model or that content is true. Documentation conflicts between 32 and 34 tests, so those totals are not capability metrics. No simulated or real-model tests were run here; provenance checks and scientific evidence remain distinct.
Reading example · Explaining the workflow
A seed-only candidate
Suppose the model does not fulfill a mechanism request and the system creates a structurally complete candidate seed.
- 01
Record the fallback origin instead of describing the node as a model research result.
- 02
Even with a ledger, the provenance gate blocks a seed-only candidate from review-budget consumption and acceptance.
- 03
Leave the unfinished candidate on the frontier for real role responses and evidence.
An unfinished object with clear provenance, without structural validity hiding missing research content.
A conceptual reading example. No generation, retrieval, review or experiment was run, and no performance result is reported.
Changes in this version
Provenance check
- Earlier approach
Model fields and seeds mix without a direct provenance gate.
- This version’s approach
Record provenance and block seed-only review.
Avoid mistaking system filling for research work.