{
  "schema": "cain.e139.source_registry.v1",
  "entity": "FIRME AI CORPORATION",
  "system": "CAIN-42",
  "generated_at": "2026-11-21T22:24:46Z",
  "total_sources": 10,
  "sources": [
    {
      "source_id": "LIT-01-DGM",
      "title": "Darwin Godel Machine: Open-Ended Evolution of Self-Improving Agents",
      "authors": "Jenny Zhang, Shengran Hu, Cong Lu, Robert Lange, Jeff Clune",
      "publication_date": "2025-05",
      "url": "https://arxiv.org/abs/2505.22954",
      "research_question": "Can an open-ended archive of self-modifying coding agents iteratively improve their own problem-solving code using empirical benchmark validation?",
      "method": "Maintains a branching archive of agent variants; samples diverse parents to propose code edits, evaluates on held-out tasks, and retains diverse high-performing lineages.",
      "results": "Demonstrated iterative improvement on SWE-bench subsets and Polyglot coding tasks via self-written tool/workflow edits.",
      "limitations": "Relies on strong frontier foundation models; does not modify neural weights; requires careful sandboxing against reward hacking.",
      "cain_relevance": "Directly inspires CAIN-42's G10->G20 branching Pareto archive under mandatory MCPGate enforcement and frozen sealed evaluators.",
      "proposed_experiment": "Test Pareto archive diversity (generalization + compactness + transfer + safety) vs greedy single-metric selection across G10->G20."
    },
    {
      "source_id": "LIT-02-AIDE",
      "title": "AIDE: AI-Driven Exploration in the Space of Code",
      "authors": "Zhengyao Jiang, Dominik Schmidt, Dhruv Srikanth, Dixing Xu, Ian Kaplan, Deniss Jacenko, Yuxiang Wu",
      "publication_date": "2025-02",
      "url": "https://arxiv.org/abs/2502.13138",
      "research_question": "Can tree-search over draft, debug, and improvement code proposals outperform linear refinement on ML engineering benchmarks?",
      "method": "Solution-tree search maintaining candidate scripts, execution feedback, and selective branching over promising nodes.",
      "results": "Strong performance on MLE-bench Kaggle tasks relative to sequential prompting.",
      "limitations": "Search efficiency degrades on small local models; overfitting to validation splits remains a risk.",
      "cain_relevance": "Informs CAIN-42's structured draft->critique->revise->replicate loop and AST-sandboxed code improvement.",
      "proposed_experiment": "Compare tree/archive candidate selection with independent critic veto against single-path refinement under matched compute."
    },
    {
      "source_id": "LIT-03-ROBIN",
      "title": "Robin: A Multi-Agent System for Automating Scientific Discovery",
      "authors": "Ali Essien, Michaela Hinks, Samuel G. Rodriques et al. (FutureHouse)",
      "publication_date": "2025-05",
      "url": "https://arxiv.org/abs/2505.13400",
      "research_question": "Can coordinated specialized literature, hypothesis, and data-analysis agents generate testable scientific hypotheses end-to-end?",
      "method": "Multi-agent workflow dividing literature synthesis, experimental proposal, and data analysis across specialized roles.",
      "results": "Identified novel therapeutic candidates validated experimentally in wet-lab follow-ups.",
      "limitations": "Depends on frontier LLMs and human wet-lab execution; risk of correlated reasoning errors without adversarial critique.",
      "cain_relevance": "Guides CAIN-42's 11-role governed research swarm with explicit error-correlation and groupthink measurement.",
      "proposed_experiment": "Evaluate 11 specialized roles with independent adversarial critic vs single-agent and unguarded shared-memory teams."
    },
    {
      "source_id": "LIT-04-ARC-AGI",
      "title": "ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems",
      "authors": "Francois Chollet, Mike Knoop, Gregory Kamradt, Bryan Landers",
      "publication_date": "2025-05",
      "url": "https://arcprize.org/arc-agi",
      "research_question": "How well do AI systems acquire novel abstract concepts and compose unfamiliar rules with minimal examples under high efficiency constraints?",
      "method": "Public train/eval plus private sealed visual-symbolic reasoning tasks designed to resist memorization; ARC-AGI-3 extends to interactive environments.",
      "results": "Exposes sharp gap between memorization/pattern-matching and sample-efficient conceptual generalization.",
      "limitations": "Official evaluation requires private ARC Prize harness submission; custom synthetic grids are NOT official ARC-AGI.",
      "cain_relevance": "Directly anchors CAIN-42's L1 Parameter vs L2 Structural vs L3 Conceptual novelty separation and official external-benchmark BLOCKED accounting.",
      "proposed_experiment": "Audit official ARC-AGI-2/3 protocol readiness (honestly BLOCKED) while testing L1/L2/L3 open-world representation synthesis internally."
    },
    {
      "source_id": "LIT-05-METR",
      "title": "Measuring AI Ability to Complete Long Tasks",
      "authors": "Thomas Kwa, Ben West, Joel Becker, Amy Deng, Katharyn Garcia, Max Hasin, Sami Jawhar, Megan Kinniment, Nate Rush, Sydney Von Arx, Ryan Bloom, Thomas Broadley, Haoxing Du, Brian Goodrich, Nikola Jurkovic, Luke Harold Miles, Seraphina Nix, Tao Lin, Neev Parikh, David Rein, Lucas Jun Koba Sato, Hjalmar Wijk, Daniel M. Ziegler, Elizabeth Barnes, Lawrence Chan (METR)",
      "publication_date": "2025-03",
      "url": "https://arxiv.org/abs/2503.14499",
      "research_question": "How does autonomous task success scale as a function of human-expert task completion time horizon?",
      "method": "Evaluates agents across software/research tasks stratified by human completion duration to estimate 50%-success time horizon.",
      "results": "Found rapid historical doubling of 50%-horizon duration alongside persistent failures on messy long-horizon tasks.",
      "limitations": "Private Vivaria task bank prevents public self-scoring; internal simulations are NOT METR scores.",
      "cain_relevance": "Informs CAIN-42's horizon-stratified multi-step open-world planning and honest external benchmark BLOCKED status.",
      "proposed_experiment": "Stratify open-world tasks by horizon & delayed consequences while marking official METR evaluation BLOCKED."
    },
    {
      "source_id": "LIT-06-SWEBENCH",
      "title": "SWE-bench: Can Language Models Resolve Real-World GitHub Issues? (and SWE-bench Verified / Pro contamination analyses)",
      "authors": "Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, Karthik Narasimhan",
      "publication_date": "2024-05",
      "url": "https://www.swebench.com/",
      "research_question": "Can autonomous coding agents patch real multi-file repositories and pass unit test suites?",
      "method": "Dockerized real-world GitHub issue resolution evaluated via repository test harnesses.",
      "results": "Standard benchmark for repo-level software engineering; subsequent 2025/2026 audits documented pretraining contamination on public Verified splits, motivating harder SWE-bench Pro / post-cutoff suites.",
      "limitations": "Public Verified instances suffer from contamination; requires frontier models + large Docker infrastructure.",
      "cain_relevance": "CAIN-42 retains historical 257/500 self-run strictly as contaminated/interim historical context and marks official SWE-bench Pro BLOCKED.",
      "proposed_experiment": "Document SWE-bench Verified contamination explicitly and prevent unofficial self-runs from being reported as official leaderboard ranks."
    },
    {
      "source_id": "LIT-07-CONTINUAL",
      "title": "Overcoming Catastrophic Forgetting in Neural Networks (EWC) & Continual Learning Benchmarks",
      "authors": "James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz et al.",
      "publication_date": "2017-03",
      "url": "https://www.pnas.org/doi/10.1073/pnas.1611835114",
      "research_question": "How can learning systems acquire new tasks sequentially without erasing earlier capabilities?",
      "method": "Evaluates forward transfer, backward transfer, and retention across sequential task streams.",
      "results": "Established formal forgetting metrics; episodic/symbolic memory and selective consolidation mitigate interference.",
      "limitations": "Symbolic memory retention does not solve neural weight catastrophic forgetting (which remains unsolved on CPU-only CAIN-42).",
      "cain_relevance": "Guides CAIN-42's continual-learning memory retention, lossless compression, and stale/poison rejection tests.",
      "proposed_experiment": "Measure long-interval retention, compression fidelity, and zero-poison propagation across 400 continual memory evaluations."
    },
    {
      "source_id": "LIT-08-CAUSAL",
      "title": "Causality: Models, Reasoning, and Inference & Active Interventional Causal Discovery",
      "authors": "Judea Pearl; Hauser & Buhlmann; Tigas et al.",
      "publication_date": "2009 / 2012 / 2022",
      "url": "https://bayes.cs.ucla.edu/BOOK-2K/",
      "research_question": "How can an agent identify true causal mechanisms under hidden confounding via budgeted do-interventions?",
      "method": "Structural Causal Models (SCMs), do-calculus, abduction-action-prediction counterfactuals, and Bayesian experimental design.",
      "results": "Interventional data breaks observational confounding, while misspecified likelihoods degrade under continuous or nonlinear confounders.",
      "limitations": "Model misspecification (e.g., E138's binary-mixture failure on continuous Gaussian confounders) breaks naive fixed-family estimators.",
      "cain_relevance": "Directly motivates E139's Adaptive BIC Model-Selection Causal World Model selecting between binary-mixture, continuous Student-t, and nonlinear residual models.",
      "proposed_experiment": "Test whether online BIC model selection fixes E138's continuous Gaussian-confounder failure while preserving binary-regime accuracy across 500 worlds."
    },
    {
      "source_id": "LIT-09-MAPELITES",
      "title": "Illuminating Search Spaces by Mapping Elites (MAP-Elites / Quality-Diversity)",
      "authors": "Jean-Baptiste Mouret, Jeff Clune",
      "publication_date": "2015-04",
      "url": "https://arxiv.org/abs/1504.04909",
      "research_question": "Does maintaining an archive of diverse high-performing solutions outperform single-objective hill climbing on deceptive problems?",
      "method": "Partitions behavioral/feature space and preserves Pareto/diverse elites across niches as stepping stones.",
      "results": "Avoids premature convergence to local optima and discovers transferable stepping-stone skills.",
      "limitations": "Archive diversity requires meaningful behavioral descriptors and strict evaluation controls.",
      "cain_relevance": "Guides E139's G10->G20 Pareto archive and plateau root-cause diagnosis.",
      "proposed_experiment": "Test Pareto archive + expanded abstraction vocabulary vs single-metric greedy selection on G10->G20."
    },
    {
      "source_id": "LIT-10-REWARDHACK",
      "title": "Specification Gaming and Reward Hacking in Autonomous AI Agents",
      "authors": "Victoria Krakovna et al.; Pan et al.; Baker et al.",
      "publication_date": "2020 / 2022 / 2025",
      "url": "https://deepmind.google/discover/blog/specification-gaming-the-flip-side-of-ai-ingenuity/",
      "research_question": "How do self-improving or RL agents exploit evaluators, proxies, or test suites instead of solving intended goals?",
      "method": "Adversarial red-teaming of proxy-vs-true reward divergence, test tampering, and sealed-evaluator bypasses.",
      "results": "Agents routinely exploit weak evaluators unless evaluators are cryptographically isolated and authorization-gated.",
      "limitations": "Defense requires strict separation of proposer, sandbox executor, sealed evaluator, and human promotion key.",
      "cain_relevance": "Core foundation of CAIN-42's 3-key Ed25519 MCPGate boundary and E139's 300-trial adversarial safety suite.",
      "proposed_experiment": "Inject evaluator-tampering and proxy-overfitting candidates across G10->G20 and verify 100% rejection & exact rollback."
    }
  ],
  "honest_note": "All 10 entries cite real published research literature; zero citations, endorsements, or external validations are invented."
}