Appendix: Implementation Evidence Map

This appendix was withheld when this site was first published on 2026-08-12, and it is published now because it was re-graded rather than repaired. Four capabilities were graded Implemented/Tested on classes that a deliberate refactor removed on 2026-08-10 — two days before launch — and one citation named a file that has never existed in any repository. The gradings below record what survived that refactor, what did not, and what was never there. Publishing the correction rather than the original is the point of keeping the record.

Version: 0.4 (2026-08-12)

This appendix maps the five-thesis suite's major claims to current Consullo repository evidence. It is not a proof of capability. It is a review aid that distinguishes implemented code, tested support, design-only specification, proposed extensions, and evidence gaps.

Verification note (2026-08-12): every cited repository path and named test in this appendix was resolved against live files across the whole workspace, against full git history, and against the agentDescriptions registry. Resolution establishes only that a named artifact exists; it does not establish that the artifact supports the sentence citing it, and it does not convert component evidence into end-to-end capability evidence.

This replaces a note claiming the paths were "re-checked against the current repository on April 24, 2026" and that the check "confirms the cited component evidence." That claim had two failures worth recording rather than deleting. It went stale on 2026-08-10, when 199ea7fd removed the Bootstrapper subsystem and four cited files with it. And it was never true of LLMNonInteractiveSession.java, which has zero commits in the history of every repository searched and no registry entry — a path that never existed passed a check that claimed to confirm it. Both citations have been corrected below.

Evidence Status

• Implemented: production or utility code exists in the Consullo codebase.
• Tested: repository tests directly exercise the implemented behavior.
• Documented: design documents or agent descriptions specify the behavior, but implementation evidence is not established here.
• Simulated: deterministic simulation code exists, but it is not operational deployment evidence.
• Proposed: feature-pending or thesis material describes a future capability.
• Gap: the thesis claim requires evidence not currently found in the Consullo codebase.

Rows may use compound tags such as Implemented/Tested when implementation code exists and repository tests exercise at least part of that behavior. Compound tags do not imply full end-to-end deployment or benchmark validation.

Capability Status remains governed by 00-vocabulary-and-invariants.md. Evidence Status is narrower: it records what the Consullo codebase currently shows.

Master Claim

Claim:

Consullo Seed AI is a specified/proposed scaffold for governed recursive capability amplification, not a system that has reached ASI.

Evidence:

• 00-master-abstract.md, 00-master-introduction.md, and 00-master-synthesis.md state the bounded master claim.
• 00-vocabulary-and-invariants.md defines Capability Status, empirical envelope, and no-ASI-status invariants.
• appendix-evidence-ledger-schema.md specifies the canonical evidence-ledger schema at design level.
• risks-and-criticisms.md records falsification signals and anti-thesis risks.

Evidence Status: Documented.

Gap:

No single repository-level benchmark report yet ties the five-thesis scaffold to observed recursive capability improvement.

Thesis 1: Validated Improvement Loop

Claim:

Candidate changes can be proposed, evaluated, validated, staged, monitored, rolled back, and recorded into method memory under explicit evidence gates.

Repository evidence:

• the Main class is the agent runtime. It performs bottom-up agent integration through refreshAgentHierarchy, runAgentRuntime, launchFriendshipBootstrapAsync, and invokeFriendshipCheckIn, which dispatches the checkIn operation. This function was previously carried by agents/Bootstrapper.java and moved here on 2026-08-10 in 199ea7fd.
• the AgentCompilationValidator class runs real Maven compilation, parses output, analyzes dependencies, and emits validation/repair recommendations.
• Dependency validation, repair orchestration, and A2A method-call testing are responsibilities of the agent-integration workstream that are not currently implemented. They were implemented and tested by agents/DependencyValidator.java, agents/RepairIntegrationManager.java, and agents/A2AIntegrationTester.java until 2026-08-10, when 199ea7fd removed the subsystem containing them. No successor class performs these functions anywhere in the workspace. They are graded Documented/Proposed below, not Implemented/Tested, and they are owed rather than abandoned.
• the AgentBuilderReport class reports agent-builder episodic memories, build status, duration, method-memory reuse, and LLM cost summaries.
• the darwinian-godel-machine design document specifies empirical self-improvement through evolved method memories and validation.
• the core-self-improvement-plan design document, self-improvement-techniques.md, learned-agents.md, behavioral-extraction-prompts.md, and method-memory-catalog.md specify future improvement-loop components.

Tests:

• the MainTest class
• the A2AHealthCheckIntegrationTest class
• the AgentBuilderReportTest class

The five Bootstrapper*Test classes previously listed here were removed with the subsystem on 2026-08-10. No test currently exercises dependency validation, repair orchestration, or A2A method-call testing.

Evidence Status: Implemented/Tested for agent integration, compilation validation, and build reporting. Documented/Proposed for dependency validation, repair orchestration, and A2A method-call testing — all three implemented and tested until 2026-08-10, removed with the Bootstrapper subsystem, and owed to the agent-integration workstream. Documented/Proposed for full recursive improvement semantics, DGM-style population update, evaluator/validator separation across all agent types, and method-memory evolution.

Gaps:

• No end-to-end benchmark yet demonstrates repeated accepted improvements improving future improvement capacity.
• The canonical evidence-ledger schema is specified in appendix-evidence-ledger-schema.md, but no repository-wide implementation is established as the operational source of improvement evidence.
• AAF gate integration into acceptance gates is specified, not implemented across the loop.

Thesis 2: Cognitive Substrate

Claim:

Consullo can organize memory, knowledge access, reasoning, attention, metacognition, perception, creativity, and executive control into measurable cognitive workflows.

Repository evidence:

• the 09-executive-functions design document, 10-knowledge-orchestration.md, 11-social-cognition.md, 12-metacognitive-systems.md, 13-perception-processing.md, and 14-creative-functions.md specify cognitive agent families.
• the 16-rapid-knowledge-access design document specifies retrieval and rapid knowledge-access substrate.
• the ChainCompositionEngine class, RoutingDecisionEngine.java, TaskArrivalEngine.java, PortfolioAccumulationEngine.java, SupersessionCascadeEngine.java, ValidationAllocationEngine.java, ModelImprovementEngine.java, and TrustTransitionEngine.java implement deterministic simulations relevant to composition, routing, validation allocation, trust, and model improvement.
• the CognitiveArtifactScenarioLibrary class and CognitiveArtifactTaskLibrary.java define simulation scenarios and task libraries.
• the CognitiveArtifactSummaryFormatter class and CognitiveArtifactResultsTableBuilder.java format simulation results.

Tests:

• src/test/java/com/consullo/cognitiveartifacts/simulation/engines/*Test.java
• the CognitiveArtifactScenarioLibraryTest class
• src/test/java/com/consullo/cognitiveartifacts/simulation/analysis/*Test.java

Evidence Status: Documented for named cognitive agents and architecture. Simulated/Tested for cognitive-artifact composition, routing, validation allocation, trust transition, model-improvement, and portfolio dynamics.

Gaps:

• Named Thesis 2 agents are mostly design artifacts, not implemented Java agents in this repository.
• Simulation evidence does not yet establish real cognitive capability amplification on external task suites.
• Model 2 now has benchmark-family measurement conventions in appendix-formal-models.md, but those units are not yet bound to an operational benchmark report or external task suite.

Thesis 3: Causal-Decision Foundations

Claim:

Recursive improvement should use causal models, counterfactuals, robust intervention choice, prediction calibration, Goodhart checks, experiment portfolios, abstention, and escalation.

Repository evidence:

• the new-causal-prediction-agents design document specifies causal prediction agents.
• the computational-decision-making design document, random-experiments.md, modelling-feature.md, and performance-monitoring-and-enhancement.md specify decision, experiment, modeling, and monitoring concepts.
• the 19-strategic-planning design document and 15-strategic-coordination.md provide planning and coordination context.
• the business.simulation package implements deterministic business/economy simulations, scenario libraries, failure injection, scoring, allocation, evaluation, and venture-funding engines.
• the ainativecommerce.simulation.scenarios package implements AI-native commerce simulation scenarios.

Tests:

• the Chapter5FormulaLibraryTest class
• the MarketStructureCaptureScenarioTest class
• the RoutingAdvantageScenarioTest class

Evidence Status: Documented/Proposed for causal-decision agents and robust causal intervention semantics. Simulated/Tested for selected business and AI-native commerce scenario modeling.

Gaps:

• No implemented Pearl-style causal graph, structural-equation, or counterfactual engine is identified in this repository.
• No prediction-calibration battery or causal prediction backtest report exists.
• No implemented Goodhart checker is wired into acceptance gates.

Thesis 4: Self-Modifying Software Substrate

Claim:

Consullo can constrain code generation, repair, testing, semantic validation, provenance, staged deployment, and permission through an executable software substrate.

Repository evidence:

• the AgentCompilationValidator class provides real compilation validation.
• Dependency and integration-order validation, A2A method-call testing for loaded agent classes, and repair-workflow coordination are specified but not currently implemented. The classes that provided them were removed on 2026-08-10 in 199ea7fd and have no successor; the responsibilities remain with the agent-integration workstream.
• the GenerateConsulloAgents class, AgentGenerationRepair.java, AgentDescriptionRepair.java, AgentCardRepair.java, ComposeVerificationMethodsForAllAgents.java, ComposeDryRunMethodsForPriorityAgents.java, EnsureCheckInMethodInserted.java, and related utilities support agent generation and repair operations.
• the internal agent-programming methodology (not published) and functional-java-subset.md specify LLM-Native Functional Java constraints.
• the atomic-prompts design document and the 04-atomic-prompt-decomposition design document specify atomic prompt and compiled-code orchestration.
• the secure-agent-builder-pipeline design document, agent-modification-specifications.md, post-verification-methods.md, and new-repair-agents-descriptions.json specify future repair and secure builder components.

Tests:

• the AgentCheckInMethodReplacerTest class
• the AgentRenamerTest class
• the AgentBuilderReportTest class
• the A2AAgentDispatchAdapterTest class
• the A2AIngressServletTest class
• the A2ACardRegistryServletTest class

Evidence Status: Implemented/Tested for parts of compilation validation, A2A dispatch, agent utilities, and reporting. Documented/Proposed for dependency validation and repair-workflow coordination, which lost their implementation on 2026-08-10. Documented/Proposed for secure staged deployment, semantic invariants, validator non-regression, provenance proofs, and full repair-pipeline recurrence.

Gaps:

• No formal semantic-validation layer is implemented as described in Thesis 4.
• ValidatorStrength is now structurally defined in appendix-formal-models.md, but held-out validator-non-regression suites and operational measurement remain unimplemented.
• Staged sandbox/canary/production deployment semantics are not established as an operational pipeline here.

Thesis 5: Alignment Invariants And Scoped Trust

Claim:

Recursive modification must be constrained by alignment invariants, AAF dissent, scoped trust, containment, interruptability, evidence preservation, ADO reporting, and human authority.

Repository evidence:

• the Constitution, published here as Constitutional Commitments specifies the constitution, single-owner Phase 1, AAF, ADO, containment, interruptability, and prohibited modifications.
• the adversarial-alignment design document, computational-trust.md, computational-trust-agent-descriptions.json, trustworthiness.md, trustworthiness-agent-descriptions.json, yolo-mode-mitigation.md, and self-preserving-consullo-agents.md specify alignment, trust, trustworthiness, and safety mechanisms.
• the A2AAuthPolicyGate class implements bearer-token authentication, token lifecycle checks, optional scope checks, rate-limit policy controls, and deny decisions for unauthorized A2A ingress.
• the A2AAgentDispatchAdapter class implements normalized dispatch, JSON-RPC validation, authentication-context checks, bounded execution, timeout/overload/error mapping, and idempotency handling.
• the PersonaManager class implements ethical/persona templates including Kantian ethics, utilitarianism, Christian natural law, and custom persona support. This is relevant to AAF critique-source diversity but is not a full AAF implementation.
• the TrustTransitionEngine class and ValidationAllocationEngine.java simulate trust transition and validation allocation.

Tests:

• the A2AAuthPolicyGateTest class
• the A2AAgentDispatchAdapterTest class
• the A2AIngressServletTest class
• the PersonaManagerTest class
• the TrustTransitionEngineTest class
• the ValidationAllocationEngineTest class

Evidence Status: Implemented/Tested for A2A ingress auth/policy gates, dispatch constraints, persona-template support, and selected trust/validation simulations. Documented/Proposed for AAF orchestration, the Friendship agent authority, ADO reporting, scoped trust estimates, containment, interruptability, and recursive alignment improvement.

Gaps:

• AdversarialAlignmentOrchestrator, Friendship, and AbundanceDistributionMonitor have design-level behavioral contracts in appendix-thesis-5-operational-contracts.md, but are not implemented operational owners in this repository.
• No implemented AAF dissent aggregation pipeline is identified.
• No implemented ADO reporting cadence, benefit/harm metric collector, or owner-facing report exists.
• The canonical evidence-ledger schema is specified in appendix-evidence-ledger-schema.md, but no implemented ledger has been identified as canonical for preserving alignment dissent, overrides, incidents, and provenance.

Substrate Context

Claim:

Specialized LLM routing, rapid knowledge access, atomic prompts, internal economy, and A2A/server infrastructure support the five theses without proving capability or alignment.

Repository evidence:

• the 02-specialized-llm-ecosystem design document, 16-rapid-knowledge-access.md, 04-atomic-prompt-decomposition.md, 01-business-orchestration.md, and 06-digital-virtual-economy.md specify major substrate areas.
• the server package implements portions of server, servlet, A2A, auth, dispatch, registry, and observability infrastructure.
• the business.simulation package implements internal-economy simulation support.

The line previously here cited llm/hooks/LLMNonInteractiveSession.java as supporting session memory and hook behavior. It is deleted rather than repointed: that file has zero commits in the history of every repository searched and no registry entry, so it was never evidence for anything. The grading below rests on the remaining infrastructure and does not need a replacement citation.

Tests:

• the server package
• the server.auth package
• the server.dispatch package

Evidence Status: Implemented/Tested for selected infrastructure. Documented/Proposed for specialized LLM ecosystem routing and rapid knowledge access as integrated Seed AI substrate.

Gaps:

• No unified substrate evidence report yet ties LLM routing, memory retrieval, A2A dispatch, internal economy, and thesis acceptance gates into one operational path.
• Internal economy simulations are not evidence that internal transfer pricing improves alignment or external welfare.

Organizational Recursive Self-Improvement

Claim:

Consullo can be interpreted as an AI-native R&D organization whose agents, workflows, benchmarks, ledgers, method memories, validators, and governance gates produce validated improvement of research, engineering, evaluation, memory, and governance processes.

Evidence Status: Documented/Proposed.

Current evidence is concentrated in software-substrate utilities, benchmark/test-plan appendices, the canonical evidence-ledger schema, cognitive-artifact simulations, and the design-level organizational appendix. No implemented weekly organizational RSI loop, frozen V_ref_0, pre-registration ledger, external-evaluator sampling pipeline, kill-switch drill, or portfolio dashboard has been identified in this repository.

Gaps:

• an internal design document (not published) is specified but not implemented.
• No Week 0 completion ledger bundle exists.
• No benchmark report yet demonstrates organizational method-memory transfer, cross-lane improvement, or second-cycle research-productivity gain.
• No implemented controls yet enforce the exploration budget, rate limits, or frontier-model default-deny policy.

Highest-Priority Evidence Gaps Before Publication

Snapshot date: April 24, 2026.

1. Implement or explicitly defer the three load-bearing Thesis 5 owner contracts specified in appendix-thesis-5-operational-contracts.md: Friendship, AdversarialAlignmentOrchestrator, and AbundanceDistributionMonitor. See risks-and-criticisms.md: Owner As Final AAF Adjudicator, AAF Decorative Failure, and Abundance Obligation Vagueness.
2. Implement the canonical evidence-ledger schema from appendix-evidence-ledger-schema.md for improvement evidence, trust evidence, provenance, dissent, incidents, overrides, and rollback state. See risks-and-criticisms.md: Evidence-Ledger Schema Unimplemented Or Unenforced and Self-Modifying Provenance Graph Integrity.
3. Produce an end-to-end improvement-loop demonstration with baseline, proposed change, evaluator evidence, validator results, staged deployment or simulation, cost/benefit report, and rollback semantics. Benchmark design is specified in appendix-thesis-1-improvement-loop-benchmarks.md; implementation evidence remains pending. See risks-and-criticisms.md: Recursive-Improvement Claim Without End-To-End Evidence.
4. Bind Model 2's benchmark-family measurement conventions to at least one external or project-local benchmark report with declared baselines, units, and cost normalization. Benchmark design is specified in appendix-thesis-2-cognitive-workflow-benchmarks.md; implementation evidence remains pending. See risks-and-criticisms.md: Cognitive Architecture Inflation.
5. Add causal-decision implementation evidence: causal graph representation, counterfactual procedure, calibration battery, and Goodhart checker. Benchmark design is specified in appendix-thesis-3-causal-decision-benchmarks.md; implementation evidence remains pending. See risks-and-criticisms.md: Causal Model Overreach and Goodhart And Validator Gaming.
6. Add validator non-regression suites for Thesis 4 and operationalize the ValidatorStrength convention from appendix-formal-models.md. Benchmark design is specified in appendix-thesis-4-software-substrate-benchmarks.md; implementation evidence remains pending. See risks-and-criticisms.md: Software Repair Overclaim and Goodhart And Validator Gaming.
7. Add AAF dissent aggregation and ADO reporting implementation evidence before treating those roles as more than specified/proposed. Benchmark design is specified in appendix-thesis-5-alignment-benchmarks.md; implementation evidence remains pending. See risks-and-criticisms.md: AAF Decorative Failure, Abundance Obligation Vagueness, and Cost Of AAF Risk.