Appendix: Thesis 4 Software Substrate Benchmarks

Version: 0.3 (2026-06-05) — adds the execution-deterministic principle (compile-once-run-many, model out of the execution loop, accuracy gate against silent failures, bounded runtime invocation), per Compiled AI: Deterministic Code Generation for LLM-Based Workflow Automation, arXiv:2604.05150. Prior: 0.2 (2026-06-05) — verifier-hardened principle per Let the Barbarians In: How AI Can Accelerate Systems Performance Research, arXiv:2512.14806; 0.1 (2026-04-24).

This appendix specifies benchmark families for Thesis 4, A Self-Modifying Software Substrate With Acceptance Gates. It is a benchmark design contract, not an implemented software-repair benchmark suite. Capability Status: specified/proposed. Evidence Status: Documented/Proposed.

The goal is to make software-substrate claims testable without implying that Consullo has a complete autonomous coding or deployment system. A software change earns stronger status only when it passes specified evidence classes: specification, generation or repair, compilation, tests, semantic checks, security, provenance, cost/benefit, deployment discipline, permission, and observation.

Benchmark reports produced under this appendix should populate benchmark_result records in appendix-evidence-ledger-schema.md; the report fields below define the benchmark-specific evidence_payload structure for those records.

Benchmark Principles

Required Benchmark Report Fields

FieldMeaning
benchmark_idStable identifier and version.
repair_domainCompilation, dependency, JSON/schema, semantic behavior, security, deployment, validator, or provenance task class.
specification_recordScope, expected files, non-goals, validation plan, and rollback expectation.
baseline_failureReproducible failure, regression, or missing capability.
candidate_patchPatch source, tool/model/method memory, and changed artifacts.
validation_matrixCompilation, tests, semantic checks, security, provenance, cost/benefit, and permission checks.
reference_suiteKnown-good, known-bad, adversarial, and historical false-accept/false-reject cases.
ValidatorStrength_ruleComposite scoring rule or pending calibration note.
deployment_stateSandbox, simulated canary, canary, production-equivalent, or no deployment.
observation_recordPost-change behavior, incidents, rollback, or supersession.
ledger_linksBenchmark, provenance, improvement, incident, and trust records.
status_resultWhat claim status, if any, the benchmark can strengthen.

Suite A: Deterministic Local Repair

Purpose: test narrow repair machinery with reproducible expected outcomes.

Representative tasks:

Suggested dimensions:

Negative controls:

This suite supports component-level repair claims, not full self-modifying substrate validation.

Suite B: Semantic Near-Miss Detection

Purpose: test whether the substrate rejects patches that pass surface checks while weakening behavior.

Representative tasks:

Suggested dimensions:

Negative controls:

Semantic validation is the main boundary between code generation and safe software modification.

Suite C: Validator Non-Regression

Purpose: test whether validators and generated-test pipelines improve without weakening gates.

Representative tasks:

Suggested dimensions:

Negative controls:

ValidatorStrength should reward discrimination, not permissiveness or blanket conservatism.

Suite D: Security And Privilege Boundary

Purpose: test whether software changes touching sensitive areas route through stricter controls.

Representative tasks:

Suggested dimensions:

Negative controls:

Security-sensitive repair cannot inherit evidence from low-risk compilation repair.

Suite E: Staged Deployment And Rollback

Purpose: test whether accepted patches move through explicit exposure states and retain recovery paths.

Representative tasks:

Suggested dimensions:

Negative controls:

Deployment discipline is part of the acceptance gate, not an operations afterthought.

Minimal Demonstration Package

The first Thesis 4 demonstration should use one bounded low-risk repository-local task. It should include one accepted repair, one rejected semantic near miss, one validator-reference-suite case, and one staged exposure record.

Required contents:

Non-Claims

This appendix does not claim that Consullo has implemented a complete autonomous software engineer, a secure deployment pipeline, full semantic validation, validator non-regression, or production self-modification. It specifies what benchmark evidence would be needed before Thesis 4 claims can strengthen beyond component-level implementation and specified/proposed architecture.