AutoResearch/Innovation Generator/V4/LA-PATS

VERSION / V4.1 / V4-pro

LA-PATS

Make research ideation persistent and inspectable

V4 expands the one-shot protocol into seven search spaces: domains, contribution lenses, kernels and anchors, mathematics, physics, experiments and reviewer attacks. The scope and budget are fixed first; candidates and refutations become structured state before review, repair and termination.

Historical version · Design & implementation

Current: Daedalus v7.4.0

V4.1 / V4-pro / ARCHITECTURE

How does an exploration move forward?

01搜索包络Search envelope02文献边界图Literature frontier03七空间搜索树Seven-space search tree04反驳台账Refutation ledger05评议与修补Review and repair06报告或失败前沿Report or failurefrontier01搜索包络Search envelope02文献边界图Literature frontier03七空间搜索树Seven-space search tree04反驳台账Refutation ledger05评议与修补Review and repair06报告或失败前沿Report or failure frontier
Architecture illustration derived from this version’s actual control flow.
  1. 01

    Search envelope

    Field, lens, boundaries and budget

  2. 02

    Literature frontier

    Saturated, underexplored and collision regions

  3. 03

    Seven-space search tree

    Operator expansion and conditional mathematics–physics bridge

  4. 04

    Refutation ledger

    Strongest priors and evidence gaps

  5. 05

    Review and repair

    Quality floors and candidate revisions

  6. 06

    Report or failure frontier

    Acceptance conditions and stop reasons

01

Make exploration leave a trail

V4.1 / V4-pro moves the model protocol toward a resumable execution system. Researchers can inspect how a direction branched, which paths were refuted and why the run stopped or repaired a candidate. Failure to find a qualifying direction becomes a legitimate, explainable handoff.

02

Lock the search envelope first

The envelope records field, subfield, contribution lens, baseline classes, forbidden regions, target and budget. A literature frontier distinguishes saturated, underexplored and collision regions. These commitments keep the run on one problem, although this execution entry still initializes some envelope fields as placeholders awaiting model content.

03

Seven spaces form the search scaffold

Seven search spaces cover domains, lenses, kernels and anchors, mathematics, physics, experiments and reviewer attacks. Python selects nodes and operators, with a conditional mathematics–physics bridge. Cheap priorities schedule exploration; refutation, review and quality gates determine final acceptance. Their roles are deliberately distinct.

04

State files are part of the design

Each run stores tree.jsonl, refutation_ledger.jsonl, run_state.json, decision_log.md and final_report.json. Nodes record ancestry, depth, operation, content, evidence and budget. Checkpoint rationales enter a decision log, and a resume entry reads existing state. These artifacts make the search a shared object rather than scattered conversation.

05

Quality floors alongside scores

Seven floors address framing, strongest-prior comparisons, failure modes, theoretical conditions, alternative explanations, blind spots and measurement protocols. Repairs target these gaps rather than add arbitrary components. A failed gate preserves a useful frontier and explanations. These are protocol conditions, not automatic evidence of real novelty.

06

Inputs, outputs and traceability

Inputs comprise a direction, optional prior work and a Lite, Standard or Heavy budget. Intended outputs include candidates, refutations, reviews, repairs and stop reasons. The supplied scaffold defines these artifacts and stage interfaces, but operators still construct placeholder content and parts of review retain scaffold scores. The design and implementation maturity should be read together.

07

State becomes a system object after V3

V3 relies on a model following a long protocol and delivering a proposal. This stage adds orchestration, tree storage, configuration and explicit outcome gates. The v4.1 ZIP also contains V4-pro and a tested summary, so it is presented as one composite snapshot rather than several independently established releases.

08

Read a run through one node

Choose one candidate identity, trace its parents to the operation that developed a mechanism from the question, follow evidence fields to the ledger and compare checkpoint decisions. This connects tree structure, research text and stop reasons. Fields awaiting role fulfillment reveal specific gaps without requiring readers to guess progress from the final report.

The early TreeStore saves JSONL nodes with an in-memory index and can rewrite storage for status changes; resumption reads run_state. Persistence preserves an object without validating it. Verified flags, scores and literature fields still require content inspection. Separating storage from research judgment provides a foundation for Hill’s semantic checks and Jupiter’s lifecycle repairs.

09

Protocol examples and capability evidence

The tested summary states that one model played the roles inline and recalled prior work without live retrieval. Source also exposes placeholder envelopes and nodes. This archive establishes how persistence and review were organized, not independent multi-model validation, live literature coverage or reliable production of high-novelty outcomes.

Reading example · Explaining the workflow

Reading a resumable search

Inputs

Suppose the input asks when a language model should stop reasoning, with a finite node budget.

  1. 01

    Lock the question, intended contribution and existing stopping strategies to compare.

  2. 02

    Save nodes and ancestry across kernel, mechanism, experiment and refutation branches.

  3. 03

    Check for real prior-work comparisons and expose placeholder content as unfinished work.

  4. 04

    Use run state and the failure frontier to record which evidence needs attention next.

Output

Traceable research artifacts and explicit reasons to continue, repair or stop.

A conceptual reading example. No generation, retrieval, review or experiment was run, and no performance result is reported.

Changes in this version

Research state

Earlier approach

Primarily a model protocol and conversation.

This version’s approach

Nodes, ledgers, checkpoints and terminal state are stored separately.

Make exploration resumable and decisions explainable.