01
Refutation must affect the candidate
Jupiter addresses a contradiction: a candidate can have a ledger whose refutation is then ignored by review. Refutation state now affects lifecycle and acceptance, with node updates, review information and resumption artifacts retained so research-content changes influence the outcome.
02
Model-first generation with seed fallback
Operators request their assigned role, wrap nodes or children responses and fall back to structurally valid seeds when none arrive. Nonempty model fields override seed fields, while defaults remain elsewhere. This supports workflow demonstrations but leaves a provenance question: format validity does not establish that the model supplied the core mechanism.
03
Draft, checked and accepted states
Candidates progress from draft to checked and then toward acceptance, repair or the frontier according to review and refutation. Directly refuted, insufficient or unbound candidates are blocked from ordinary review. Tree-update events preserve changes so later readers can inspect the candidate’s history rather than only its creation record.
04
Consume each role’s output contract
Review consumes compatible score fields, including lens_specific_score and meta_score, and detects framing mismatch. Five roles cover lens, subfield, venue, meta-review and skeptical refutation. Per-role artifacts and valid-response counts remain visible; unresolved responses are not full reviews, and separated roles do not prove independent models were used.
05
Acceptance checks both content and ledger
Acceptance combines internal thresholds, complete role responses, strongest prior, decisive test, low framing mismatch and ledger state. Refutation can block promotion rather than merely appear in the report. Review data is saved and reread from the node. These thresholds define an internal selection policy, not measured novelty or publication probability.
06
Inputs, state and handoff
Inputs are a locked direction and budget with role responses or simulated workflow fixtures. Outputs retain tree-update events, ledger links, per-role review and accepted/frontier lists. Researchers can trace a candidate to its criticism and review, distinguishing content generation, format checks and internal promotion.
07
From correct advancement to correct promotion
Compared with Stars, Jupiter repairs model-first candidate generation, refutation blocking, review-field consumption and lifecycle persistence. A ledger that invalidates a candidate now has consequences. Mixed seed and model content still leaves room for acceptance errors, motivating Galaxy’s provenance flags and seed-only gate.
08
Trace a promotion back to its basis
Read backward from acceptance through complete role responses, ledger permission and the candidate’s mechanism and prior-work difference. This explains why an apparently adequate score may still not permit handoff. Coverage by prior work, missing decisive tests or severe framing mismatch must have consequences beyond a risk paragraph at the end of the report.
Generic updates and lifecycle events persist revised objects. Readers can distinguish creation, review write-back, refutation transitions and acceptance rather than treating them as one generation. The audit explains where changes occurred, while ledger citations, review rationales and model responses still need to substantiate those changes.
09
Workflow validity and content validity remain separate
The archive’s simulated and seed materials check how the workflow connects. They do not establish complete real-model candidates, independent cross-model review or comprehensive live mathematics–physics refutation. This account statically checks source and documents. Jupiter’s historical contribution is connecting refutation to promotion rather than guaranteeing high-novelty results.
Reading example · Explaining the workflow
A candidate refuted by prior work
Suppose a candidate’s core mechanism is fully covered by the ledger’s strongest prior.
- 01
Bind the ledger state to the candidate identity instead of only writing criticism.
- 02
The review entry checks direct refutation and blocks promotion based on self-scores.
- 03
Save the state change and retain frontier material for repair or stopping.
A refutation record with runtime consequences; the candidate needs a changed mechanism or framing.
A conceptual reading example. No generation, retrieval, review or experiment was run, and no performance result is reported.
Changes in this version
Refutation effect
- Earlier approach
Ledger and acceptance can become disconnected.
- This version’s approach
Refutation blocks review and promotion and is persisted.
Let evidence affect selection as well as report wording.