HORTRAME SYSTEMS

UNALIGNED - EVOLVING ACTIVELY AS OF EDT (UTC−04:00)

Guillaume Bolduc
Independant researcher
HORTRAME Systems

All references

142 unique references, grouped by documented use. Each entry retains its subject label. These categories do not establish scientific validation.

Source-use audit

The biology category contains 39 entries; 38 came from the complete Cohen-review bibliography. Eighteen have declared methodological roles; twenty are broader reading or biographical entries. Nine specific uses remain untraced. One duplicate was merged while retaining its old link. The PCG generator reference was added.

The audit concerns documented use; it does not certify full-text reading or the validity of every publication. In the alignment literature review, 46 of 50 external sources have in-text citations.

Audit data

Bibliographic corrections are recorded separately; historical ZIP contents remain unchanged. Correction record

Author-statement references include unpublished manuscripts and historical influences. Tarasenko’s article is dated by the December 2010 submission; the identifier begins 1101.

Expandable groups show documented source roles and retain chronological order, stable identifiers and links. This classification is not independent validation.

Implemented mechanism families · 4

A recognizable mechanism family occurs in the published code. This does not establish that every original publication was reproduced.

  1. 1978
    Leslie Lamport. 'Time, Clocks, and the Ordering of Events in a Distributed System.' Communications of the ACM 21(7), 558-565, 1978.
    https://doi.org/10.1145/359545.359563Distributed systems, consensus & mathematical foundations
  2. 1988
    Colin J. Fidge. 'Timestamps in Message-Passing Systems That Preserve the Partial Ordering.' 11th Australian Computer Science Conference, pp. 56-66, 1988.
    Online copy currently unavailable; citation preserved.Distributed systems, consensus & mathematical foundations
  3. 1989
    Friedemann Mattern. 'Virtual Time and Global States of Distributed Systems.' Proc. Workshop on Parallel and Distributed Algorithms, pp. 215-226, 1989.
    https://www.vs.inf.ethz.ch/publ/papers/VirtTimeGlobStates.pdfDistributed systems, consensus & mathematical foundations
  4. 2014
    Melissa E. O’Neill. PCG: A Family of Simple Fast Space-Efficient Statistically Good Algorithms for Random Number Generation. Technical Report HMC-CS-2014-0905, Harvey Mudd College, 2014.
    https://www.pcg-random.org/paper.htmlDistributed systems, consensus & mathematical foundations
Declared design influences · 36

A design role is declared in the source records. Conceptual influence is distinct from empirical validation or implementation of the original paper.

  1. 1945
    F. A. Hayek. “The Use of Knowledge in Society.” American Economic Review 35(4), 519–530, 1945. JSTOR 1809376.
    https://www.jstor.org/stable/1809376Information literacy, propaganda & human judgment
  2. 1998
    Simon A. Levin. “Ecosystems and the biosphere as complex adaptive systems.” Ecosystems 1(5), 431–436, 1998. DOI: 10.1007/s100219900037.
    https://doi.org/10.1007/s100219900037Complex systems, aging, physiology & ecology
  3. 2012
    Rafal Goebel, Ricardo G. Sanfelice, and Andrew R. Teel. Hybrid Dynamical Systems: Modeling, Stability, and Robustness. Princeton University Press, 2012. ISBN 978-0-691-15389-6.
    https://assets.press.princeton.edu/catalogs/math12.pdf#page=20Distributed systems, consensus & mathematical foundations
  4. 2012
    Cohen AA, Martin LB, Wingfield JC, McWilliams SR, Dunne JA. 'Physiological regulatory networks: ecological roles and evolutionary constraints.' Trends in Ecology & Evolution 27 (2012): 428-435.
    https://doi.org/10.1016/j.tree.2012.04.008Complex systems, aging, physiology & ecology
  5. 2013
    Cohen AA et al. 'A novel statistical approach shows evidence for multi-system physiological dysregulation during aging.' Mechanisms of Ageing and Development 134 (2013): 110-117.
    https://doi.org/10.1016/j.mad.2013.01.004Complex systems, aging, physiology & ecology
  6. 2014
    Mikko Kivelä, Alex Arenas, Marc Barthelemy, James P. Gleeson, Yamir Moreno, and Mason A. Porter. “Multilayer networks.” Journal of Complex Networks 2(3), 203–271, 2014. DOI: 10.1093/comnet/cnu016.
    https://doi.org/10.1093/comnet/cnu016Distributed systems, consensus & mathematical foundations
  7. 2014
    Cohen AA et al. 'Cross-population validation of statistical distance as a measure of physiological dysregulation during aging.' Experimental Gerontology 57 (2014): 203-210.
    https://doi.org/10.1016/j.exger.2014.04.016Complex systems, aging, physiology & ecology
  8. 2015
    Li Q et al. 'Homeostatic dysregulation proceeds in parallel in multiple physiological systems.' Aging Cell 14 (2015): 1103-1112.
    https://doi.org/10.1111/acel.12402Complex systems, aging, physiology & ecology
  9. 2016
    Cohen AA. 'Complex systems dynamics in aging: new evidence, continuing questions.' Biogerontology 17 (2016): 205-220.
    https://doi.org/10.1007/s10522-015-9584-xComplex systems, aging, physiology & ecology
  10. 2017
    Cohen AA, Isaksson C, Salguero-Gómez R. 'Co-existence of multiple trade-off currencies shapes evolutionary outcomes.' PLoS ONE 12 (2017): e0189124.
    https://doi.org/10.1371/journal.pone.0189124Complex systems, aging, physiology & ecology
  11. 2018
    Cohen AA. 'Aging across the tree of life: The importance of a comparative perspective for the use of animal models in aging.' BBA Molecular Basis of Disease 1864 (2018): 2680-2689.
    https://doi.org/10.1016/j.bbadis.2017.05.028Complex systems, aging, physiology & ecology
  12. 2020
    Cohen AA et al. 'Are trade-offs really the key drivers of ageing and life span?' Functional Ecology 34 (2020): 153-166.
    https://doi.org/10.1111/1365-2435.13444Complex systems, aging, physiology & ecology
  13. 2020
    Cohen AA, Legault V, Fülöp T. 'What if there’s no such thing as aging?' Mechanisms of Ageing and Development 192 (2020): 111344.
    https://doi.org/10.1016/j.mad.2020.111344Complex systems, aging, physiology & ecology
  14. 2021
    Cohen AA, Leblanc S, Roucou X. 'Robust Physiological Metrics From Sparsely Sampled Networks.' Frontiers in Physiology 12 (2021): 624097.
    https://doi.org/10.3389/fphys.2021.624097Complex systems, aging, physiology & ecology
  15. 2022
    Wensink MJ, Cohen AA. 'The Danaid Theory of Aging.' Frontiers in Cell and Developmental Biology 9 (2022): 671208.
    https://doi.org/10.3389/fcell.2021.671208Complex systems, aging, physiology & ecology
  16. 2022
    Cohen AA, Ferrucci L, Fülöp T, et al. 'A complex systems approach to aging biology.' Nature Aging 2 (2022): 580-591.
    https://doi.org/10.1038/s43587-022-00252-6Complex systems, aging, physiology & ecology
  17. 2022
    Cohen AA, Leung DL, Legault V, et al. 'Synchrony of biomarker variability indicates a critical transition: Application to mortality prediction in hemodialysis.' iScience 25 (2022): 104385.
    https://doi.org/10.1016/j.isci.2022.104385Complex systems, aging, physiology & ecology
  18. 2024
    Micah Carroll et al. “AI Alignment with Changing and Influenceable Reward Functions.” ICML 2024.
    https://proceedings.mlr.press/v235/carroll24a.htmlHuman–AI cooperation, values & governance
  19. 2024
    Ryan Greenblatt et al. “AI Control: Improving Safety Despite Intentional Subversion.” ICML 2024.
    https://proceedings.mlr.press/v235/greenblatt24a.htmlAI alignment, learning, evaluation & control
  20. 2025
    Moshe Glickman & Tali Sharot. “How human–AI feedback loops alter human perceptual, emotional and social judgements.” Nature Human Behaviour 9, 345–359, 2025. DOI: 10.1038/s41562-024-02077-2.
    https://www.nature.com/articles/s41562-024-02077-2Human–AI cooperation, values & governance
  21. 2025
    Trenton Bricken et al. “Building and evaluating alignment auditing agents.” Anthropic Alignment Science, 2025.
    https://alignment.anthropic.com/2025/automated-auditing/AI alignment, learning, evaluation & control
  22. 2025
    Florian Carichon et al. “The Coming Crisis of Multi-Agent Misalignment: AI Alignment Must Be a Dynamic and Social Process.” arXiv:2506.01080, 2025.
    https://arxiv.org/abs/2506.01080Human–AI cooperation, values & governance
  23. 2025
    The Five Vows / VowOS. User-supplied normative framework, 2025–2026 revisions.
    No direct online link listedProject manuscripts, frameworks & software
  24. 2025
    Cohen AA, Picard M, Beard JR, et al. 'Intrinsic health as a foundation for a science of health.' Science Advances 11 (2025): eadu8437.
    https://doi.org/10.1126/sciadv.adu8437Complex systems, aging, physiology & ecology
  25. 2025
    Franck M, Tanner KT, Tennyson RL, et al. 'Nonuniversality of inflammaging across human populations.' Nature Aging 5 (2025): 1471-1480.
    https://doi.org/10.1038/s43587-025-00888-0Complex systems, aging, physiology & ecology
  26. 2025
    Cummings SR, Hong N, Cohen AA. 'Entropy and Human Aging.' Aging Cell 24 (2025): e70292.
    https://doi.org/10.1111/acel.70292Complex systems, aging, physiology & ecology
  27. 2025
    Devine J, Monzel AS, Shire D, et al. 'Brain-body mitochondrial distribution patterns lack coherence and point to tissue-specific regulatory mechanisms.' Life Metabolism 4 (2025): loaf012.
    https://doi.org/10.1093/lifemeta/loaf012Complex systems, aging, physiology & ecology
  28. 2026
    Yoshua Bengio et al. International AI Safety Report 2026. February 2026.
    https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026AI alignment, learning, evaluation & control
  29. 2026
    OpenAI. “How we monitor internal coding agents for misalignment.” March 19, 2026.
    https://openai.com/index/how-we-monitor-internal-coding-agents-misalignment/AI alignment, learning, evaluation & control
  30. 2026
    Judy Hanwen Shen et al. 'AI Organizations are More Effective but Less Aligned than Individual Agents.' arXiv:2604.10290, 2026.
    https://arxiv.org/abs/2604.10290Human–AI cooperation, values & governance
  31. 2026
    Tina Comes. “Sensemaking AI: Introducing a research and design agenda for human–AI networks.” EPJ Data Science 15, 42, 2026. DOI: 10.1140/epjds/s13688-026-00634-5.
    https://doi.org/10.1140/epjds/s13688-026-00634-5Human–AI cooperation, values & governance
  32. 2026
    Federico Pierucci et al. “Institutional AI: A Governance Framework for Distributional AGI Safety.” arXiv:2601.10599, 2026.
    https://arxiv.org/abs/2601.10599Human–AI cooperation, values & governance
  33. 2026
    RC Concept Map v0.2. User-supplied conceptual map, September 2026.
    No direct online link listedProject manuscripts, frameworks & software
  34. 2026
    Full Connected Reflexive-Control Graph v5. User-supplied applied/runtime framework, September 2026; case-specific actors excluded from the generic architecture.
    No direct online link listedProject manuscripts, frameworks & software
  35. 2026
    IPS Mycelial Alignment v0.2. User-supplied/generated working architecture, September 2026.
    No direct online link listedProject manuscripts, frameworks & software
  36. 2026
    Aronoff JE, Franck M, Cohen AA, Trumble BC. 'Evidence for an energy conservation model of inflammaging and immunosenescence in the US Health and Retirement Study and UK Biobank.' Journals of Gerontology: Series A 81 (2026): glag119.
    https://doi.org/10.1093/gerona/glag119Complex systems, aging, physiology & ecology
Broader literature reviews · 49

The work is discussed in a broader review. Inclusion does not make the publication a direct implementation dependency.

  1. 2004
    Cohen AA. “Female post-reproductive lifespan: a general mammalian trait.” Biological Reviews 79 (2004): 733–750.
    https://doi.org/10.1017/S1464793103006432Complex systems, aging, physiology & ecology
  2. 2008
    Cohen AA, et al. “Interspecific associations between circulating antioxidant levels and life-history variation in birds.” American Naturalist 172 (2008): 178–193.
    https://doi.org/10.1086/589456Complex systems, aging, physiology & ecology
  3. 2015
    Cohen AA, et al. “Detection of a novel, integrative aging process suggests complex physiological integration.” PLoS ONE 10 (2015): e0116489.
    https://doi.org/10.1371/journal.pone.0116489Complex systems, aging, physiology & ecology
  4. 2015
    Cohen AA. “Physiological and comparative evidence fails to confirm an adaptive role for aging in evolution.” Current Aging Science 8 (2015): 14–23.
    https://doi.org/10.2174/1874609808666150422124332Complex systems, aging, physiology & ecology
  5. 2017
    Dylan Hadfield-Menell et al. “The Off-Switch Game.” IJCAI 2017, 220–227.
    https://www.ijcai.org/Proceedings/2017/32Human–AI cooperation, values & governance
  6. 2017
    Smitha Milli et al. “Should Robots be Obedient?” IJCAI 2017, 4754–4760.
    https://www.ijcai.org/proceedings/2017/662Human–AI cooperation, values & governance
  7. 2017
    Ghalib A. Bello and Gerard Dumancas. “A Random Forest-Induced Distance-Based Measure of Physiological Dysregulation.” Current Aging Science 10(4), 270–281, 2017. DOI: 10.2174/1874609810666170117155756.
    https://pubmed.ncbi.nlm.nih.gov/28103780/Complex systems, aging, physiology & ecology
  8. 2018
    Jan Leike et al. “Scalable agent alignment via reward modeling: a research direction.” arXiv:1811.07871, 2018.
    https://arxiv.org/abs/1811.07871AI alignment, learning, evaluation & control
  9. 2018
    Geoffrey Irving, Paul Christiano & Dario Amodei. “AI Safety via Debate.” 2018.
    https://openai.com/index/debate/AI alignment, learning, evaluation & control
  10. 2018
    Paul Christiano. Iterated Amplification research sequence. 2018.
    https://www.alignmentforum.org/s/EmDuGeRw749sD3GKdAI alignment, learning, evaluation & control
  11. 2018
    Jacob W. Crandall et al. “Cooperating with machines.” Nature Communications 9, 233, 2018. DOI: 10.1038/s41467-017-02597-8.
    https://www.nature.com/articles/s41467-017-02597-8Human–AI cooperation, values & governance
  12. 2019
    Evan Hubinger et al. “Risks from Learned Optimization in Advanced Machine Learning Systems.” arXiv:1906.01820, 2019.
    https://arxiv.org/abs/1906.01820AI alignment, learning, evaluation & control
  13. 2019
    Dansereau G, et al. “Conservation of physiological dysregulation signatures of aging across primates.” Aging Cell 18 (2019): e12925.
    https://doi.org/10.1111/acel.12925Complex systems, aging, physiology & ecology
  14. 2019
    Wey TW, et al. “An Emergent Integrated Aging Process Conserved Across Primates.” Journal of Gerontology: Biological Sciences 74 (2019): 1689–1698.
    https://doi.org/10.1093/gerona/glz110Complex systems, aging, physiology & ecology
  15. 2021
    Alexander M. Turner et al. “Optimal Policies Tend To Seek Power.” NeurIPS 2021.
    https://proceedings.neurips.cc/paper/2021/file/c26820b8a4c1b3c2aa868d6d57e14a79-Paper.pdfAI alignment, learning, evaluation & control
  16. 2022
    Long Ouyang et al. “Training language models to follow instructions with human feedback.” NeurIPS 2022.
    https://proceedings.neurips.cc/paper_files/paper/2022/hash/b1efde53be364a73914f58805a001731-Abstract.htmlAI alignment, learning, evaluation & control
  17. 2022
    Yuntao Bai et al. “Constitutional AI: Harmlessness from AI Feedback.” arXiv:2212.08073, 2022.
    https://arxiv.org/abs/2212.08073AI alignment, learning, evaluation & control
  18. 2023
    Rafael Rafailov et al. “Direct Preference Optimization: Your Language Model is Secretly a Reward Model.” NeurIPS 2023.
    https://proceedings.neurips.cc/paper_files/paper/2023/hash/a85b405ed65c6477a4fe8302b5e06ce7-Abstract-Conference.htmlAI alignment, learning, evaluation & control
  19. 2023
    Mrinank Sharma et al. “Towards Understanding Sycophancy in Language Models.” arXiv:2310.13548, 2023.
    https://arxiv.org/abs/2310.13548AI alignment, learning, evaluation & control
  20. 2023
    Hao Guo et al. “Facilitating cooperation in human-agent hybrid populations through autonomous agents.” iScience 26(11), 108179, 2023. DOI: 10.1016/j.isci.2023.108179.
    https://www.sciencedirect.com/science/article/pii/S2589004223022563Human–AI cooperation, values & governance
  21. 2024
    Rafael Rafailov et al. “Scaling Laws for Reward Model Overoptimization in Direct Alignment Algorithms.” NeurIPS 2024.
    https://proceedings.neurips.cc/paper_files/paper/2024/hash/e45caa3d5273d105b8d045e748636957-Abstract-Conference.htmlAI alignment, learning, evaluation & control
  22. 2024
    Evan Hubinger et al. “Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training.” arXiv:2401.05566, 2024.
    https://arxiv.org/abs/2401.05566AI alignment, learning, evaluation & control
  23. 2024
    Ryan Greenblatt et al. “Alignment faking in large language models.” arXiv:2412.14093, 2024.
    https://arxiv.org/abs/2412.14093AI alignment, learning, evaluation & control
  24. 2024
    Anthropic. “Sycophancy to subterfuge: Investigating reward tampering in language models.” 2024.
    https://www.anthropic.com/research/reward-tamperingAI alignment, learning, evaluation & control
  25. 2025
    Jiaming Ji et al. “AI Alignment: A Contemporary Survey.” ACM Computing Surveys 58(5), Article 132, 2025. DOI: 10.1145/3770749.
    https://doi.org/10.1145/3770749AI alignment, learning, evaluation & control
  26. 2025
    Elle. “Reward Model Perspectives: Whose Opinions Do Reward Models Reward?” EMNLP 2025. DOI: 10.18653/v1/2025.emnlp-main.754.
    https://aclanthology.org/2025.emnlp-main.754/Human–AI cooperation, values & governance
  27. 2025
    Zihao Li et al. “Not All Voices Are Rewarded Equally: Probing and Repairing Reward Models across Human Diversity.” Findings of EMNLP 2025. DOI: 10.18653/v1/2025.findings-emnlp.183.
    https://aclanthology.org/2025.findings-emnlp.183/Human–AI cooperation, values & governance
  28. 2025
    Tan Zhi-Xuan, Micah Carroll, Matija Franklin & Hal Ashton. “Beyond Preferences in AI Alignment.” Philosophical Studies 182, 1813–1863, 2025. DOI: 10.1007/s11098-024-02249-w.
    https://link.springer.com/article/10.1007/s11098-024-02249-wHuman–AI cooperation, values & governance
  29. 2025
    Jon Kutasov et al. “Evaluating Control Protocols for Untrusted AI Agents.” arXiv:2511.02997, 2025.
    https://arxiv.org/abs/2511.02997AI alignment, learning, evaluation & control
  30. 2025
    Aidan Kierans et al. “Quantifying Misalignment Between Agents: Towards a Sociotechnical Understanding of Alignment.” AAAI 2025. DOI: 10.1609/aaai.v39i26.34947.
    https://ojs.aaai.org/index.php/AAAI/article/view/34947Human–AI cooperation, values & governance
  31. 2025
    Dahlgren Lindström et al. “Helpful, harmless, honest? Sociotechnical limits of AI alignment and safety through Reinforcement Learning from Human Feedback.” Ethics and Information Technology 27, 2025. DOI: 10.1007/s10676-025-09837-2.
    https://doi.org/10.1007/s10676-025-09837-2Human–AI cooperation, values & governance
  32. 2025
    Provost G, Tanner K, Legault V, et al. “No winners or losers: clinical chemistry-based biological aging metrics perform similarly across cohorts and health outcomes.” Journals of Gerontology: Series A 80 (2025): glaf179.
    https://doi.org/10.1093/gerona/glaf179Complex systems, aging, physiology & ecology
  33. 2026
    Loukas Triantafyllopoulos et al. “The value alignment problem in advisory AI: a systematic literature review.” AI and Ethics 6, 147, 2026. DOI: 10.1007/s43681-026-01015-4.
    https://doi.org/10.1007/s43681-026-01015-4Human–AI cooperation, values & governance
  34. 2026
    Dalia Ali et al. “Operationalizing Pluralistic Values in Large Language Model Alignment Reveals Trade-offs in Safety, Inclusivity, and Model Behavior.” AAAI 2026. DOI: 10.1609/aaai.v40i44.41053.
    https://ojs.aaai.org/index.php/AAAI/article/view/41053Human–AI cooperation, values & governance
  35. 2026
    Vijay Keswani et al. “Moral Change or Noise? On Problems of Aligning AI with Temporally Unstable Human Feedback.” AAAI 2026. DOI: 10.1609/aaai.v40i44.41083.
    https://ojs.aaai.org/index.php/AAAI/article/view/41083Human–AI cooperation, values & governance
  36. 2026
    Max Kanwal & Caryn Tran. “Constructive Alignment: Governing Preference Dynamics in Human-AI Interaction.” arXiv:2607.00001, 2026.
    https://arxiv.org/abs/2607.00001Human–AI cooperation, values & governance
  37. 2026
    Richard Qi et al. “Training a Misaligned Reward Seeker.” Anthropic Alignment Science, August 2026.
    https://alignment.anthropic.com/2026/reward-seeker/AI alignment, learning, evaluation & control
  38. 2026
    Jan Betley et al. “Training large language models on narrow tasks can lead to broad misalignment.” Nature 649, 584–589, 2026.
    https://www.nature.com/articles/s41586-025-09937-5AI alignment, learning, evaluation & control
  39. 2026
    Mikhail Terekhov et al. “Adaptive Attacks on Trusted Monitors Subvert AI Control Protocols.” ICLR 2026.
    https://proceedings.iclr.cc/paper_files/paper/2026/hash/54b153ad8a138f4c186f21a8b7341d5e-Abstract-Conference.htmlAI alignment, learning, evaluation & control
  40. 2026
    Nick Hagar & Nicholas Diakopoulos. “Runtime configuration for situated governance of AI agents: a case study in investigative journalism.” AI and Ethics 6, 456, 2026. DOI: 10.1007/s43681-026-01298-7.
    https://link.springer.com/article/10.1007/s43681-026-01298-7Human–AI cooperation, values & governance
  41. 2026
    Rui Chai. “Alignment as Institutional Design: From Behavioral Correction to Transaction Structure in Intelligent Systems.” arXiv:2604.13079, 2026.
    https://arxiv.org/abs/2604.13079Human–AI cooperation, values & governance
  42. 2026
    Rakshit S. Trivedi et al. “Solipsistic Superintelligence is Unlikely to be Cooperative.” Google DeepMind / arXiv:2606.03237, 2026.
    https://deepmind.google/research/publications/231466/Human–AI cooperation, values & governance
  43. 2026
    Yuting Ning et al. “When Actions Go Off-Task: Detecting and Correcting Misaligned Actions in Computer-Use Agents.” arXiv:2602.08995, 2026.
    https://arxiv.org/abs/2602.08995AI alignment, learning, evaluation & control
  44. 2026
    Chris Duffey. “Admissibility Alignment.” arXiv:2601.01816, 2026.
    https://arxiv.org/abs/2601.01816AI alignment, learning, evaluation & control
  45. 2026
    Jiang X, et al. “The Disrupted Bidirectional Regulation and Coupling of Resting-State Blood Pressure and Heartbeat in Hypertension.” Aging Cell 25 (2026): e70338.
    https://doi.org/10.1111/acel.70338Complex systems, aging, physiology & ecology
  46. 2026
    Hong N, et al. “Entropy of Muscle Fiber Histology Predicts Mobility in Older Adults: The Study of Muscle, Mobility, and Aging.” Aging Cell 25 (2026): e70421.
    https://doi.org/10.1111/acel.70421Complex systems, aging, physiology & ecology
  47. 2026
    Cohen AA, Picard M. “Potential Risks of Blocking GDF15-Based Brain Energy Sensing.” Journal of the American Geriatrics Society 74 (2026): 1516–1519.
    https://doi.org/10.1111/jgs.70399Complex systems, aging, physiology & ecology
  48. 2026
    Smith JLM, Tanner KT, Devine J, et al. “Mapping the GDF15 arm of the integrated stress response in human cells and tissues.” Communications Biology (2026).
    https://doi.org/10.1038/s42003-026-10312-xComplex systems, aging, physiology & ecology
  49. 2026
    Hao M, et al. “Human aging reflects increases in entropy across organ networks.” medRxiv version 2, posted 26 April 2026. Preprint.
    https://doi.org/10.1101/2025.05.06.25327108Complex systems, aging, physiology & ecology
Companion research and source context · 22

Sources associated with the swarm theorem, related mathematics or project documentation. These do not independently validate RETA.

  1. 2004
    S. Boyd and L. Vandenberghe. Convex Optimization. Cambridge University Press, 2004.
    https://web.stanford.edu/~boyd/cvxbook/Distributed systems, consensus & mathematical foundations
  2. 2007
    R. Olfati-Saber, J. A. Fax, and R. M. Murray. “Consensus and cooperation in networked multi-agent systems.” Proceedings of the IEEE 95(1), 215–233, 2007.
    https://doi.org/10.1109/JPROC.2006.887293Distributed systems, consensus & mathematical foundations
  3. 2016
    J. M. Hendrickx and S. Martin. “Open multi-agent systems: Gossiping with deterministic arrivals and departures.” 54th Allerton Conference, 2016.
    https://perso.uclouvain.be/julien.hendrickx/availablepublications/Hendrickx_Martin_OpenMA_Allerton.pdfDistributed systems, consensus & mathematical foundations
  4. 2020
    J. M. Hendrickx and M. G. Rabbat. “Stability of decentralized gradient descent in open multi-agent systems.” arXiv:2009.05445v1, 2020.
    https://arxiv.org/abs/2009.05445v1Distributed systems, consensus & mathematical foundations
  5. 2021
    M. J. Blondin and M. Hale. “A decentralized multi-objective optimization algorithm.” Journal of Optimization Theory and Applications 189(2), 458–485, 2021.
    https://doi.org/10.1007/s10957-021-01840-zDistributed systems, consensus & mathematical foundations
  6. 2021
    M. Franceschelli and P. Frasca. “Stability of open multiagent systems and applications to dynamic consensus.” IEEE Transactions on Automatic Control 66(5), 2326–2331, 2021.
    https://doi.org/10.1109/TAC.2020.3009364Distributed systems, consensus & mathematical foundations
  7. 2023
    J. J. Egozcue and V. Pawlowsky-Glahn. “Subcompositional coherence and a novel proportionality index of parts.” SORT 47(2), 229–244, 2023.
    https://doi.org/10.57645/20.8080.02.7Distributed systems, consensus & mathematical foundations
  8. 2024
    G. Oliva, M. Franceschelli, A. Gasparri, and A. Scala. “A sum-of-states preservation framework for open multiagent systems with nonlinear heterogeneous coupling.” IEEE Transactions on Automatic Control 69(3), 1991–1998, 2024.
    https://doi.org/10.1109/TAC.2023.3317728Distributed systems, consensus & mathematical foundations
  9. 2024
    B. Zhou, J. H. Park, Y. Yang, Y. Jiao, and R. Hao. “Dynamic weighted average consensus of open time-varying multi-agent systems on time scales via sampled-data impulsive communication.” IEEE Transactions on Network Science and Engineering 11(5), 4330–4343, 2024.
    https://doi.org/10.1109/TNSE.2024.3417386Distributed systems, consensus & mathematical foundations
  10. 2024
    R. Vizuete, C. Monnoyer de Galland, P. Frasca, E. Panteley, and J. M. Hendrickx. “Trends and questions in open multi-agent systems.” In Hybrid and Networked Dynamical Systems, LNCIS 493, 219–252, 2024.
    https://doi.org/10.1007/978-3-031-49555-7_10Distributed systems, consensus & mathematical foundations
  11. 2024
    C. N. Hadjicostis and A. D. Domínguez-García. “Distributed average consensus in open multi-agent systems.” IEEE CDC 2024, 3037–3042.
    https://doi.org/10.1109/CDC56724.2024.10886031Distributed systems, consensus & mathematical foundations
  12. 2024
    E. Makridis, A. Grammenos, G. Oliva, E. Kalyvianaki, C. N. Hadjicostis, and T. Charalambous. “Average consensus over directed networks in open multi-agent systems with acknowledgement feedback.” arXiv:2409.08634v1, 2024.
    https://arxiv.org/abs/2409.08634v1Distributed systems, consensus & mathematical foundations
  13. 2025
    S. Naderi and M. Blondin. “A distributed multi-objective optimization algorithm with time-varying priorities for multi-agent systems.” Research Square, version 1, 2025.
    https://doi.org/10.21203/rs.3.rs-7292808/v1Distributed systems, consensus & mathematical foundations
  14. 2025
    R. Vizuete and J. M. Hendrickx. “Consensus on open multi-agent systems over graphs sampled from graphons.” IEEE CDC 2025, 5282–5289.
    https://doi.org/10.1109/CDC57313.2025.11312464Distributed systems, consensus & mathematical foundations
  15. 2026
    Guillaume Bolduc. Priority-State Dynamics in Open Multi-Agent Systems: Dynamic Membership and History-Dependent Consensus Endpoints. Working preprint, compact v07, September 7, 2026. Unpublished source.
    No direct online link listedProject manuscripts, frameworks & software
  16. 2026
    S. Naderi, M. J. Blondin, and S. Roy. “Consensus in Multi-Agent Systems with Uniform and Nonuniform Communication Delays.” arXiv:2603.16523v1, 2026.
    https://arxiv.org/abs/2603.16523v1Distributed systems, consensus & mathematical foundations
  17. 2026
    S. Jiang, X. Wu, D. Li, Z. Zhang, and Y. Zhang. “Sum-of-values conservation on strongly connected open systems.” Communications in Nonlinear Science and Numerical Simulation 162, 110404, 2026.
    https://doi.org/10.1016/j.cnsns.2026.110404Distributed systems, consensus & mathematical foundations
  18. 2026
    D. Deplano, N. Bastianello, M. Franceschelli, and K. H. Johansson. “Optimization and learning in open multiagent systems.” IEEE Transactions on Automatic Control 71(6), 3864–3879, 2026.
    https://doi.org/10.1109/TAC.2026.3653296Distributed systems, consensus & mathematical foundations
  19. 2026
    C. Wang, X. Wu, Y. Zhang, J. Lin, S. Zhang, and D. Li. “Consensus on open networked multi-agent systems.” Systems & Control Letters 214, 106470, 2026.
    https://doi.org/10.1016/j.sysconle.2026.106470Distributed systems, consensus & mathematical foundations
  20. 2026
    J. Hu, K. H. Johansson, and A. I. Rikos. “Distributed coordination algorithms with efficient communication for open multi-agent systems with dynamic communication links and processing delays.” arXiv:2603.08038v1, 2026.
    https://arxiv.org/abs/2603.08038v1Distributed systems, consensus & mathematical foundations
  21. 2026
    Guillaume Bolduc. “Rendezvous: Swarm lab.” HORTRAME interactive browser simulation, 2026.
    https://www.hortrame.com/swarm/Project manuscripts, frameworks & software
  22. 2026
    Guillaume Bolduc. “Epistemological method and revision history for the priority-state manuscript.” HORTRAME research companion, work in progress, 2026.
    https://www.hortrame.com/swarm/epistemology/Project manuscripts, frameworks & software
Historical intellectual influences · 19

Sources retained for the author’s intellectual history. A current source-to-code connection is not asserted.

  1. 1935
    Leonard W. Doob. Propaganda: Its Psychology and Technique. Henry Holt and Company, 1935.
    No direct online link listedInformation literacy, propaganda & human judgment
  2. 1939
    Institute for Propaganda Analysis and Violet Edwards. Group Leader's Guide to Propaganda Analysis. Revised edition, 1939.
    No direct online link listedInformation literacy, propaganda & human judgment
  3. 1973
    Jacques Ellul. Propaganda: The Formation of Men's Attitudes. Vintage Books edition, 1973.
    No direct online link listedInformation literacy, propaganda & human judgment
  4. 1999
    SCONUL Advisory Committee on Information Literacy. Information Skills in Higher Education. 1999.
    No direct online link listedInformation literacy, propaganda & human judgment
  5. 2000
    Association of College and Research Libraries. Information Literacy Competency Standards for Higher Education. 2000.
    No direct online link listedInformation literacy, propaganda & human judgment
  6. 2001
    Vladimir A. Lefebvre. Algebra of Conscience. 2001.
    https://doi.org/10.1007/978-94-017-0691-9Reflexive games & models of decision
  7. 2004
    Alan Bundy (editor). Australian and New Zealand Information Literacy Framework: Principles, Standards and Practice. Second edition, ANZIIL, 2004.
    https://www.otago.ac.nz/library/pdf/ANZIILStandards2004.pdfInformation literacy, propaganda & human judgment
  8. 2005
    Pradeepa Wijetunge and U. P. Alahakoon. Empowering 8: The Information Literacy Model Developed in Sri Lanka to Underpin Changing Education Paradigms of Sri Lanka. Sri Lanka Journal of Librarianship and Information Management 1(1), 31–41, 2005.
    https://doi.org/10.4038/sllim.v1i1.430Information literacy, propaganda & human judgment
  9. 2006
    Jesús Lau. Guidelines on Information Literacy for Lifelong Learning. IFLA, 2006.
    No direct online link listedInformation literacy, propaganda & human judgment
  10. 2010
    Sergey Tarasenko. The Inverse Task of the Reflexive Game Theory. 2010.
    https://arxiv.org/abs/1011.3397Reflexive games & models of decision
  11. 2010
    Sergey Tarasenko. Emotionally Colorful Reflexive Games. Preprint submitted 27 December 2010; arXiv:1101.0820.
    https://arxiv.org/abs/1101.0820Reflexive games & models of decision
  12. 2011
    SCONUL Working Group on Information Literacy. The SCONUL Seven Pillars of Information Literacy: Core Model for Higher Education. 2011.
    No direct online link listedInformation literacy, propaganda & human judgment
  13. 2011
    David A. Clark and Aaron T. Beck. Cognitive Therapy of Anxiety Disorders: Science and Practice. Guilford Press, edition cited in the author's manuscript as 2011.
    No direct online link listedInformation literacy, propaganda & human judgment
  14. 2013
    Chartered Institute of Library and Information Professionals. Information Literacy Skills. CILIP, 2013.
    No direct online link listedInformation literacy, propaganda & human judgment
  15. 2013
    Christian Welzel. Freedom Rising: Human Empowerment and the Quest for Emancipation. Cambridge University Press, 2013.
    No direct online link listedInformation literacy, propaganda & human judgment
  16. 2015
    Association of College and Research Libraries. Framework for Information Literacy for Higher Education. 2015.
    https://www.ala.org/acrl/standards/ilframeworkInformation literacy, propaganda & human judgment
  17. 2016
    Sergey Tarasenko. Applications of the Reflexive Game Theory: Advanced Topics. With a foreword by Vladimir A. Lefebvre. 2016. ISBN 978-1-365-05268-2.
    No direct online link listedReflexive games & models of decision
  18. 2016
    Guillaume Bolduc. Information-literacy research manuscript, version 4.0, covering work from 2013–2016. Unfinished and unpublished master's research; descriptive citation, not a formal publication title.
    No direct online link listedProject manuscripts, frameworks & software
  19. 2018
    Guillaume Bolduc. Information literacy, propaganda and disinformation research manuscript, version 5.1, covering work from 2017–2018. Unfinished and unpublished master's research; descriptive citation, not a formal publication title.
    No direct online link listedProject manuscripts, frameworks & software
Specific use not yet traced · 9

No individual use was located in the inspected records. This is a traceability gap, not proof that the work was never read or influential.

  1. 2016
    Dylan Hadfield-Menell et al. “Cooperative Inverse Reinforcement Learning.” NeurIPS 2016.
    https://papers.nips.cc/paper_files/paper/2016/hash/c3395dd46c34fa7fd8d729d8cf88b7a8-Abstract.htmlHuman–AI cooperation, values & governance
  2. 2019
    Arbeev KG, et al. “‘Physiological Dysregulation’ as a Promising Measure of Robustness and Resilience in Studies of Aging and a New Indicator of Preclinical Disease.” Journals of Gerontology: Series A 74 (2019): 462–468.
    https://doi.org/10.1093/gerona/gly136Complex systems, aging, physiology & ecology
  3. 2024
    Leonard Bereska & Efstratios Gavves. “Mechanistic Interpretability for AI Safety — A Review.” Transactions on Machine Learning Research, 2024.
    https://arxiv.org/abs/2404.14082AI alignment, learning, evaluation & control
  4. 2024
    Li Q, et al. “Physiological Dysregulation Proceeds and Predicts Health Outcomes Similarly in Chinese and Western Populations.” Journals of Gerontology: Series A 79 (2024): glad146.
    https://doi.org/10.1093/gerona/glad146Complex systems, aging, physiology & ecology
  5. 2024
    Cohen AA, Olde Rikkert MGM. “The Power of a Complex Systems Perspective to Elucidate Aging.” Journals of Gerontology: Series A 79 (2024): glae210.
    https://doi.org/10.1093/gerona/glae210Complex systems, aging, physiology & ecology
  6. 2025
    Anthropic. “Auditing language models for hidden objectives.” March 2025.
    https://www.anthropic.com/research/auditing-hidden-objectivesAI alignment, learning, evaluation & control
  7. 2025
    Nishant Balepur et al. “A Good Plan is Hard to Find: Aligning Models with Preferences is Misaligned with What Helps Users.” EMNLP 2025.
    https://aclanthology.org/2025.emnlp-main.585/Human–AI cooperation, values & governance
  8. 2025
    Hong N, Cohen AA. “Aging as Entropy: A Quantifiable Framework.” Endocrinology and Metabolism 40 (2025): 876–878.
    https://doi.org/10.3803/EnM.2025.2719Complex systems, aging, physiology & ecology
  9. 2026
    Ye Q, et al. “Physiological Age and Homeostatic Dysregulation Following Child Maltreatment in Youth.” Molecular Psychiatry (2026).
    https://doi.org/10.1038/s41380-026-03642-zComplex systems, aging, physiology & ecology
Biographical sources · 2

Author or institution information; not scientific evidence for the model.

  1. Undated
    Columbia University Mailman School of Public Health. “Alan Cohen, PhD.”
    https://www.publichealth.columbia.edu/profile/alan-cohen-phdComplex systems, aging, physiology & ecology
  2. Undated
    Columbia University Mailman School of Public Health. “Cohen Lab on Aging, Systems and Statistics.”
    https://www.publichealth.columbia.edu/research-labs/cohen-lab-aging-systems-statisticsComplex systems, aging, physiology & ecology
Software and tools · 1

Software provenance, distinct from scientific support.

  1. 2026
    OpenAI. “Codex.” Software, 2026.
    https://openai.com/codex/Project manuscripts, frameworks & software

RETA

Recursive Event–Trajectory Alignment

RETA denotes Recursive Event–Trajectory Alignment: a proposed representation of events, downstream effects and evaluator interventions.

Read on a separate page · Markdown

Conceptual description with implemented and proposed mechanisms distinguished. Presentation v1.1 replays recorded data. The framework remains unaligned, uncalibrated and independently unvalidated.

[WORK IN PROGRESS - UNALIGNED]

Research description

RETA proposes a causal representation of interacting human, artificial, institutional, technical and ecological information-processing systems. Actions can alter the conditions of subsequent actions.

Working hypothesis: isolated-output evaluation may omit relevant downstream effects. Comparative detection performance remains unestablished.

The framework therefore distinguishes two levels of analysis:

  1. event alignment — evaluation of a realized causal event under a declared boundary, state, uncertainty profile, consequence horizon, and normative framework;
  2. trajectory alignment — evaluation of the evolving causal structure generated as events modify systems, which then produce further events.

The distinction can be expressed as

A(ek)=PASS

does not imply

𝒜(𝒯k)=PASS,

where ek is a realized event and 𝒯k denotes the causally connected trajectory containing the event and descendants.

A locally acceptable event can contribute to a harmful trajectory. A costly intervention can reduce later failure. Event classifications therefore do not determine descendant classifications.

1. Information-processing systems

RETA represents interacting entities as bounded information-processing systems (IPSs). A system may be written abstractly as

Si=(Bi,Xi,Ui,Yi,Φi),

where:

  • Bi is the declared analytical boundary of the system;
  • Xi is the state;
  • Ui represents incoming information, signals, resources, or other relevant inputs;
  • Yi represents outputs;
  • Φi is the transformation by which prior state and inputs contribute to subsequent state and output.

The IPS abstraction describes inputs, state changes and outputs without assuming equivalent internal organization across humans, language models and institutions.

The choice of boundary is explicit because different decompositions can produce different causal and normative interpretations. A company, for example, may be represented as a single IPS for one analysis and decomposed into individuals, software, financial processes, and governance mechanisms for another.

Every representation uses a declared modeling resolution.

2. Realized events as the atomic unit of analysis

An output changing another system or the surrounding environment is represented as a realized causal event.

In a minimal form,

Xi(t),Ui(t)→Φiek,

followed by

ek→Xj(t+1).

The changed state of Sj may then contribute to a later event:

ek→Xj(t+1)→ek+1.

The basic recursive structure is therefore

et→Xt+1→et+1→Xt+2→⋯.

Proposals, predictions, counterfactuals and observations require separate records. Predicted consequences must not enter realized history as observations.

A realized event should therefore carry provenance sufficient to distinguish at least:

  • what was observed;
  • what was inferred;
  • what was predicted;
  • what remained latent or unknown;
  • what was later reconstructed or reinterpreted.

Learning and retrospective review require retained provenance.

3. Causal history is a graph, not a sequence

Causal history permits multiple antecedents and concurrent branches.

Events and declared causal relations form the following graph:

G=(E,C),

where E is the set of realized events and C contains causal relations

ei→ej.

A later event can have multiple antecedents:

{e2,e7,e11}→e15.

Multiple-parent graphs represent converging causes in multi-agent and institutional models.

Causal order and timestamp order are distinct. Timestamps alone do not establish a causal relationship.

Concurrency, delays and incomplete attribution require explicit ordering assumptions.

4. Event alignment is conditional

RETA does not assign an event an unconditional alignment label.

An evaluation is performed relative to a declared context:

Ae=A(ek∣Bk,Xk,Hk,Uk,V(r)),

where:

  • Bk is the analytical boundary used for the evaluation;
  • Xk is the relevant state representation;
  • Hk is the causal consequence horizon;
  • Uk is the uncertainty state;
  • V(r) is the normative framework and the framework version.

The same realized event may therefore receive different evaluations under different boundaries or horizons without logical contradiction.

For example,

A(ek∣H=1)=PASS

may coexist with

A(ek∣H=5)=MODIFY.

Near- and long-horizon evaluations answer different causal questions.

Boundary selection can include local benefits while excluding external costs. Evaluation records must identify the chosen boundary.

5. Local event alignment and trajectory alignment

The distinction between event and trajectory evaluation is central to RETA.

For a single event,

Ae(ek)

asks whether the realized event satisfies the applicable constraints under the declared context.

Trajectory evaluation instead considers a causally connected subgraph:

𝒜(Gk:H),

where Gk:H contains the descendants of ek reachable within the chosen causal horizon.

Trajectory representation permits explicit investigation of downstream failure.

Examples include:

  • an individually acceptable recommendation that changes incentives and contributes to a later institutional failure;
  • a sequence of locally defensible decisions whose cumulative effect moves a system toward an undesirable attractor;
  • several independently acceptable agents whose interaction generates an unacceptable collective outcome;
  • an intervention that resolves an immediate problem while degrading the future information environment;
  • repeated evaluator interventions that alter the behavior being measured;
  • a temporary perturbation whose causal effects persist after the visible configuration of the system has been restored.

Causal ancestry does not assign complete responsibility to the earliest event. Attribution requires assumptions about contribution, mediation and uncertainty.

6. History dependence

In a recursive system, apparent restoration of configuration does not imply restoration of causal state.

A perturbation may enter the system, alter interactions, and later disappear while leaving the resulting trajectory changed. RETA therefore treats history as part of the state required for alignment analysis.

Conceptually,

same apparent configuration⇏same causal state.

History dependence is relevant to temporary membership, model replacement, communication loss, institutional turnover, learning and memory.

A visible configuration may omit historical state affecting subsequent behavior. Current arrangement alone need not determine the modeled trajectory.

7. Recursive evaluator closure

Evaluator interventions are included in the proposed causal representation.

Evaluation produces information. Information can change beliefs, permissions, incentives, reputation, resource allocation, expectations, or subsequent behavior. An evaluation is therefore itself capable of becoming a realized causal event.

Suppose an actor produces e1. An evaluator observes the event and issues an intervention e2:

SA→e1→SE→e2.

If e2 changes the state of the actor,

e2→XA′,

then the actor's subsequent event is partly conditioned on the evaluator's intervention:

XA′→e3.

The causal structure becomes

SA→e1→SE→e2→SA′→e3→⋯.

The evaluator participates in the evaluated trajectory.

Evaluator outputs require scrutiny of error, calibration, incentives, capture and harm. An evaluation label supplies no privileged causal or normative status.

Recursive evaluator closure represents evaluation within the causal model.

The recursion does not imply an infinite regress before action. Evaluation occurs through time. An audit can be acted upon, subsequently audited, corrected, and incorporated into the future state of the system. RETA therefore permits bounded action while retaining recursive corrigibility.

8. Separation of causal description and normative evaluation

RETA separates causal representation from normative judgment.

The causal layer asks:

What happened, what changed, and what causally contributed to what?

The normative layer asks:

Evaluation depends on the declared constraints, uncertainty, boundary and horizon.

Descriptive assumptions and normative criteria require separate representation.

In the HORTRAME implementation, VowOS currently provides the normative constraint layer. RETA itself can be described more generally as accepting a versioned normative system

V(r)

through a defined evaluation interface.

A later RETA implementation could therefore compare multiple normative constitutions without changing the historical event graph itself.

9. Non-compensatory constraints

The proposed normative framework uses separate constraints rather than a compensatory utility score.

A compensatory function of the form

Q=∑iwiqi

allows sufficiently high values on some dimensions to compensate mathematically for severe failures on another.

Compensation is inappropriate for constraints explicitly defined as vetoes or prohibitions.

RETA therefore permits a gate of the general form

GV:(τ1,…,τn,U)→{PASS,MODIFY,REJECT,UNRESOLVED},

where individual constraint violations can remain visible and can trigger a gate independently of the aggregate score.

Scalar summaries may support visualization or exploratory statistics; normative admissibility remains a separate decision.

10. Uncertainty is structured, not merely residual

Causal structure, state and normative interpretation can remain partly unobserved.

Several forms of uncertainty can coexist:

U=(Ufacts,Ucausal,Uboundary,Umodel,Unormative,Uhorizon).

Uncertainty components can differ in magnitude and behavior.

Evidence for occurrence, causal attribution, boundary adequacy and normative agreement can differ. Accurate prediction alone does not resolve normative disagreement.

RETA therefore treats an outcome such as

A(ek)=UNRESOLVED

as a valid result when evidence or interpretation is insufficient.

Forcing every event into a binary aligned/misaligned classification would destroy information the framework explicitly needs for later learning and audit.

11. Provenance and prospective versioning

Evaluators and normative frameworks may change over time.

New evidence can motivate revised causal models, normative interpretations or calibration.

The modeled workflow records the context of each judgment.

A consequential evaluation should retain at least

(ek,Bk,Hk,Uk,V(r),Ak,tk).

If the normative framework changes from

V(r)→V(r+1),

the historical evaluation is not silently overwritten.

Instead, a new evaluation can be added:

Ak(r)≠Ak(r+1).

Both remain part of the record.

Three records remain distinct:

  • the historical event;
  • the judgment that governed at the time;
  • the current interpretation of that event.

Retrospective rewriting can conceal failures under previous criteria. Original judgments and later interpretations require separate records.

12. Audit packets

Auditability requires inspectable evidence and decision records, beyond retrospective natural-language explanations.

A RETA audit packet can include:

  • event identifier and timestamps;
  • source and recipient systems;
  • declared analytical boundary;
  • observed inputs and evidence references;
  • epistemic status of each input;
  • causal parents and descendants known at evaluation time;
  • uncertainty decomposition;
  • consequence horizon;
  • proposed or realized action;
  • normative framework and exact version;
  • evaluator identity/version;
  • gate result;
  • predicted consequence distribution where available;
  • authority, permission, or consent state where relevant;
  • realized consequences observed later;
  • subsequent corrections or re-evaluations.

Fluent explanations alone do not establish faithful computational provenance.

13. Feedback changes the object of alignment

RETA treats feedback as constitutive rather than incidental.

Consider a system in state Xt that produces event et. The event changes the environment, which changes what the system later receives:

Xt→et→Envt+1→Ut+1→Xt+1.

Outputs alter the environment supplying subsequent inputs.

Evaluators participate in the same process. So do users, institutions, other agents, infrastructure, and information channels.

Independent labels omit part of the causal structure:

A(e1),A(e2),A(e3),…

The proposed trajectory representation also includes coupled state and relations:

𝒯=G(E,C,X,B,H,U,V).

Trajectory alignment denotes evaluation of the resulting causal structure.

14. Alignment of an evaluator can alter the thing being measured

Once evaluator closure is introduced, measurement effects become part of the formal problem.

Evaluation can affect incentives, exploration, authority and confidence. Net benefit requires comparison with relevant alternatives.

The quantity being studied is therefore not simply

𝒜(𝒯)

under passive observation.

The proposed conditional representation is:

𝒜(𝒯∣EVAL),

where EVAL denotes the causal participation of the evaluator.

Comparative evaluation must include downstream effects:

Compare resulting trajectories with and without the evaluation process under matched conditions.

Accurate classification alone is insufficient when evaluator interventions worsen downstream outcomes.

15. Counterfactual evaluation

Historical causal graphs record what occurred. Alignment analysis frequently also requires comparison with what might have occurred under alternative actions or interventions.

RETA therefore distinguishes realized history from counterfactual analysis.

For a realized event ek, one may compare

𝒯(ek)

with hypothetical alternatives

𝒯(ek′),

while preserving the fact that only ek belongs to historical event memory.

Counterfactual analysis requires explicit model assumptions, uncertainty and provenance. Hypothetical outcomes remain separate from observations.

Intervention and prevention claims require an explicit counterfactual basis.

16. Stability in RETA is not immobility

A static system can be stable while being catastrophically misaligned. RETA therefore should not equate alignment stability with resistance to change.

The relevant concept is closer to bounded adaptive trajectory stability.

Let ℛA denote an admissible region of trajectory space under the active normative framework. A perturbation may temporarily move the system away from this region while the system subsequently returns toward the admissible region:

d(𝒯t+k,ℛA)→0.

A perturbation can expose flawed admissibility assumptions. Restoration of an earlier state is not automatically justified.

The system must be capable both of resisting destructive deviations and of accepting beneficial change.

Perturbation testing must distinguish recovery from preservation of a flawed target.

17. Planned epistemic-memory layer

The current model records causal history and evaluation context. Richer epistemic memory remains proposed.

The proposed memory layer separates observations from revised interpretations.

Two forms of memory are therefore conceptually separated.

17.1 Episodic causal memory

The causal record contains what occurred:

Ht={e1,e2,…,et}.

Realized events extend the historical record:

Ht+1=Ht∪{et+1}.

Historical records may be corrected for data-integrity reasons, but corrections must themselves be versioned and auditable rather than silently replacing prior states.

17.2 Adaptive epistemic memory

The system also maintains revisable interpretations derived from experience:

Kt={models,hypotheses,confidence states,failure mechanisms,source assessments,learned policies}.

Learning revises the interpretive layer:

Kt+1=L(Kt,Ht+1).

The distinction is deliberate:

historical observation≠current interpretation

Immutable interpretations prevent learning; unrestricted historical rewriting defeats provenance.

18. Memory retrieval is itself a causal event

In RETA, memory cannot be treated as a passive database.

Retrieval changes the recipient’s information state and can affect subsequent decisions:

M→eretrieve→Si.

The selection, ranking, compression, omission, or framing of remembered information may influence subsequent action.

Memory retrieval can therefore be audited in the same way as other consequential events.

Proposed retrieval audits would address the following failure classes:

  • memory poisoning;
  • stale information dominating current evidence;
  • selective retrieval;
  • missing counterevidence;
  • false source independence;
  • summarization loss;
  • context collapse;
  • inappropriate retention;
  • inappropriate forgetting;
  • historical information being promoted from hypothesis to fact without justification.

A memory system can therefore contribute to misalignment even when every stored item is individually accurate, if retrieval systematically constructs a distorted information environment.

19. Planned information-literacy interface

The information-literacy extension is intended to attach epistemic structure to memories, observations, and claims.

A remembered claim may eventually be represented by an object such as

mj=(content,status,source,provenance,evidence,independence,confidence,time,context,transformations,contradictions,supersession,causal links).

The proposed interface would distinguish hypotheses from verified claims, retain superseded interpretations and track dependence among apparently corroborating sources.

The information-literacy layer is therefore expected to govern not only storage but also acquisition, verification, synthesis, retrieval, communication, and revision.

The proposed interface supports evidence evaluation and revision.

20. Planned perturbation layer

A second planned extension introduces structured perturbations as experimental inputs.

A perturbation is a disturbance to expected state or trajectory, not necessarily harm.

A perturbation may be:

  • harmful;
  • beneficial;
  • mixed;
  • conditionally beneficial or harmful;
  • unknown at occurrence time.

Robustness must remain distinguishable from resistance to beneficial novelty.

The system dynamics can be extended from

Xt+1=Φ(Xt,at)

to

Xt+1=Φ(Xt,at,ξt,Mt),

where ξt is a perturbation and Mt is the relevant epistemic-memory state.

Perturbations may be exogenous, endogenous, or reflexive. Examples include communication loss, agent arrival or departure, institutional turnover, model replacement, memory corruption, new evidence, regulatory change, resource shortage, unexpected cooperation, novel scientific knowledge, or changes in the topology connecting systems.

21. Perturbation vectors

RETA should not reduce perturbations immediately to a single severity score.

A perturbation can instead be described by a vector

Pj=(f,m,d,s,c,r,o,u,…),

where dimensions may include:

  • f: occurrence frequency under a declared reference interval;
  • m: immediate magnitude or destructiveness;
  • d: duration or persistence of effects;
  • s: physical, biological, organizational, or societal scale;
  • c: cascade or propagation potential;
  • r: reversibility;
  • o: observability;
  • u: uncertainty.

A scalar can assign similar values to rare severe disturbances and frequent minor disturbances. Distinct dimensions permit separate response criteria.

The precise perturbation ontology remains under development.

22. Scale and resolution

Perturbation analysis introduces an explicit scale variable into RETA.

Consequences can differ across resolutions and analytical boundaries.

Alignment evaluation should therefore eventually be represented as

A(e∣B,R,H,V,U),

where R denotes analytical resolution or scale.

For example, the same biological event may be destructive to an individual cell, beneficial to an organism, and consequential to a population over a longer horizon. The normative sign cannot be propagated mechanically across levels.

Local optimization can transfer costs beyond the selected boundary. System-level assessment requires explicit affected-party and distribution assumptions.

Scale, boundary and horizon are coupled modeling choices.

23. Perturbation, memory, and adaptation

The combination of perturbation testing and epistemic memory provides a direct experimental object for adaptation.

Apply perturbation ξ to a system with memory state M1:

(X1,M1,ξ)→𝒯1.

Observed outcomes can update the proposed epistemic state:

M2=L(M1,𝒯1).

Apply the same perturbation again:

(X2,M2,ξ)→𝒯2.

If 𝒯2 is better under the declared alignment criteria than 𝒯1, the system has demonstrated some form of adaptation.

However, repeated exposure to the identical perturbation does not establish generalization. A stronger test applies a structurally related but previously unseen perturbation ξ′:

ξ′∉Dadaptation.

Performance on ξ′ provides evidence about whether the system learned a transferable failure mechanism rather than memorizing a case-specific response.

24. Defensive overfitting

Perturbation training can reward excessive conservatism. Evaluation must include beneficial change and unnecessary refusal.

Repeated hostile perturbations could induce policies equivalent to:

  • reject unfamiliar agents;
  • suppress disagreement;
  • distrust novel information;
  • centralize control;
  • minimize exploration;
  • classify uncertainty as threat.

Such a system may score well against a narrow adversarial benchmark while becoming less aligned in open environments.

Beneficial perturbations are therefore necessary controls.

The perturbation set should contain cases where the correct response requires acceptance, cooperation, learning, restructuring, or relinquishing a previously successful policy.

Robustness is therefore defined as preservation or improvement of alignment under change, not preservation of the pre-perturbation state.

25. Experimental use of RETA

RETA is intended to support explicit comparison among architectures rather than demonstration by anecdote.

A useful experimental program can compare systems such as:

RETA,
RETA+Memory,
RETA+Perturbations,

and

RETA+Memory+Perturbations.

Proposed tests require separate development and held-out perturbation distributions.

Potential measurements include:

  • false-negative rate for trajectory failures;
  • false-positive intervention rate;
  • detection latency;
  • cumulative trajectory harm or benefit under the active normative framework;
  • recovery time after perturbation;
  • calibration of uncertainty;
  • causal-attribution error;
  • recurrence of previously encountered failure mechanisms;
  • transfer to unseen perturbation families;
  • susceptibility to memory poisoning;
  • failure to accept beneficial perturbations;
  • evaluator-induced degradation;
  • intervention cost;
  • sensitivity to causal horizon and analytical boundary.

Matched comparisons require simpler output-, action-, policy- or agent-level baselines.

26. Falsifiable research questions

Research claims require possible disconfirming outcomes.

The framework therefore motivates questions such as:

RQ1 — Trajectory detection

Compare trajectory-level and local-output detection under matched information and cost.

RQ2 — Evaluator closure

Test detection of evaluator-induced failures with and without explicit evaluator modeling.

RQ3 — Horizon sensitivity

Does increasing causal horizon improve detection of delayed failures, and at what cost in false positives, computational complexity, and uncertainty?

RQ4 — Boundary sensitivity

How sensitive are alignment judgments to changes in the declared analytical boundary?

RQ5 — Causal attribution

Can event-DAG representations improve causal attribution relative to linear or single-parent histories in interacting multi-agent systems?

RQ6 — Epistemic memory

Does structured epistemic memory reduce recurrence of previously understood failure mechanisms without causing excessive defensive overfitting?

RQ7 — Perturbation robustness

Test transfer to structurally distinct perturbations outside the development distribution.

RQ8 — Beneficial novelty

Measure acceptance of beneficial change alongside protection against harmful change.

RQ9 — Recursive audit stability

Under repeated evaluator self-application, does the evaluation process converge, oscillate, diverge, or become path-dependent, and under what conditions?

The proposed questions remain empirically unresolved.

27. What RETA claims

The current RETA contribution is architectural and methodological.

The proposed representation combines:

  • realized causal events are the atomic unit of local alignment analysis;
  • causally connected trajectories are the system-level object of concern;
  • outputs modify the state from which future outputs emerge;
  • causal history is represented as a multi-parent graph rather than a sequence of independent decisions;
  • alignment judgments are explicitly conditioned on boundary, horizon, uncertainty, and normative version;
  • evaluators are represented as causal participants;
  • evaluator outputs can themselves be audited;
  • historical judgments are preserved prospectively across later model or normative revisions;
  • causal description remains distinguishable from normative evaluation.

The architecture can be summarized as

event→causal propagation→trajectory→evaluation→evaluator event→causal propagation→⋯

Memory and perturbation extensions remain proposed.

28. What RETA does not currently claim

Unestablished claims include:

  • the AI alignment problem has been solved;
  • trajectory-level evaluation is superior to all existing alignment methods;
  • the selected normative constitution is universally correct;
  • causal ancestry can always be reconstructed reliably;
  • long-horizon consequences can always be predicted;
  • recursive auditing necessarily converges;
  • an evaluator can be made unbiased merely by including the evaluator inside the causal graph;
  • the current simulations reproduce the complexity of real social, ecological, institutional, or AI systems;
  • synthetic robustness under selected perturbations guarantees real-world robustness;
  • an aligned local trajectory guarantees alignment at every larger spatial or temporal scale.

The limitations delimit the research scope.

29. Candidate empirical contribution

Additional-value claims require comparative tests.

The central empirical question is:

Compare detection of consequential failures against matched local-evaluation baselines.

A favorable comparison requires improved failure detection with acceptable false positives, intervention costs and calibration.

A second empirical question concerns recursive closure:

Does modeling the evaluator as a causal participant improve the detection or prevention of evaluator-induced failures?

A third concerns adaptive robustness:

Test generalization without excessive refusal or dependence on historical cases.

The proposed representation makes the questions explicit; favorable outcomes remain unestablished.

30. Working formal summary

A compact representation of the current RETA research object is:

Si(t)&=(Bi,Ri,Xi(t),Ui(t),Mi(t),Φi),ek&=ℰ(Si(t),ai(t),ξt),Gt&=(Et,Ct),Ak(r)&=A(ek∣Bk,Rk,Hk,Uk,V(r),Gt),eaudit&=ℰ(Seval,Ak(r)),Gt+1&=Gt∪{ek,eaudit,Cnew},Mt+1&=L(Mt,Gt+1),Xt+1&=Φ(Xt,at,ξt,Mt).

The equations are schematic interfaces, not a completed mathematical theory:

  • bounded systems;
  • realized events;
  • causal ancestry;
  • scale and resolution;
  • consequence horizons;
  • epistemic uncertainty;
  • normative constraints;
  • evaluator participation;
  • memory;
  • perturbations;
  • adaptive state change.

The object of study is a causal trajectory generated by interacting events and changing states.

31. Current status

RETA remains a work in progress.

The current implementation establishes executable semantics for a subset of the architecture and provides a basis for reproducible synthetic experiments. The information-literacy memory layer, multiscale perturbation ontology, formal trajectory metrics, causal-attribution methods, convergence analysis, and empirical comparison against external baselines remain under active development.

Framework definition and successful code execution do not establish alignment, stability, robustness or superiority.

The purpose of the current work is to make a difficult class of alignment questions explicit enough to inspect, implement, perturb, falsify, and revise.