Thermodynamic Evolution

Problem

Automated software improvement needs more than a plausible patch. Diagnostics, candidate generation, validation, and rollback must retain evidence and represent uncertainty.

Architecture

The Go backend separates repository intelligence, diagnostics, remediation planning, and execution. Semantic analysis adds call resolution, dataflow, resource-lifecycle reasoning, and cached function summaries. Rust AST components and an administrative frontend support the broader system.

Implementation

Function summaries encode parameter-to-return taint relationships, resource acquisition/release effects, and error behavior. Summary entries record source hashes and completeness. The executor checks task eligibility and approval before creating a workspace and choosing a registered strategy. The Rust engine computes a weighted structural score from cyclomatic complexity, maximum nesting, log AST size, and concurrency primitives. The Go annealer requires configured candidate and energy evaluators before starting.

Engineering decisions

The LLM mutation agent proposes replacement function bodies; its instructions are not a proof of correctness. The Docker evaluator specifies network isolation, a read-only root, a non-root user, dropped capabilities, resource limits, and bounded output. These are inspected implementation controls, not a claim that every deployment enforces them.

Contribution

This is collaborative work. Merged pull requests #26 and #27 are authored by Zylus08; inspected commits attribute semantic-analysis and verification changes to Saksham Mishra. That evidence supports those contributions, not sole ownership of every subsystem.

Results

VERIFIED by source inspection: semantic summaries, eligibility checks, isolated execution orchestration, and Docker containment arguments are implemented. DOCUMENTED: the latest Phase 1.1 report describes counterfactual and adversarial verification, bounded recursion, and reused evaluation corpora. Those experiments were not rerun here.

Limitations

The latest inspected report explicitly limits the milestone to verification and does not authorize general autonomous remediation. Polymorphic dispatch, ownership escape, multi-hop aliases, and corpus reuse constrain generalization. Small fixture suites do not establish universal diagnostic accuracy or optimization superiority.

Reproducibility

Reviewed at commit d5ae3189dcbc on 2026-11-09. Source inspection establishes implementation structure; repository-authored results are not independent measurements. The upstream application and benchmark suite were not rerun during this review.

• go-backend/agent/repo_intel/diagnostic/semantic/function_summary.go
• go-backend/agent/repo_intel/execution/executor.go
• go-backend/agent/evaluator.go
• go-backend/agent/mutation.go
• reports/milestone3_5_phase1_1_independent_verification.md
• thermodynamic-ast-engine/src/main.rs
• go-backend/agent/annealer.go

The semantic-analysis research note focuses on the scientific question and evaluation limits.

Next investigation

Resolve the limitations above, preserve the reviewed commit and exact experiment configuration, and reproduce correctness checks before promoting performance or deployment claims.