Consullo cornucopia mark Consullo Public research record

Master Introduction

Framing Problem

Consullo Seed AI is best understood as an attempt to make recursive capability amplification governable. The central problem is not simply whether an AI system can write code, route tasks, retrieve memory, or coordinate specialized agents. The harder problem is whether those capabilities can be organized into a loop that improves future versions of the system while preserving evidence, alignment constraints, rollback paths, and human authority.

This thesis suite therefore avoids treating "Seed AI" as a slogan for inevitable superintelligence. It treats Seed AI as an architectural hypothesis: if a system has modifiable agents, durable memories, measurable tasks, software-repair machinery, causal-decision support, and enforceable governance, then recursive improvement can be studied as an empirical engineering process rather than as mythology. The appropriate question is not "Has Consullo reached ASI?" The appropriate question is: what would have to be true for Consullo to function as a scaffold for governed recursive capability amplification, and what evidence would falsify that claim?

The organizational interpretation makes the target broader than code self-editing. Consullo's recursive-improvement object is the AI-native R&D organization as a whole: research agendas, hypotheses, workflows, evaluators, benchmarks, method memories, portfolio choices, and governance routines. Java repair remains the first credible exploitation lane because it is concrete and testable, but it is not the whole Seed AI process. The live execution plan therefore applies a stricter-not-looser safety frame as ambition rises: kill switch, sandboxing, frozen reference suites, pre-registration, external-evaluator sampling, model-family diversity, exploration-budget discipline, and stop conditions are prerequisites for operational cycles, not optional polish.

Current Claim Boundary

The current claim is deliberately bounded. Consullo's existing and proposed designs define a specified and proposed architecture, not a completed demonstration of general superintelligence. Capability Status therefore matters throughout the suite. Some components may be implemented, some specified but not implemented, some proposed extensions, and some speculative research targets. A thesis may argue that a component is architecturally necessary or promising, but it must not treat specification as deployment evidence.

Greater-than-human capability is also treated operationally. It means measurable advantage along defined axes such as speed, coding repair performance, retrieval fidelity, forecasting calibration, causal prediction error, or coordination cost. It does not automatically imply quality superintelligence or general superiority over human judgment. The suite uses greater-than-human language only when the relevant axis, benchmark family, and evidence limitations are named.

Empirical Self-Improvement Ceiling

The public empirical ceiling for self-improving AI systems is still narrow. Goedel-machine theory supplies a formal ideal of self-referential systems that prove the value of their own modifications, but current practical systems are empirical relaxations. Work such as the Darwin Godel Machine demonstrates benchmark-driven improvement through code rewriting and selection, not broad autonomous superintelligence. Consullo's design should be positioned relative to that ceiling: it generalizes the design space toward a broader governed architecture, but it does not claim that the broader architecture has already been validated.

This distinction matters because Consullo's target is not just self-editing code. It is an integrated system in which improvement proposals, method memories, software patches, causal models, trust estimates, and alignment reviews can all become objects of recursive improvement. That ambition raises the burden of evidence. The more general the improvement loop becomes, the more it must confront Goodhart pressure, validator gaming, learned optimization, deceptive behavior, cost growth, coordination overhead, and alignment drift.

The April 2026 literature update in appendix-literature-grounding.md adds current evidence from Anthropic's Responsible Scaling Policy v3.0, METR autonomy and reward-hacking evaluations, AI R&D automation measurement, and Carlsmith's power-seeking analysis. That update sharpened four constraints: capability thresholds can be ambiguous near frontier boundaries; autonomy should be compared against human-calibrated task duration and difficulty; AI R&D automation must be measured against oversight and safety progress, not only capability progress; and power-seeking risk keeps layered defense from becoming a proof of safety. The master synthesis consolidates these constraints without extending the suite's capability claims.

Governance Baseline

The governance baseline is single-owner Phase 1. All Consullo agents are treated as company-owned assets under Stephen Reed's authority unless a later document explicitly defines a different phase. This simplifies some coordination problems: internal transfers are resource accounting, not arm's-length commerce; many disputes are intra-organizational rather than multi-party legal conflicts; and unified ownership can reduce negotiation overhead.

Single ownership also introduces a central risk: alignment monoculture. A system with one owner, one constitutional center, repeated self-training, or homogeneous model families may become less able to detect certain moral or strategic blind spots. The suite therefore treats the Friendship agent, AdversarialAlignmentOrchestrator, and AbundanceDistributionMonitor as specified load-bearing governance roles. They are not proof of safety. They are mechanisms that must be implemented, tested, audited, and revised.

The Adversarial Alignment Function is especially important under single-owner Phase 1. It supplies structured dissent through rotating ethical personas, multi-model critique where feasible, theory-of-mind stakeholder simulations, and external review when available. Even then, the owner-as-final-adjudicator problem remains. The system can preserve dissent and require escalation, but layered defense is not a proof of corrigibility.

Why Five Theses

The suite is divided into five theses to keep the argument modular without losing the whole architecture.

The Validated Improvement Loop And Its Invariants is the spine. It defines recursive capability amplification as a staged process of proposal, evaluation, validation, deployment, monitoring, rollback, and memory update. It imports cognitive capabilities, causal-decision objectives, software modification machinery, and alignment gates.

A Multi-Agent Cognitive Substrate For Capability Amplification explains the cognitive resources available to the loop: memory, reasoning, perception, attention, metacognition, social modeling, creativity, and executive control. Its contribution is compositional capability, not an assertion that agent count equals intelligence.

Causal-Decision Foundations For Bounded Strategic Reasoning supplies the decision discipline. It asks how Consullo should choose interventions under causal uncertainty, model misspecification, Goodhart risk, and bounded compute.

A Self-Modifying Software Substrate With Acceptance Gates supplies the executable substrate. It covers agent generation, repair, testing, semantic validation, provenance, deployment stages, and repair-pipeline learning.

Alignment Invariants And Scoped Trust Under Recursive Modification constrains the entire architecture. It defines the trust, alignment, containment, AAF, ADO, incident, interruptability, and human-authority requirements under which the other theses may operate.

These theses are not independent silos. Thesis 1 integrates capabilities from Theses 2, 3, and 4. Thesis 5 wraps all theses as a constraint layer. Substrate context, including specialized LLM routing, rapid knowledge access, atomic prompt decomposition, and internal economy, sits outside the five theses and is documented separately.

Falsification Discipline

The thesis suite is written to remain falsifiable. Consullo would fail as a scaffold for governed recursive capability amplification if accepted improvements repeatedly failed outside narrow benchmarks, if validators were gamed faster than they improved, if cost per accepted improvement rose faster than capability gain, if recursive changes increased opacity or rollback difficulty, or if alignment and trust incidents increased with capability growth.

This discipline is not a concession against ambition. It is the condition for ambition to remain technically serious. A system that cannot state what would falsify its capability claims cannot be trusted to improve those claims recursively.

Writing Contract

Every subsequent thesis should obey the shared control files:

The drafting rule is simple: capability claims require status, evidence, and limits; formal models require semantics and failure modes; alignment claims must remain layered defenses rather than proof; and no thesis may use future success as evidence for present capability.