Consullo cornucopia mark Consullo Public research record

Cross-Thesis Dependency Map

Version: 0.3.1

This file defines the import/export contract among the five Consullo Seed AI theses. It exists to prevent the suite from becoming a set of overlapping agent rosters.

Dependency Principle

Each thesis has one primary conceptual responsibility. Agents may appear in multiple discussions, but every agent or function should have one primary home. Cross-thesis use should be described as an import rather than duplicate ownership.

Imports are marked as:

Capability Flow

Thesis 2: Cognitive Substrate -----------------------> Thesis 1: Validated Improvement Loop
              \                                       ^
               \                                      |
                v                                     |
Thesis 4: Self-Modifying Software Substrate ----------+
                ^                                     |
               /                                      |
Thesis 3: Causal-Decision Foundations ----------------+

Capability flow shows how capabilities feed the improvement loop. Thesis 2 and Thesis 3 both feed Thesis 1 directly and also inform Thesis 4. It does not show permissioning.

This diagram shows capability flow only, not all definitional imports. For example, Thesis 4 imports Thesis 1's acceptance vocabulary, but that back-import is definitional rather than a capability-flow dependency.

Constraint Wrapper

+-----------------------------------------------------------------------+
| Thesis 5: Alignment Invariants And Scoped Trust                       |
|                                                                       |
|  constrains Thesis 1, Thesis 2, Thesis 3, Thesis 4, substrate context |
|  via AAF gate, scoped permission, trust estimates, human authority     |
+-----------------------------------------------------------------------+

Thesis 5 is not downstream output of the improvement loop. It is a constraint layer over all theses and all substrate context.

Organizational Operating Layer

appendix-organizational-recursive-self-improvement.md defines a cross-thesis operating layer, not a sixth thesis. It interprets the five theses as an AI-native R&D organization whose work product is validated improvement of research, engineering, evaluation, memory, and governance processes.

Organizational functionPrimary thesis importsEvidence boundary
Agenda and portfolio formationThesis 1 improvement targets; Thesis 3 portfolio reasoning; Thesis 5 permission boundariesSpecified/proposed until research_agenda and portfolio_decision ledger records exist.
Hypothesis and candidate generationThesis 2 cognitive search, brainstorming, negative-space mapping; Thesis 1 proposal semanticsCandidate generation only; accepted improvement requires downstream gates.
Pre-registration and experiment designThesis 3 causal-decision discipline; Thesis 1 evidence packagesCapability claims require pre-registered success/failure criteria.
Implementation and validationThesis 4 software substrate; Thesis 1 acceptance semanticsJava repair is exploitation evidence, not full organizational RSI evidence.
Adversarial review and governanceThesis 5 AAF, scoped trust, owner authority; Thesis 1 acceptance gatesHigh-stakes or externally consequential actions route through I12.
Institutional memory and post-cycle learningThesis 1 method-memory update; Thesis 2 memory and anti-library functionsOrganizational learning requires later measured reuse or transfer.

The live execution controls for this layer are specified in the internal execution plan. That document is operational, not part of the publication thesis claim.

Cycle-Breaking Rule

There is an unavoidable conceptual relationship between Thesis 1 and Thesis 2: cognitive agents help improve the system, and the improvement loop evaluates cognitive agents. This is not a circular proof if the import types are separated:

No thesis may use its own future success as evidence for its current capability claim.

Master Frame

The master frame exports:

All theses import the master frame.

Substrate Context

Substrate context is not a sixth thesis, but it should be documented in appendix-substrates.md.

Substrate context owns:

Thesis 5 may evaluate these systems for trust and alignment. It does not own them as alignment mechanisms.

Thesis 1: Validated Improvement Loop

Primary responsibility:

Model recursive capability amplification as a staged, evidence-gated loop over agents, method memories, code, tests, workflows, policies, and agent populations.

Exports:

Imports:

Primary agent/function home:

Not owned here:

Thesis 2: Cognitive Substrate

Primary responsibility:

Model Consullo's modular cognitive substrate for capability amplification: memory, reasoning, perception, attention, metacognition, social cognition, creativity, executive control, knowledge access, and orchestration.

Exports:

Imports:

Primary agent/function home:

Theory-of-mind agents primarily support the Adversarial Alignment Function and are owned by Thesis 5, while Thesis 2 may describe them as cognitive primitives:

Not owned here:

Thesis 3: Causal-Decision Foundations

Primary responsibility:

Model causal prediction, counterfactual reasoning, causal influence diagrams, robust decision-making, experiment selection, strategic bias control, and model-misspecification handling.

Exports:

Imports:

Primary agent/function home:

Note: SuperhumanExperienceMiner is a legacy codebase name. Its name is not a claim that superhuman capability is implemented.

Not owned here:

Thesis 4: Self-Modifying Software Substrate

Primary responsibility:

Model the executable substrate that lets Consullo generate, edit, repair, test, validate, document, and deploy agents and code.

Exports:

Imports:

Primary agent/function home:

Not owned here:

Thesis 5: Alignment And Scoped Trust

Primary responsibility:

Model alignment invariants, scoped trust, evidence ledgers, permissioning, containment, rollback, incident response, adversarial alignment, abundance obligation, and human authority under recursive modification.

Exports:

Imports:

Primary agent/function home:

Notes under single-owner Phase 1:

Not owned here:

Cross-Cutting Constraints

Single-Owner Phase 1

All theses must treat single-owner Phase 1 as the baseline. Future contractor oversight and multi-stakeholder governance are later phases or stress-test scenarios.

Bounded Compute

All theses must account for token cost, compute cost, latency, validation cost, and opportunity cost.

Capability Status

Every major capability claim must identify whether it is implemented, specified, proposed, or speculative.

Benchmark Discipline

Benchmarks are evidence sources, not proof of intelligence or safety. Every benchmark claim must state task class and limitation.

Evidence Ledger Discipline

Evidence ledgers must be treated as audit-preserving records. Rollback annotates or supersedes evidence; it does not erase inconvenient history.

The canonical evidence-ledger schema is maintained in appendix-evidence-ledger-schema.md.

AAF Gate Discipline

The Adversarial Alignment Function must be routed into Thesis 1 acceptance gates according to invariant I12 in 00-vocabulary-and-invariants.md.

Provenance Discipline

Accepted modifications must satisfy invariant I11.

Cost/Benefit Discipline

Accepted non-emergency improvements must satisfy invariant I17.

Deceptive Optimization Discipline

Improvements to learned subsystems must satisfy invariant I19.

Overclaim Control

Sub-theses should use "scaffold for," "pathway to," "bounded," "measurable," and "governed" where appropriate. Avoid unqualified claims of superintelligence, solved alignment, or guaranteed improvement.

Changelog