01
Verify the candidate itself
EulerForge addresses two remaining gaps in Dedalus: hardened modules are not the default route, and mathematical checks can run a fixed task unrelated to the proposed equation. V6.2 adds a real entry route and candidate compilation, connecting equations, algorithms and architectures to their own executors and linking evidence, prior deltas, benchmarks and verdicts through candidate_id.
02
Connect semantics, execution and evidence
Compilation makes the interpretation of a research object explicit. Unknown mathematical operators are rejected, architecture graphs produce executable forward paths, and algorithm source faces property and metamorphic tests. Topical overlap alone cannot support a core claim: background, nearest priors, support and refutations are separated, with unresolved refutations blocking evidence binding.
03
Three compiled routes
Bilingual hard requirements and a domain contract lead to genomes and candidate compilation. Candidate-bound solvers, limits, invariants, code tests, benchmarks and ablations follow. Recurrent refinement and DiscoForge interfaces operate above these objects; Aurora aggregates required routes. Package construction, auditing and replay summaries preserve traceability to candidate identities.
04
Background is not claim support
EquationCompiler uses registered operators to produce rhs_fn and stamps reports with compiled_equation_id and replay_hash. RelationAwareBinding separates supporting, refuting, background, nearest-prior and resolution IDs. Algorithm benchmarks deep-copy and hash inputs before and after execution to detect mutation of shared inputs; permitted mutation must be explicit in the contract.
05
One generation, three packaging variants
The original, Claude Code-adapted and Full-stack packages belong to one EulerForge generation. The adapted package retains discovery_v62 and tests_v62 while adding request-ID-matched host-agent and AnySearch bridges. Full-stack coordinates existing subsystems and host review through a manifest. These are delivery and wiring variants, not three algorithmic generations.
- Original package
Core candidate compilation, execution and evidence binding.
- Host adaptation
Request/response bridges require actual host scheduling.
- Full-stack
Coverage manifests and review gates do not establish causal usefulness.
06
Conceptual walkthrough: a sorting candidate
The default lite sorting candidate wraps Python sorted(), and its priors include Python sorting. The system can compile, test, benchmark and identify overlap with the prior, but the candidate is not a new sorting algorithm. Default mathematical candidates similarly derive from preset graph-diffusion variants. A complete execution chain and scientific originality remain separate questions.
07
Corrected wiring and interpretation
Relative to V6.1, V6.2 wires EulerForge into the CLI, adds Chinese hard hints and runs numerical checks against the candidate's compiled right-hand side. Program checks add layered timeouts and mutation auditing; missing evidence audits fail closed. V6.3 then shifts emphasis toward axiom transformations, quality-diversity archives and generation strategies.
08
What replay means here
Inspected code includes EquationCompiler, subprocess benchmarking, relation binding and verdict aggregation. replay_package_v62 re-canonicalizes artifacts_by_candidate and compares its hash with the manifest; tests cover intact and tampered artifacts. Replay here checks artifact integrity, not re-execution of every experiment. Import-based Full-stack coverage is also not a test of causal effectiveness.
09
Deterministic demonstration and research
The lite route is a deterministic offline demonstration; a host bridge does not automatically launch external agents. Passed gates, structural coverage and consistent artifact hashes support engineering-integrity claims, not originality, performance breakthroughs or third-party replication. The archive preserves all three variants while treating EulerForge as one candidate-compilation stage.
Reading example · Explaining the workflow
From source code to a research package
The default lite candidate wrapping sorted().
- 01
Hard requirements fix the algorithm route.
- 02
Compiled code enters a time-limited subprocess.
- 03
Property, metamorphic and benchmark checks bind its ID.
- 04
Prior comparison exposes overlap with Python sorting.
- 05
Persist the package and recompute artifact hashes.
Inspectable execution and prior records, not a new sorting algorithm.
Conceptual account of the default source; no program or speedup measurement was run here.
Changes in this version
Entry route
- Earlier approach
Dedalus modules not wired into the default CLI
- This version’s approach
EulerForge route and bilingual hard hints
Make the selected flow reachable
Mathematical execution
- Earlier approach
A fixed graph task can be unrelated to the candidate
- This version’s approach
Execute the candidate's compiled rhs_fn
Inspect candidate semantics
Evidence
- Earlier approach
Source structure or background can be misused
- This version’s approach
Require core support and block unresolved refutations
Match evidence to the claim