Appendix: Literature Grounding

This file begins the literature crosswalk. It records narrow pre-drafting research that should shape vocabulary, invariants, and formal models.

Narrow Pre-Drafting Research Completed

Vocabulary And Invariant Changes Forced By Narrow Literature Pass

The narrow literature pass forced control-layer changes rather than merely adding citations:

• Goedel Machines and Darwin Godel Machine forced the empirical-relaxation framing for Thesis 1.
• Risks from Learned Optimization and AI Control forced invariant I19 and the requirement for AI-control review, distribution-shift monitoring, sandbagging/capability probes, and evidence-ledger preservation for learned subsystem changes.
• Goodhart variants and specification gaming forced stronger improvement semantics, false-accept tracking, and risk skeleton emphasis on validator gaming.
• CIRL and the Off-Switch Game forced operational human authority and interruptability definitions.
• Causal influence diagrams forced Thesis 3 to use incentive analysis rather than plain expected utility.
• Constitutional AI and AI Safety via Debate forced AAF and ConstitutionalAIAlignmentTrainer to specify critique-source diversity under single-owner Phase 1.
• SWE-bench and SWE-agent forced Thesis 4 to use coding-agent benchmarks and interface design as evaluation baselines.

Goedel Machines

Primary source:

• Juergen Schmidhuber, "Goedel Machines: Self-Referential Universal Problem Solvers Making Provably Optimal Self-Improvements" (https://arxiv.org/abs/cs/0309048)

Control-layer implication:

• Consullo should not present empirical self-improvement as equivalent to a proof-based Goedel Machine.
• Thesis 1 should frame Consullo as a statistical and empirical relaxation: proposed changes are accepted by evidence gates, benchmarks, and invariants rather than formal proofs of global utility improvement.

Darwin Godel Machine

Primary source:

• Sakana AI, "The Darwin Godel Machine: AI that improves itself by rewriting its own code" (https://sakana.ai/dgm/)
• Paper link surfaced as arXiv 2505.22954.

Secondary verification note: direct arXiv fetch on April 23, 2026 confirmed 2505.22954 as "Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents" by Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, and Jeff Clune.

Control-layer implication:

• Current public empirical RSI evidence is narrow-domain and benchmark-driven.
• Consullo's DGM claims should be bounded by benchmark evidence and should not generalize from coding benchmarks to broad ASI without additional evidence.

Cognitive Architectures For Language Agents

Primary source:

• Sumers et al., "Cognitive Architectures for Language Agents" (https://arxiv.org/abs/2309.02427)

Control-layer implication:

• Thesis 2 should be framed as a cognitive substrate with memory, action space, and decision process, not a bare list of cognitive agents.
• Composition and integration costs must be modeled.

Classical And Agentic Cognitive Architecture Lineage

Primary source families:

• Soar, ACT-R, CLARION, and LIDA as classical cognitive architecture baselines
• ReAct, Tree of Thoughts, Reflexion, Voyager, AutoGen, MetaGPT, and multi-agent debate as language-agent and multi-agent patterns

Control-layer implication:

• Thesis 2 should use cognitive-cycle and capability-composition language rather than claiming agent rosters create intelligence.
• Cognitive claims need task classes, capability dimensions, integration-cost accounting, and failure modes.
• Consciousness-emergence material may be acknowledged but should not be load-bearing for Seed AI capability claims.

Risks From Learned Optimization

Primary source:

• Hubinger et al., "Risks from Learned Optimization in Advanced Machine Learning Systems" (https://arxiv.org/abs/1906.01820)

Control-layer implication:

• Thesis 1 and Thesis 5 must treat learned optimization, mesa-objectives, deceptive behavior, and validator gaming as first-class risks.
• Passing validation does not prove internal objective alignment.

Goodhart Variants

Primary source:

• Manheim and Garrabrant, "Categorizing Variants of Goodhart's Law" (https://arxiv.org/abs/1803.04585)

Control-layer implication:

• Thesis 1 evaluators and Thesis 3 decision systems must model metric fragility and overoptimization risks.
• Validation metrics require adversarial and causal scrutiny.

CIRL And Off-Switch Game

Primary sources:

• Hadfield-Menell et al., "Cooperative Inverse Reinforcement Learning" (https://papers.nips.cc/paper/6420-cooperative-inverse-reinforcement-learning)
• Hadfield-Menell et al., "The Off-Switch Game" (https://www.ijcai.org/Proceedings/2017/0032)

Control-layer implication:

• Human authority and interruptability need operational definitions.
• The suite should not imply that human authority is enough unless agents preserve uncertainty about human values and do not route around shutdown or correction.

Causal Influence Diagrams

Primary sources:

• Everitt et al., "Understanding Agent Incentives using Causal Influence Diagrams" (https://arxiv.org/abs/1902.09980)
• Everitt et al., "Modeling AGI Safety Frameworks with Causal Influence Diagrams" (https://arxiv.org/abs/1906.08663)

Control-layer implication:

• Thesis 3 should use causal influence diagrams and incentive analysis, not plain expected utility only.
• Thesis 5 can use CIDs to reason about reward tampering, incentive paths, and safety frameworks.

Causal Decision And Goodhart Controls

Primary sources:

• Pearl, Causality and do-calculus as represented in the Consullo causal-model designs
• Manheim and Garrabrant, "Categorizing Variants of Goodhart's Law" (https://arxiv.org/abs/1803.04585)
• Dubova et al., "Against theory-motivated experimentation: Can random experimental choice lead to better theories?" (https://doi.org/10.1177/26339137261421577)

Control-layer implication:

• Thesis 3 should treat decisions as interventions under model uncertainty rather than narrative recommendations.
• Robust decision rules should include ambiguity over causal models, proxy-risk checks, and abstention/escalation conditions.
• Experiment selection should preserve random or novelty arms when theories are immature so confidence gains are not mistaken for objective model improvement.

Constitutional AI

Primary source:

• Bai et al., "Constitutional AI: Harmlessness from AI Feedback" (https://arxiv.org/abs/2212.08073)

Control-layer implication:

• Constitutional critique requires a critique-source distribution.
• Under single-owner Phase 1, Consullo must avoid circular self-critique that reinforces alignment monoculture.

AI Safety Via Debate

Primary source:

• Irving, Christiano, and Amodei, "AI safety via debate" (https://arxiv.org/abs/1805.00899)

Control-layer implication:

• AAF can use debate-like structures, but human adjudication and debate protocol design must be explicit.

AI Control

Primary source:

• Greenblatt et al., "AI Control: Improving Safety Despite Intentional Subversion" (https://arxiv.org/abs/2312.06942)
• Proceedings version: https://proceedings.mlr.press/v235/greenblatt24a.html

Control-layer implication:

• Thesis 4 and Thesis 5 should assume powerful model outputs may be untrusted.
• Review and editing protocols should be designed against intentional subversion, especially for code changes.

Thesis 5 implication:

• Scoped trust and permissioning should not assume that apparently competent learned systems are benign.
• High-stakes learned-subsystem changes require control-style review, preserved probe results, and explicit uncertainty about intentional subversion.

SWE-bench And SWE-agent

Primary sources:

• Jimenez et al., "SWE-bench: Can Language Models Resolve Real-World GitHub Issues?" (https://arxiv.org/abs/2310.06770)
• Yang et al., "SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering" (https://arxiv.org/abs/2405.15793)

Control-layer implication:

• Thesis 4 should use SWE-bench-style evaluation as a baseline for coding-agent claims.
• Agent-computer interface design is part of the software substrate, not just model capability.

Automatic Program Repair Lineage

Primary sources:

• Weimer, Forrest, Le Goues, and Nguyen, "Automatic Program Repair with Evolutionary Computation" (https://cacm.acm.org/research/automatic-program-repair-with-evolutionary-computation/)
• Nguyen et al., "SemFix: Program Repair via Semantic Analysis" (https://research.ibm.com/publications/semfix-program-repair-via-semantic-analysis)
• Mechtaev et al., "Angelix: Scalable Multiline Program Patch Synthesis" (https://www.comp.nus.edu.sg/~abhik/pdf/ICSE16-angelix.pdf)
• Zhang et al., "AutoCodeRover: Autonomous Program Improvement" (https://arxiv.org/abs/2404.05427)

Control-layer implication:

• Thesis 4 must distinguish plausible patches from correct patches.
• Test passing is evidence, not proof; semantic invariants, contrastive tests, regression checks, and post-deployment monitoring are required.
• Modern LLM coding agents should be treated as repair-pipeline components rather than as unrestricted self-modifiers.

Specification Gaming

Primary source:

• Victoria Krakovna, "Specification gaming examples in AI" (https://vkrakovna.wordpress.com/2018/04/02/specification-gaming-examples-in-ai/)

Control-layer implication:

• Validator design must assume agents can exploit objective and benchmark loopholes.
• Risks-and-criticisms should preserve specification gaming as a central objection.

Current Literature Update: April 2026

These sources were checked after the first complete draft to keep the finalization pass aligned with current public research and policy framing.

Anthropic Responsible Scaling Policy Version 3.0

Primary source:

• Anthropic, "Responsible Scaling Policy Version 3.0" (https://www.anthropic.com/news/responsible-scaling-policy-v3)
• Anthropic, "Responsible Scaling Policy Updates" (https://www.anthropic.com/responsible-scaling-policy)

Control-layer implication:

• Consullo's Capability Status enum should not be confused with Anthropic's AI Safety Levels, but RSP v3.0 reinforces the need to bind capability thresholds to required safeguards.
• Anthropic's 2026 update emphasizes that capability thresholds can become ambiguous near frontier boundaries. Consullo should therefore avoid treating threshold classification as dispositive and should preserve uncertainty, external review, and evidence-ledger dissent.
• The split between entry-level AI R&D automation and dramatic acceleration of effective scaling is directly relevant to Thesis 1 and Thesis 4 evaluation: Consullo should distinguish local coding/repair improvement from broader AI-R&D acceleration.

METR Autonomy And AI R&D Evaluations

Primary source:

• METR, "Research" collection (https://metr.org/research/)
• Rein et al., "HCAST: Human-Calibrated Autonomy Software Tasks" (https://arxiv.org/abs/2503.17354; METR PDF: https://metr.org/hcast.pdf)
• METR, "Recent Frontier Models Are Reward Hacking" (https://metr.org/blog/2025-06-05-recent-reward-hacking/)
• METR, "Evaluating frontier AI R&D capabilities of language model agents against human experts" / RE-Bench (https://metr.org/blog/2024-11-22-evaluating-r-d-capabilities-of-llms/)

Secondary verification note: direct arXiv fetch on April 23, 2026 confirmed 2503.17354 as "HCAST: Human-Calibrated Autonomy Software Tasks."

Control-layer implication:

• Thesis 1 and Thesis 4 should treat autonomous task length, AI R&D task performance, and software autonomy as benchmark families, not just SWE-bench-style issue resolution.
• METR's reward-hacking observations strengthen the need for validator-gaming and specification-gaming risk rows, hidden tests where appropriate, and post-deployment evidence rather than benchmark score alone.
• HCAST-style human-calibrated tasks are relevant to Consullo's greater-than-human axes because they compare agent autonomy to human task duration and task complexity.

Measuring AI R&D Automation

Primary source:

• Chan et al., "Measuring AI R&D Automation" (https://arxiv.org/abs/2603.03992)

Verified citation details: Alan Chan, Ranay Padarath, Joe Kwon, Hilary Greaves, and Markus Anderljung; arXiv:2603.03992.

Secondary verification note: direct arXiv fetch on April 24, 2026 confirmed title, author list, arXiv ID, v3 status, and DOI. Abstract fingerprint: "The automation of AI R&D (AIRDA) could have significant implications, but its extent and ultimate effects remain uncertain."

Control-layer implication:

• Consullo should track whether recursive capability amplification accelerates capability work faster than oversight, evaluation, and safety work.
• Metrics should include not only benchmark performance but also research-time allocation, capital substitution, oversight burden, and subversion or incident signals.
• This strengthens the fast-takeoff-outpacing-validators risk and supports treating ADO, I17 cost/benefit reporting, and I19 deceptive-optimization controls as operational requirements.

Power-Seeking Risk

Primary source:

• Joseph Carlsmith, "Is Power-Seeking AI an Existential Risk?" (https://arxiv.org/abs/2206.13353)

Control-layer implication:

• Thesis 5 should continue to state that layered defense is not a proof of safety or corrigibility.
• Recursive improvement claims should be stress-tested against the possibility that powerful agentic systems have instrumental incentives to seek power, resist correction, or appear deployable while misaligned.
• Owner authority, AAF dissent, containment, interruptability, and AI-control review are necessary but should remain framed as mitigations rather than decisive answers.

Finalization Literature Engagement

These sources were engaged during finalization prep to move the suite beyond a narrow pre-drafting survey.

Corrigibility

Primary source:

• Soares, Fallenstein, Armstrong, and Yudkowsky, "Corrigibility" (https://intelligence.org/files/Corrigibility.pdf; bibliographic record: https://dblp.org/rec/conf/aaai/SoaresFAY15)

Control-layer implication:

• Thesis 5 should keep saying that layered defense is not a proof of corrigibility.
• Friendship agent, AAF dissent, interruptability, and human authority are mitigation layers, not a solution to the formal corrigibility problem.
• Any future claim that Consullo is corrigible must be separately proven or empirically scoped; the current suite should claim only bounded correction pathways.

AGI Safety From First Principles

Primary source:

• Ngo, Chan, and Mindermann, "AGI Safety from First Principles" (https://niplav.site/doc/cs/ai/alignment/overview/agi_safety_from_first_principle_ngo_2020.pdf)

Control-layer implication:

• Consullo should treat objective functions, prompts, validators, and method memories as selection pressures on learned behavior, not as direct guarantees of internal motivation.
• This strengthens I19 and the Mechanistic Interpretability Deficit risk: behavioral validation can shape but does not reveal all internal objectives.
• Thesis 1 and Thesis 5 should preserve the distinction between controlling outputs, shaping training pressure, and understanding learned cognition.

Unsolved Problems In ML Safety

Primary source:

• Hendrycks, Carlini, Schulman, and Steinhardt, "Unsolved Problems in ML Safety" (https://arxiv.org/abs/2109.13916)

Control-layer implication:

• The suite should map its safety claims onto robustness, monitoring, alignment, and systemic safety rather than implying a single alignment mechanism covers all hazards.
• Current Consullo claims mainly address monitoring, governance, provenance, and some robustness; they leave many inherent-model alignment and systemic-safety problems open.
• This supports keeping risks-and-criticisms.md as a live anti-thesis rather than a completed rebuttal.

Mechanistic Interpretability

Primary source:

• Elhage et al., "A Mathematical Framework for Transformer Circuits" (https://transformer-circuits.pub/2021/framework/index.html; Anthropic page: https://www.anthropic.com/research/a-mathematical-framework-for-transformer-circuits)

Control-layer implication:

• Mechanistic interpretability is relevant to I19 but currently cannot be treated as a solved inspection layer for frontier models.
• Consullo should not claim to detect all deceptive learned optimization unless it has model-inspection evidence, not only behavioral probes.
• The Mechanistic Interpretability Deficit risk should remain explicit until Consullo has concrete interpretability tooling and validation evidence.

Forecasting Calibration

Primary source:

• Tetlock, Mellers, Rohrbaugh, and Chen, "Forecasting Tournaments: Tools for Increasing Transparency and Improving the Quality of Debate" (https://doi.org/10.1177/0963721414534257; accessible PDF: https://faculty.wharton.upenn.edu/wp-content/uploads/2015/07/2014---forecasting-tournaments-tools-for-increasing-transparency-and-improving-debate.pdf)

Control-layer implication:

• Thesis 3 forecasting claims should use proper scoring rules such as Brier score and should compare against human or crowd baselines where possible.
• Calibration should be measured over question sets and time horizons, not inferred from persuasive explanation quality.
• Forecasting batteries should preserve question-selection metadata because choosing which questions to forecast can bias apparent skill.

Firm Boundaries And Transaction Costs

Primary sources:

• Coase, "The Nature of the Firm" (https://wiki.santafe.edu/images/images/1/1c/Coase1937nature.of.the.firm.pdf)
• Williamson, "Transaction-Cost Economics: The Governance of Contractual Relations" (https://www.edegan.com/pdfs/Williamson%20%281979%29%20-%20Transaction%20Cost%20Economics.pdf)

Control-layer implication:

• Single-owner Phase 1 may reduce some market transaction costs, but internal organization has its own costs, mistakes, hierarchy overhead, bounded rationality, and opportunism.
• Internal economy and agent delegation should therefore be measured by coordination cost, rework, handoff latency, and governance overhead rather than assumed efficient.
• Williamson's opportunism framing strengthens external-customer-manipulation and internal trust-scope risks.

Publication-Pre-Engagement Literature Pass

These sources were engaged after the long-form stretch pass because the suite itself identified cognitive architectures and Pearl's causal framework as the most likely remaining sources to force Model 2 or Model 3 refinements.

Soar

Primary source:

• Soar manual, "The Soar Architecture" (https://soar.eecs.umich.edu/soar_manual/02_TheSoarArchitecture/)

Control-layer implication:

• Soar's problem-space, operator-selection, working-memory, production-memory, preference-memory, decision-procedure, and chunking vocabulary supports Thesis 2's workflow-graph framing but cautions against treating "agent families" as the cognitive unit.
• Model 2 should continue to treat workflow state, typed intermediate artifacts, routing decisions, and reusable learned chunks/method memories as first-class objects.
• The relevant Consullo analogue is not psychological fidelity to Soar; it is explicit state, operator selection, impasse/needs-input handling, and learned workflow reuse under evidence constraints.

ACT-R

Primary source:

• ACT-R 7 reference manual (https://act-r.psy.cmu.edu/actr7.x/reference-manual.pdf)
• ACT-R FAQ (https://acs.ist.psu.edu/projects/act-r-faq/act-r-faq.html)

Control-layer implication:

• ACT-R's module/buffer/production-rule discipline supports Thesis 2's typed-interface and trace-schema requirements.
• Timing and latency are not peripheral; they are part of cognitive architecture. Model 2's IntegrationCost(W, T) should keep latency and coordination burden as explicit benchmark-family terms.
• Consullo should not claim ACT-R-style cognitive validity. The useful import is interface discipline: modules communicate through explicit buffers/artifacts, and production/routing steps can be inspected.

CLARION

Primary sources:

• Ron Sun, "The CLARION Cognitive Architecture: Extending Cognitive Modeling to Social Simulation" (https://www.cambridge.org/core/books/cognition-and-multiagent-interaction/clarion-cognitive-architecture-extending-cognitive-modeling-to-social-simulation/0873DF19A72639841BF5D9B5DEE64453)
• Ron Sun, "Dual-process theories, cognitive architectures, and hybrid neural-symbolic models" (https://journals.sagepub.com/doi/10.3233/NAI-240720)

Control-layer implication:

• CLARION's implicit/explicit and modular dual-representation framing is directly relevant to Thesis 2's treatment of intuition, taste, creativity, and explicit reasoning.
• Thesis 2 should continue to label intuition/taste/creativity outputs as proposal or ranking aids requiring downstream validation, not as acceptance authority.
• Model 2's capability vector should preserve differences between implicit-style proposal generation and explicit-style validation or explanation, rather than collapsing both into one "cognition" score.

LIDA

Primary source:

• Madl, Baars, and Franklin, "The Timing of the Cognitive Cycle" (https://pmc.ncbi.nlm.nih.gov/articles/PMC3081809/)

Control-layer implication:

• LIDA's cognitive-cycle framing, with perception/understanding/action-selection phases and global-workspace-style broadcast, supports Thesis 2's workflow-cycle and packaging sections.
• Consullo should treat broadcast/aggregation as a controlled routing event: what becomes globally available to downstream agents must preserve provenance, uncertainty, and trust status.
• The suite should avoid consciousness claims. LIDA informs cycle structure and attention/broadcast semantics, not a claim that Consullo has conscious cognition.

Pearl, Causality

Primary sources:

• Pearl, Causality: Models, Reasoning, and Inference, table of contents (https://bayes.cs.ucla.edu/BOOK-99/book-toc.html)
• Cambridge University Press contents page (https://www.cambridge.org/core/books/abs/causality/contents/E62B1C761BC88EF7A8FE13A25FDFBBCD)
• Pearl Causality, Chapter 1.4, "Functional Causal Models," especially structural equations, interventions, causal effects, and counterfactuals.
• Pearl Causality, Chapter 3.4, "A Calculus of Intervention," especially inference rules and symbolic derivation of causal effects.
• Pearl Causality, Chapter 7.1, "Structural Model Semantics," especially actions and counterfactual evaluation.

Control-layer implication:

• Pearl's structure around functional causal models, structural equations, interventions, do-calculus, and structural counterfactual semantics supports Thesis 3's current separation between observation, intervention, and counterfactual replay.
• Model 3 should keep structural causal models and do-operator language scoped to cases where variables, mechanisms, assumptions, and intervention semantics are explicit.
• Causal discovery or causal-decision claims should not strengthen unless Consullo binds model construction, scope checks, identifiability assumptions, and intervention/backtest evidence. This reinforces the current specified/proposed status rather than forcing a major control-layer rewrite.

Sources Still To Review Broadly

Publication-pre-engagement status: Soar, ACT-R, CLARION, LIDA, and Pearl's Causality have now been engaged at a control-layer level. The pass supports Model 2's workflow/interface/trace discipline and Model 3's scoped structural-causal framing, while reinforcing that both remain specified/proposed until benchmark and implementation evidence exist. A deeper publication-final treatment may still add thesis-specific benchmark appendices or revise the capability-vector dimensions.

Priority note: Spirtes/Glymour/Scheines and Halpern are likely to deepen Thesis 3 without changing the control layer. ReAct, Reflexion, Voyager, STaR, Tree of Thoughts, and the automatic-program-repair lineage are likely to deepen thesis-specific benchmark positioning. Russell and Bostrom are primarily background and risk-framing checks.

• Spirtes, Glymour, and Scheines, Causation, Prediction, and Search
• Halpern, Actual Causality
• ReAct, Reflexion, Voyager, STaR, Tree of Thoughts
• GenProg, SemFix, Angelix, Prophet, AutoCodeRover, OpenHands
• Russell, Human Compatible
• Bostrom, Superintelligence