01
Problem and philosophy
Aurora addresses responsibility after proposal generation. A polished, highly scored idea may still lack its equation, prior comparison or failure history. V5.0 turns those gaps into inputs to a separate verification layer, so a final status can be traced to its evidence.
Its central commitments are graded claims, recorded failures and explicit precedence. Physics models, equations and conjectures have different evidence requirements. Reviewer agreement cannot replace missing checks, and candidates begin unverified.
02
From input to candidate
The entry path remains V4-Nova’s persisted search: envelope lock, frontier tensions, tree search, refutation ledger, panel evaluation, repair and stopping. Aurora is appended as an eighth phase and consumes the resulting artifacts.
The runner checks the contract and workspace, then conditionally dispatches gates for the supplied artifacts. The terminal aggregator handles blocking failures and unresolved warnings before budget exhaustion and missing required passes; absence of a check is not treated as success.
03
Modules and state
Propose first, verify explicitly. Reviewer agreement cannot override tool failure; software evidence gates make status auditable rather than establish scientific truth.
Terminal ordering matters: excessive review rounds or a missing failure ledger can refuse output; blocking gates, unresolved warnings and exhausted budgets follow distinct paths. Status describes evidence within a call, not universal scientific certification.
- Inherited search
Retains V4-Nova’s seven-phase persisted search and appends an Aurora verification phase.
- Verification inputs
AuroraInput gathers the contract, idea, priors, models, equation, theorem, counterexample search, reviewer records, claim registry and replay lock.
- Claim-sensitive dispatch
Gate 0–12 checks run according to supplied claim types; optional payloads control which validators execute.
- Terminal and records
The terminal aggregator records accepted, promising, refuted, honest-failure or refused status alongside gate and warning logs.
04
Inputs, outputs and contracts
The research contract specifies field, subfield, contribution lens, output types, success criteria, falsification conditions and claims forbidden without evidence. AuroraInput joins that contract to the run directory, agent manifest and optional candidate payloads.
An idea need not include a theorem, but adding mathematical or physical claims activates additional requirements. Outputs retain gate results, high warnings and failure records, allowing readers to follow a verdict back to its checks.
- Inputs
Research direction, mode and search budget; candidates and supporting artifacts from a resumable workspace.
- Outputs
A v5_verification report block plus gate_results.jsonl, high_warn_register.jsonl and final_terminal_gate.json.
05
A conceptual reading scenario
For a candidate claiming that memory improves a dynamics model, inspect the definition of improvement, then follow the model, equation, prior and counterexample records. Reviewer approval without an equation payload should leave an unresolved candidate; an exhausted run preserves its failure trail.
06
Changes from the preceding stage
Compared with score- and panel-driven stopping, Aurora assigns final reporting authority to independent gates. It keeps the older search tree and recovery mechanisms; the change is primarily in evidence status and reporting authority.
The design exposes an upstream limitation: validators can inspect only artifacts that exist. More gates cannot supply missing equations, definitions or conjectures. V5.2 subsequently expands the generation layer to produce these objects.
07
What the source establishes
Conditional dispatch and terminal precedence are the clearest implementation evidence. The default backend remains a stub, and the contemporary Claude Code integration is primarily hook-based. Archived software-test claims do not measure scientific discovery success.
Python gate orchestration and terminal classification; structural dimension/limit/conservation checks, circularity detection, failure records and claim linkage.
08
Historical scope and next steps
The default backend is a stub; the contemporary Claude Code backend is hook-only and requires external subagent orchestration. Validators inspect supplied evidence and fields; they do not imply external domain validation, machine-checked proof or an exhaustive prior-art search. Documented software-test outcomes are not scientific novelty or discovery-success results.
Reading example · Explaining the workflow
Reading example: trace one candidate
For a candidate claiming that memory improves a dynamics model, inspect the definition of improvement, then follow the model, equation, prior and counterexample records. Reviewer approval without an equation payload should leave an unresolved candidate; an exhausted run preserves its failure trail.
- 01
Read the contract: define improvement and forbidden unsupported claims.
- 02
Check whether the candidate actually supplies its model, equation and priors.
- 03
Check that required gates exist and unresolved warnings are visible.
- 04
Distinguish continued exploration, refutation and exhausted budget.
Locate the saved candidate, diagnostics and status, and identify evidence that remains absent.
This is a conceptual reading scenario derived from archive structures, not a new run, discovery result or performance experiment.
Changes in this version
Final authority
- Earlier approach
Score and panel stopping
- This version’s approach
Independent gate aggregation
Bind reporting to inspectable evidence.
Candidate status
- Earlier approach
A complete report could obscure gaps
- This version’s approach
Unverified by default; claim-sensitive promotion
Distinguish ideas, equations and mathematical claims.
Failure record
- Earlier approach
Failure could remain narrative-only
- This version’s approach
Terminal checks require a failure ledger
Preserve a trail for later work.