AutoResearch/Innovation Generator/V4/Jupiter

VERSION / V4.5

Jupiter

Let refutation change a candidate’s fate

Jupiter separates draft, checked, refuted and internally accepted candidate states. Ledger decisions return to the tree and block unsuitable candidates; state changes and events are saved so the tree and final report remain consistent.

Historical version · Design & implementation

Current: Daedalus v7.4.0

V4.5 / ARCHITECTURE

How does an exploration move forward?

01模型响应或 seedModel response or seed02候选草稿Candidate draft03台账反驳Ledger refutation04五类评议契约Five review contracts05生命周期晋升Lifecycle promotion06接受或待修补前沿Accepted or repair-pending frontier01模型响应或 seedModel response or seed02候选草稿Candidate draft03台账反驳Ledger refutation04五类评议契约Five review contracts05生命周期晋升Lifecycle promotion06接受或待修补前沿Accepted or repair-pendingfrontier
Architecture illustration derived from this version’s actual control flow.

Paths & conditions

03 → 04
Ledger permits review
03 → 06
Refuted or insufficient
  1. 01

    Model response or seed

    Nonempty responses override defaults

  2. 02

    Candidate draft

    Save content and update events

  3. 03

    Ledger refutation

    Block refuted, insufficient or unbound candidates

  4. 04

    Five review contracts

    Read role fields and framing mismatch

  5. 05

    Lifecycle promotion

    Combine content, ledger and internal conditions

  6. 06

    Accepted or repair-pending frontier

    Retain per-role reasons

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

Inputs

Suppose a candidate’s core mechanism is fully covered by the ledger’s strongest prior.

  1. 01

    Bind the ledger state to the candidate identity instead of only writing criticism.

  2. 02

    The review entry checks direct refutation and blocks promotion based on self-scores.

  3. 03

    Save the state change and retain frontier material for repair or stopping.

Output

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.