01
First, persist exploration
V4.1 / V4-pro establishes the envelope, literature frontier, seven-space tree and run artifacts. Candidates, criticism, decisions and failure frontiers become resumable objects, while placeholders and scaffold scores expose the distinction between a complete workflow and completed research.
02
Then, check content and phases
Hill adds a bridge and semantic gates, Ocean connects major operators to model roles, and Stars repairs pause advancement and literature fallback. The snapshots are not simple supersets: Stars returns to local tree skeletons, requiring separate inspection of generation and control flow.
03
Make evidence affect acceptance
Jupiter connects refutation to candidate lifecycle and persists updates. Galaxy adds seed provenance, ledger matching, reviewer vetoes and call budgets. Even complete formatting cannot let a mean score erase refutation or system-filled content.
04
Advance candidates and revalidate repairs
Cosmos prioritizes later stages, audits core-field provenance, refutes repairs again and strengthens UUID identity. Nova then handles random bridges, first-candidate starvation and budget growth with deterministic triggers and finite terminal progress. Research objects become more traceable while scientific value still needs independent assessment.
05
Contracts retained across the generation
The tree records where exploration went; the ledger records prior-work refutation; run state records the current blocker; role reviews explain continuation or stopping. V4 separates these responsibilities and links them through candidate identity, helping distinguish execution defects from failures of the research question.
06
How to read the engineering progress
Simulated and golden-test records help explain the workflow, not reliable production of real high-novelty results. Snapshots retain placeholders, seed fallback, incomplete live retrieval and budget exceptions. These limits accompany the architecture; sequencing uses versions rather than invented public release dates.
How this series evolves
- V4.1 / V4-pro
V4.1 / V4-pro · LA-PATS
Compared with the V3 model protocol, this archive adds persistent state, structured gates and engineering orchestration. Although the ZIP is named v4.1, it also contains the V4-pro execution layer and a V4-pro-tested summary.
- V4.2
V4.2 · Hill
Adds runtime gates, SubagentBridge and SemanticValidator to the V4-pro scaffold, moving the focus from protocol completeness to exposing execution defects.
- V4.3
V4.3 · Ocean
Compared with Hill, main search operators stop directly fabricating placeholder children and model calls enter the research-generation path.
- V4.4
V4.4 · Stars
Compared with Ocean, repairs scope and literature gates and adds type-driven scheduling and complete report fields. This snapshot’s policy also returns to local seed construction, so it should not be described as fully retaining Ocean’s content-generation capability.
- V4.5
V4.5 · Jupiter
Compared with Stars, repairs default candidate generation, ledger blocking, state persistence and the draft→checked→accepted lifecycle.
- V4.6
V4.6 · Galaxy
Compared with Jupiter, focuses on provenance, output contracts, rejection and budget constraints rather than merely passing an idealized simulated path.
- V4.7
V4.7 · Cosmos
Compared with Galaxy, adds stage scheduling, core-field provenance audits, strict UUID binding and a ledger lifecycle for repairs.
- V4.8
V4.8 · Nova
Compared with Cosmos, addresses random triggering, candidate-path starvation and repair-budget growth, with separate checks for seed drafts, model-filled candidates and seed-only runs.
Versions in this series
LA-PATS
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.
Read the full introductionHill
Hill focuses on engineering reliability. A model-call bridge and structural and semantic validation turn missing responses, placeholders and empty refutation ledgers into pauses or explicit invalid states, with review records saved separately.
Read the full introductionOcean
Ocean assigns tree expansion to specialized model roles for field splitting, kernel mapping, anchor transfer, mechanism construction, experiment design and refutation. Python wraps and records their responses; the mathematics–physics bridge becomes a staged model-call pipeline.
Read the full introductionStars
Stars repairs advancement after a pause, frontier construction without a literature response and incomplete final-report fields. Operators are scheduled by node type, with visits and attempted actions recorded to make process state explicit.
Read the full introductionJupiter
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.
Read the full introductionGalaxy
Galaxy accepts both single-node and list model responses, injects kernel and anchor seeds and uses provenance flags to keep seed-only candidates out of the internal acceptance path. Reviewer rejection, candidate–ledger binding and call budgets become runtime constraints.
Read the full introductionCosmos
Cosmos uses stage-bucket scheduling to connect mechanism, experiment, attack and candidate stages, reducing the starvation of later-stage nodes by exploration scores. It also tightens model-field coverage, ledger identity binding and renewed refutation after repairs.
Read the full introductionNova
Nova replaces random mathematics–physics triggering with a deterministic depth condition, prioritizes the main candidate path until a first candidate exists and limits bridge branches. Terminal progress and repair-budget guards reduce branch starvation and unchecked repair growth.
Read the full introductionWhat carries forward
Persist research objects with identity and ancestry.
Constrain promotion with criticism, provenance and role decisions.
Repair changes the object and requires renewed evidence.
Allow failure-frontier handoffs and explicit stop reasons.
Implementation & evidence across the series
These eight local snapshots are identified through changelogs and source. No independent reruns or scientific-effectiveness measurements were performed, and each version is not claimed to be a strict superset of its predecessor.
