01
Advance the main path more reproducibly
Nova targets remaining scheduling randomness in Cosmos: early mathematics–physics branches can proliferate while experiments and attacks never reach a candidate. It prioritizes the first candidate path, uses deterministic bridge triggers and permits finite terminal progress. The objective is reproducible execution rather than a claimed research success rate.
02
Priorities before the first candidate
While no candidate exists, the scheduler prefers non-mathematics–physics buckets so MathFormalization, PhysicsInterpretation and NovelMapping do not dominate. It falls back when no other bucket exists and restores normal scheduling after the first candidate. This is a stage policy rather than a permanent ban on cross-domain exploration.
03
Deterministic bridge triggers and early quotas
A depth condition replaces random triggering: eligible lenses with depth % 4 == 0 can consider the bridge, subject to pre-candidate quotas. Budget tiers guide quotas, which relax after a candidate forms. This explains why expansion occurs more consistently, without promising deterministic research outputs from variable model responses.
04
Finite terminal progress
After the main loop, existing experiments or attacks without an IdeaCandidate can receive finite terminal steps. The runtime prefers idea_finalize from an attack or reviewer_attack from an experiment, with at most four such attempts and candidate recounting. It closes near-complete paths without guaranteeing a valid candidate every run.
05
Repair budgets and provenance checks
Repair checks global node counts and review budgets and records stop reasons at limits. Core-field coverage and ledger lifecycle continue from Cosmos, distinguishing seed drafts, model-filled candidates and seed-only paths. Terminal progress occurs outside the main loop, so max_nodes is not an absolute whole-run budget bound.
06
Inputs, execution artifacts and outputs
Inputs remain a locked envelope, budget and role responses. Outputs add clearer selection, bridge-trigger, terminal-progress and budget-stop records. Readers can inspect priority reasons, field sources and repair stops. The outcome can still be unfinished frontier material rather than seed content presented as a finding.
07
From stage scheduling to deterministic progress conditions
Relative to Cosmos, Nova addresses random bridge triggering, candidate-path starvation and repair growth while retaining provenance, ledger and review gates. The strategy becomes interpretable: first form a reviewable object, then restore broader bridge exploration. This provides engineering experience for later systems without implying a quantified novelty improvement.
08
What reproducible progress means
Deterministic conditions remove a random trigger from eligible bridge nodes, while candidate-first ordering gives the early main path clearer priority. Reproducibility mainly concerns execution choices: why a depth triggers the bridge and why experiments or attacks receive attention. Content still depends on models, retrieval and evidence, so repeatable routing is not a promise of identical candidates or conclusions.
Stopping needs equally precise explanation. Main-loop limits, finite terminal progress and repair guards are checks at different points. Reports should identify the stop point and retain unfinished fields and reviewable objects. These materials support resumption or a human decision to stop investing. Nova preserves failure outcomes while making their execution reasons inspectable.
09
The actual scope of stability records
The changelog reports simulated workflow stability and leaves real eight-hour model runs and full live retrieval unverified. None were run here. Source establishes candidate-first selection, depth triggering, terminal progress and budget guards, not reliable high-novelty results or independent scholarly conclusions from accepted or confirmed_novel states.
Reading example · Explaining the workflow
Candidate progress near the budget limit
Suppose the main search has experiments and reviewer attacks but no deliverable candidate.
- 01
Inspect whether candidate-first scheduling protects the main path and check depth conditions and bridge quotas.
- 02
Attempt finite terminal actions after the main loop and recount candidates.
- 03
If repair follows, check node and review budgets and save the actual reason to continue or stop.
A search record explaining progress and stop conditions, still requiring real-model, literature and experimental assessment.
A conceptual reading example. No generation, retrieval, review or experiment was run, and no performance result is reported.
Changes in this version
Scheduling stability
- Earlier approach
Random bridge triggers can divert the first candidate path.
- This version’s approach
Candidate-first selection, deterministic depth conditions and early bridge quotas.
Reduce random branch selection and protect the main path.
Budget interpretation
- Earlier approach
Repair growth lacks a global-node guard.
- This version’s approach
Guard repair by node budget and account for terminal progress separately.
Explain real stop conditions without an absolute budget guarantee.