Index

This is the navigation map for the LLM-maintained Obsidian research wiki.

Areas

• Safe Multi-Agent Reinforcement Learning: safety interventions, constraints, verification, robustness, and evaluation in multi-agent RL settings.
• AI Evaluation and Benchmarking: benchmark design, saturation, calibration, monitorability, and capability measurement for modern AI systems.
• Logic and Action Formalisms: situation calculus, action theories, executability, and reasoning about change.
• Spatial Algorithms and Data Structures: spatial indexing, locality-aware search, and geometric query procedures.
• Temporal Logic: temporal specification languages, automata-shaped tasks, monitors, and time-indexed objectives.
• Formal Methods: synthesis, verification, abstraction, and specification-driven analysis of learning systems.
• Causality: causal structure, interventions, and explanation-oriented reasoning.
• Continual Learning: long-lived adaptation, non-stationary training, task-model revision, and transfer across tasks.
• Explainable AI: interpretable models, mechanistic explanation, and trustworthy model behavior.
• Strategic Reasoning: incentive, equilibrium, information, and adaptation structure in multi-agent interaction.

Foundations

• Algorithms and Data Structures: reusable data structures, search procedures, and complexity tradeoffs across AI and computing.
• Deep Learning Fundamentals: representation learning, optimization, architectures, and training dynamics across ML, interpretability, and sequential decision-making.
• Probability and Statistics: uncertainty, estimation, inference, calibration, causality, and statistical reasoning across AI and science.
• Control Theory: dynamical systems, feedback, stability, estimation, control, and robotics-oriented reasoning.
• Decision Theory: preferences, utility, uncertainty, trade-offs, and rational choice in sequential and multi-agent settings.

Concepts

• Probabilistic Model Checking: verification of stochastic models against quantitative temporal, reward, and risk properties.
• Constrained Markov Decision Processes: MDPs with explicit expected-cost feasibility constraints in addition to reward maximization.
• Constrained Policy Optimization: trust-region safe-RL algorithm for optimizing neural policies under CMDP cost constraints.
• Sound Value Iteration: value-iteration method that computes explicit upper/lower bounds for probabilistic reachability and expected-reward queries.
• Distributional Value Iteration: Bellman-style dynamic programming over return distributions rather than scalar expected values.
• Non-Markovian Reinforcement Learning: RL problems where reward, validity, or constraints depend on event histories.
• Compositional Reinforcement Learning: modular RL systems whose subsystem interfaces support decomposed training, verification, and system-level task guarantees.
• Automata Learning: inference of finite automata from positive/negative traces, queries, or interaction data.
• Reward Machines: finite-state reward specifications that expose temporal reward structure over high-level event traces.
• Alternating-Time Temporal Logic: temporal logic for reasoning about what coalitions of agents can force.
• Branching-Time Temporal Logic: temporal logic family for reasoning over trees of possible futures rather than one execution trace.
• Dynamic Epistemic Logic: modal logic for modeling how agents’ knowledge changes after information-changing events.
• Intuitionistic Logic: proof-centered non-classical logic that rejects unrestricted excluded middle without constructive evidence.
• Chain-of-Thought Monitorability: whether monitors can infer behavior-relevant properties from reasoning traces.
• Benchmark Gaming: benchmark score inflation through exploiting the evaluator rather than solving the intended task.
• Calibration: alignment between predicted confidence and observed frequencies, distinct from informativeness.
• Brier Score: proper scoring rule for probabilistic forecasts that combines calibration-sensitive error with broader forecast quality.
• Counting in Probability: finite-sample-space probability calculations via exact combinatorial counts.
• Opponent Shaping: multi-agent learning in which an agent acts partly to influence another learner’s future updates.
• Aspiration-Based Reinforcement Learning: satisficing reinforcement rule where actions are strengthened or weakened by payoff relative to aspiration levels.
• Responsibility Anticipation: ex ante strategy evaluation by the avoidable values an agent may be responsible for under possible choices by others.
• Dyadic Morality: moral-cognition model compressing judgment into perceived agency, patient vulnerability, and causal harm.
• Corrigibility: oversight-preservation property where an agent prefers waiting for human approval to acting or disabling intervention.
• Goal Alignment: objective-matching problem where the agent’s optimized utility should track designer-intended true utility.
• Guaranteed Safe AI: safety-case framework combining world models, safety specifications, and verifiers to produce quantitative safety guarantees.
• Reward Tampering: RL failure mode where observed reward diverges from true reward through signal, function, observation, or data corruption.
• Emergent Communication: learned signaling protocols developed through agent interaction rather than fixed in advance.
• Successful Misunderstandings: coordination success despite incompatible signal interpretations, often exposed by new interaction partners.
• Multi-Agent Non-Stationarity: instability in MARL caused by concurrently adapting agents changing each other’s induced environment.
• Stochastic Games: multi-agent sequential decision model with state-dependent joint actions, stochastic transitions, player-specific objectives, and equilibrium reasoning.
• Cooperative Multi-Agent Reinforcement Learning: shared-objective MARL where agents coordinate through local observations, policies, and actions.
• Multi-Agent Coordination: choosing compatible actions, roles, subtasks, messages, or influence strategies across multiple agents.
• Ad Hoc Teamwork: cooperation with teammates that were not known, trained with, or pre-coordinated with before deployment.
• Cooperative MARL Benchmarking: evaluation of shared-objective MARL algorithms across task families, observation modalities, variance, and training cost.
• Social Learning in MARL: direct reward-reshaping and incentive channels between learning agents, including deception risks.
• Temporal Difference Learning: bootstrapped value-learning methods that update from later value estimates.
• Q-Learning: off-policy temporal-difference control algorithm for learning optimal action values from sampled transitions.
• Value Decomposition Networks: additive team-value factorization for cooperative centralized-training/decentralized-execution MARL.
• Multi-Agent PPO: PPO-style cooperative-MARL baselines using independent learners or decentralized actors with centralized critics.
• Mean-Field Reinforcement Learning: large-population MARL approximation using a representative agent and state-distribution dynamics.
• Constrained Markov Potential Games: constrained multi-agent games where unilateral incentives align with a potential while policies must satisfy coupled safety constraints.
• Causal Decision Making: decision framework that evaluates actions and policies by identifiable intervention effects rather than observed associations alone.
• Counterfactual Simulation: sampling counterfactual distributions from structural causal models when analytic conditioning is difficult.
• Causal Games: game-theoretic causal models that distinguish prediction, intervention, and counterfactual queries over rational outcomes.
• Multi-Agent Influence Diagrams: DAG-based graphical game models for representing decision contexts, utilities, policies, and strategic relevance.
• Structural Equation Models: causal model language representing variables as structural assignments that can be replaced under interventions.
• Actual Causality: analysis of whether an event caused an outcome in a specific actual model/context.
• Temporal Causal Models: structural causal models with explicit time-indexed update equations and temporal interventions.
• Sigmoid Functions: S-shaped bounded functions used as activations, CDFs, and saturating growth curves.
• S-Curves and Saturating Growth: growth models that look locally exponential while approaching a finite capacity or saturation level.
• Universal Approximation Theorem: neural-network expressivity result for approximating continuous functions under suitable architecture and activation assumptions.
• Double Descent: nonmonotone test-error behavior around interpolation and overparameterization.
• Beta Distribution: two-parameter probability family on [0,1] used for proportions and Bayesian uncertainty over probabilities.
• EXPTIME: deterministic exponential-time complexity class used to mark hard worst-case decision procedures.
• Floyd’s Sampling Algorithm: exact k-subset sampling procedure that avoids shuffling the full population.
• Segmented Iterators: iterator abstraction that exposes segment/local structure so generic algorithms can exploit contiguous local ranges.
• Shielding: runtime intervention that blocks unsafe actions while preserving as much learner autonomy as possible.
• Probabilistic Shielding: runtime safety filtering with live probability or expected-cost budgets for stochastic RL constraints.
• Shield Synthesis: formal construction of runtime shields that correct unsafe reactive-system outputs with minimum interference.
• Explainable Shielding: methods for explaining shield decisions through risk categories, allowed-action trees, and unsafe-action evidence.
• Probabilistic Controlled Invariant Sets: safe sets from which a controller can keep an MDP inside a safe region with high probability over a horizon.
• Hamilton-Jacobi Reachability: control-theoretic safe-set analysis where the sign of a reachability value function separates safety-preserving states from failure-inevitable states.
• Automata-Theoretic Logic: automata/logic correspondence layer where formulas define languages and automata provide decision procedures.
• Second-Order Logic: logic that quantifies over properties, relations, and functions, with full/Henkin semantic tradeoffs.
• Higher-Order Logic: typed extension of second-order logic to functions, predicates, and objects over lower types.
• Lean Theorem Prover: proof assistant and functional language for machine-checked proof terms, tactics, and formalized mathematics.
• Linear Temporal Logic: temporal specification language for expressing safety and ordering constraints over trajectories.
• Duration Calculus: interval temporal logic for real-time requirements involving accumulated durations of conditions.
• Safety and Co-Safety Properties: temporal language classes where violation or satisfaction is witnessed by finite prefixes.
• Reactive Synthesis: automatic construction of controllers that satisfy temporal specifications against all environment behaviors.
• Safety Games: two-player games where the controller wins by keeping execution inside a safe region forever.
• Emerson-Lei Objectives: regular infinite-duration game objectives expressed with Boolean combinations of infinitely/finitely recurring colors.
• Probabilistic Strategic Timed CTL: branching-time logic combining coalition ability, real-time bounds, and probability thresholds.
• Quadtrees: adaptive spatial partitioning structure for pruning 2D queries.
• Nearest-Neighbor Search: query primitive for retrieving the closest points under a distance metric.
• Situation Calculus: first-order action formalism for dynamic domains and executable action sequences.
• Runtime Verification: monitoring system traces against formal specifications during execution.
• Predicate Abstraction: finite abstraction technique for turning rich state spaces into tractable logical summaries.
• Invariant Synthesis: construction of state or strategy invariants that remain true across execution.
• Safe Reinforcement Learning: RL methods that respect safety constraints during learning or deployment.

Sources

• Alshiekh2018 - Safe Reinforcement Learning via Shielding: partial-ingest paper note for shielding as a temporal-logic-based safety wrapper around RL.
• Fisac2019 - Bridging Hamilton-Jacobi Safety Analysis and Reinforcement Learning: partial-ingest note on a discounted Safety Bellman Equation that makes Hamilton-Jacobi safety analysis amenable to RL algorithms.
• Neary2022 - Verifiable and Compositional Reinforcement Learning Systems: partial-ingest note on pMDP-based decomposition of task reachability specifications into verifiable RL subsystem success thresholds.
• HamelDeLeCourt2025 - Probabilistic Shielding for Safe Reinforcement Learning: partial-ingest note on known-safety-dynamics probabilistic shielding via risk-budget state augmentation and certified unsafe-reachability upper bounds.
• HamelDeLeCourt2025 - ProSh Probabilistic Shielding for Model-free Reinforcement Learning: partial-ingest note on model-free CMDP shielding with a learned backup cost critic and differentiable policy-distribution shield.
• Achiam2017 - Constrained Policy Optimization: partial-ingest note on CMDP trust-region policy search with per-iteration near-constraint satisfaction.
• Everitt2019 - Towards Safe Artificial General Intelligence: partial-ingest thesis note on formal goal alignment, reward tampering, corrupt reward MDPs, sequential decision theory, and corrigibility.
• Dalrymple2024 - Towards Guaranteed Safe AI: partial-ingest note on world-model, safety-specification, and verifier architectures for high-assurance quantitative AI safety guarantees.
• Bloem2015 - Shield Synthesis: partial-ingest note on generic reactive-system shields, minimum interference, and k-stabilizing safety-game synthesis.
• Rieder2025 - Explainably Safe Reinforcement Learning: partial-ingest note on probabilistic shield explanations using risk categories and hierarchical decision trees.
• ElSayed-Aly2021 - Safe Multi-Agent Reinforcement Learning via Shielding: partial-ingest note on centralized and factored LTL shields for safe MARL.
• Gu2022 - Multi-Agent Constrained Policy Optimisation: partial-ingest note on MACPO, MAPPO-Lagrangian, and safety-aware MuJoCo/Robosuite MARL benchmarks.
• Agrawal2026 - The Multi-Agent Off-Switch Game: partial-ingest note on individual corrigibility failing to compose into group corrigibility under non-additive multi-agent utilities.
• Jusup2024 - Safe Model-Based Multi-Agent Mean-Field Reinforcement Learning: partial-ingest note on Safe-M3-UCRL, distributional population constraints, and safe vehicle repositioning.
• Alatur2024 - Provably Learning Nash Policies in Constrained Markov Potential Games: partial-ingest note on CMPGs, failure of strong duality, and coordinate-ascent learning of Nash policies under coupled constraints.
• Alinejad2026 - Dynamic Automaton Refinement and Planning for Non-Markovian RL: partial-ingest note on DARP, confidence-weighted automata refinement, and product-MDP planning for non-Markovian RL.
• Bendor2001 - Aspiration-Based Reinforcement Learning in Repeated Interaction Games: partial-ingest overview of aspiration levels, satisfaction-driven action reinforcement, and cooperation in repeated games.
• Fink1964 - Equilibrium in a Stochastic n-Person Game: partial-ingest note on stationary mixed-strategy equilibrium existence for finite discounted stochastic n-person games.
• Brorholt2025 - Compositional Shielding and Reinforcement Learning for Multi-Agent Systems: partial-ingest note on distributed shielding, assume-guarantee local obligations, and cascading shielded learning.
• Furelos-Blanco2023 - Hierarchies of Reward Machines: partial-ingest note on hierarchical reward machines, call-stack semantics, option learning, and ILASP-based HRM induction.
• Toro Icarte2022 - Reward Machines: partial-ingest note on reward machines, MDPRMs, counterfactual experiences, HRM/QRM, and RM-based reward shaping.
• Varricchione2023 - Synthesising Reward Machines for Cooperative MARL: partial-ingest note on ATL-guided synthesis of agent-local reward machines for cooperative MARL.
• Bartocci2018 - Introduction to Runtime Verification: seeded overview source on runtime verification and monitoring.
• Latvala2002 - Efficient Model Checking of Safety Properties: partial-ingest note on finite automata for LTL safety bad prefixes, pathological formulas, and scheck experiments.
• Zhu2020 - A Symbolic Approach to Safety LTL Synthesis: partial-ingest note on Safety LTL synthesis via bad-prefix automata, Horn-SAT, and symbolic safety games.
• Artale2023 - Complexity of Safety and coSafety Fragments of Linear Temporal Logic: partial-ingest note on satisfiability, validity, and realizability complexity for safety/co-safety LTL fragments.
• Hausmann2024 - Symbolic Solution of Emerson-Lei Games for Reactive Synthesis: partial-ingest note on Zielonka-tree fixpoint solving for Emerson-Lei games and Safety-and-EL synthesis.
• Cerexhe2011 - Executability in the Situation Calculus: partial-ingest paper note on executability and automata-theoretic expressivity in situation calculus.
• Hopkins2005 - Causality and Counterfactuals in the Situation Calculus: partial-ingest note on embedding situation-calculus histories into structural causal models.
• Liu2026 - A Counterfactual Cause in Situation Calculus: partial-ingest note on cause definitions for situation-calculus narratives using necessary element sets.
• Elsayed-Aly2024 - Distributional Probabilistic Model Checking: partial-ingest note on distributional reward queries, forward DTMC distribution generation, and DVI for expected/CVaR objectives.
• Quatmann2018 - Sound Value Iteration: partial-ingest note on certified value-iteration bounds for MC/MDP reachability probabilities and expected rewards.
• Dolgorukov2024 - Dynamic Epistemic Logic of Resource Bounded Information Mining Agents: partial-ingest note on semi-public query actions and resource-bounded epistemic update.
• Galimullin2025 - Changing the Rules of the Game: partial-ingest note on ATL model-building actions that dynamically revise game graphs.
• Gladyshev2025 - Temporal Causal Reasoning with Non-Recursive SEMs: partial-ingest note on temporal SEM syntax, semantics, interventions, and causal reasoning over traces.
• Gladyshev2026 - Temporal Causal Models as a Model of Computation: partial-ingest note on TCMs as computation models with Turing-completeness-style expressivity.
• Ge2026 - A Review of Causal Decision Making: partial-ingest JAIR survey note on CSL, CEL, CPL, six causal decision-making paradigms, assumption violations, and causal views of OPE, bandits, and RL.
• Karvanen2024 - Simulating Counterfactuals: partial-ingest JAIR note on simulating counterfactual distributions in known SCMs via conditional simulation, particle filtering, and fairness audits.
• Fox2024 - Causality and Strategic Reasoning: partial-ingest thesis note on causal games, mechanised MAIDs, subgames, imperfect recall, and causal queries in strategic settings.
• Hofmann2025 - Automata Theory and Logic: partial-ingest textbook note on finite/infinite-word automata, MSO, LTL model checking, tree automata, and parity games.
• Goranko2023 - Temporal Logics: partial-ingest Cambridge Elements survey on models of time, Prior tense logic, LTL, branching-time semantics, first-order temporal logic, and applications.
• Vaananen2024 - Second-Order and Higher-Order Logic: partial-ingest SEP note on second-order quantification, full versus Henkin semantics, categoricity, MSO decidability, finite model theory, and higher-order/type-theoretic logic.
• Karunus2026 - All Lean Books And Where To Find Them: partial-ingest resource guide connecting Lean learning paths to proof terms, tactics, inductive types, Mathlib-style formalization, and metaprogramming.
• AbuseOfNotation2026 - The Case Against Boolean Logic: partial-ingest essay note on Boolean thinking, context, and intuitionistic logic as a proof-oriented alternative.
• AISafetyCom2026 - Map of AI Existential Safety: seeded directory-style note on organizations, programs, funders, media, and training resources in the AI safety ecosystem.
• Ghasemi2025 - Toward Virtuous Reinforcement Learning: seeded critique and roadmap for virtue-oriented RL and machine ethics.
• Varshney2026 - An Algebraic Exposition of the Theory of Dyadic Morality: partial-ingest note on SCM-style dyadic morality, psychological operators, node compression, and patient-centric AI policy design.
• Guan2026 - Open Sourcing Monitorability Evaluations: partial-ingest note on CoT monitorability datasets, g-mean^2, and cross-fit filtering for intervention evals.
• Kondylidis2025 - Successful Misunderstandings: Learning to Coordinate Without Being Understood: partial-ingest note on emergent communication protocols that coordinate successfully while hiding non-shared interpretations.
• Hashimoto2026 - Data-Driven Synthesis of Probabilistic Controlled Invariant Sets for Linear MDPs: partial-ingest note on data-derived PCIS safe sets, conservative predecessor operators, and runtime shields for unknown linear MDPs.
• GrowingSWE2026 - An Interactive Intro to Quadtrees: interactive blog walkthrough of quadtrees, range queries, pruning, and nearest-neighbor search.
• Jamroga2026 - Towards Probabilistic Strategic Timed CTL: partial-ingest note on branching-time probabilistic strategic reasoning in continuous-time asynchronous multi-agent systems.
• Luo2022 - Automated Synthesis of Generalized Invariant Strategies: partial-ingest paper note on CEGIS-style invariant-strategy synthesis.
• Llb4ll2026 - k-Nearest-Neighbor Search using D3 Quadtrees (Gist): compact implementation-oriented note on quadtree k-NN search.
• Llb4ll2026 - k-Nearest-Neighbor Search using D3 Quadtrees (Observable): code-heavy notebook clip for visualizing quadtree traversal and pruning.
• LeroyOd2019 - Universal Approximation Theorem Proof Pointers: clipped Q&A mapping classical, architectural, and quantitative universal-approximation proof references.
• LessWrong2026 - All Exponentials are Eventually S-Curves: clipped essay on finite-resource growth, stacked S-curves, and limits of pure exponential extrapolation.
• Phan2026 - Humanity’s Last Exam: partial-ingest note on a hard frontier benchmark with calibration-aware reporting.
• MacKinlay2026 - Opponent Shaping as a Model for Manipulation and Cooperation: partial-ingest note on LOLA, Advantage Alignment, and strategic influence in MARL.
• Parker2024 - Responsibility in a Multi-Value Strategic Setting: partial-ingest note on multi-value responsibility attribution and anticipation, LTLf value bases, symbolic regret, and non-dominated strategy choice.
• Papoudakis2019 - Dealing with Non-Stationarity in Multi-Agent Deep Reinforcement Learning: partial-ingest survey note on centralized critics, self-play, opponent modelling, meta-learning, communication, and open MARL non-stationarity problems.
• Rother2023 - Disentangling Interaction using Maximum Entropy Reinforcement Learning in Multi-Agent Systems: partial-ingest note on coexistence environments, impact-aware policies, and entropy/JSD task-impact policy blending.
• Wang2020 - Too Many Cooks: Coordinating Multi-agent Collaboration Through Inverse Planning: partial-ingest note on Bayesian Delegation, inverse-planning subtask inference, and decentralized cooking-task coordination.
• Ahmed2022 - Deep Reinforcement Learning for Multi-Agent Interaction: partial-ingest portfolio overview of cooperative MARL benchmarks, ad hoc teamwork, emergent communication, and agent modelling.
• Papadopoulos2025 - An Extended Benchmarking of Multi-Agent Reinforcement Learning Algorithms in Complex Fully Cooperative Tasks: partial-ingest benchmark note on PyMARLzoo+, fully cooperative task diversity, image observations, and training-time reporting.
• Chelarescu2021 - Deception in Social Learning: partial-ingest review note on social-incentive MARL, reward-reshaping channels, and deception risk.
• Watkins1992 - Q-Learning: partial-ingest note on tabular Q-learning, action-value optimality, and almost-sure convergence conditions.
• Dole2023 - Correct-by-Construction Reinforcement Learning of Cardiac Pacemakers from Duration Calculus Requirements: partial-ingest note on DC pacemaker requirements, stopwatch automata, safety games, and shielded adaptive pacing.
• Sunehag2017 - Value-Decomposition Networks for Cooperative Multi-Agent Learning: partial-ingest note on VDN, additive team-value factorization, lazy-agent failure modes, and cooperative gridworld experiments.
• Yu2022 - The Surprising Effectiveness of PPO in Cooperative Multi-Agent Games: partial-ingest benchmark note on MAPPO/IPPO, cooperative MARL implementation details, and PPO baseline strength.
• Gallici2025 - Simplifying Deep Temporal Difference Learning: partial-ingest note on TD stability, LayerNorm plus regularization, and PQN/PQN-VDN without target networks or large replay buffers.
• MIT6041 - Probability Counting: lecture-note ingest on finite sample spaces, binomial coefficients, binomial probabilities, and conditional counting.
• Pearl2026 - Structural Equation Models and Intervention Assumptions Thread: short thread note on SEM assumptions and intervention-guiding usefulness.
• PredictAddict2026 - Calibration vs Refinement Thread: short synthesis on why calibration and informativeness come apart.
• RDI2026 - How We Broke Top AI Agent Benchmarks: partial-ingest note on exploit-driven benchmark failure modes and adversarial evaluation design.
• Rutherford2024 - JaxMARL: partial-ingest note on JAX-native MARL environments and baselines for reproducible experimentation.
• BoostedCpp2026 - Neoclassical C++ Segmented Iterators Revisited: partial-ingest blog note on segmented iterators, hierarchical algorithms, Boost.Container experiments, and compiler-dependent vectorization gains.
• Varricchione2024 - Pure-Past Action Masking: seeded paper note on action masking with temporally specified safety information.
• Vinzent2026 - Probabilistic Safety Verification of Neural Policies via Predicate Abstraction: seeded paper note on abstraction-based neural policy verification.
• Jaffray2026 - Floyd’s Sampling Algorithm: partial-ingest note on exact subset sampling via Floyd’s incremental selection rule and its Fisher-Yates connection.
• Wikipedia2026 - Beta Distribution: partial-ingest reference note on the beta family for proportions and conjugate Bayesian updating.
• Wikipedia2026 - EXPTIME: partial-ingest reference note on deterministic exponential time and its role in worst-case tractability.
• Wikipedia2026 - Brier Score: reference note on proper scoring, calibration, and Brier-score decomposition.
• Wikipedia2026 - Sigmoid Function: partial-ingest reference note on sigmoid curves, logistic activations, CDFs, and S-shaped growth functions.
• Wikipedia2026 - Double Descent: partial-ingest reference note on interpolation, overparameterization, and nonmonotone generalization curves.
• Zhang2025 - Trustworthy Reinforcement Learning under Constraints and Perturbations: seeded thesis note on trustworthy RL under constraints and perturbations.

Artifacts

• Monitorability Metrics and Cross-Fit Filtering: formal annex on intervention total-effect estimates, Wald eligibility, and sample-split filtering.
• Advantage Alignment Update: formal annex on the first-order opponent-shaping advantage update and its LOLA connection.
• Aspiration Dynamics in Repeated Games: formal annex on aspiration updates, satisfaction-driven reinforcement, and stability of repeated-game outcomes.
• Discounted Stochastic Game Equilibrium: formal annex on Fink’s contraction-plus-Kakutani existence proof for finite discounted stochastic n-person games.
• Intuitionistic Proof Rules: formal annex on constructive proof judgments, disjunction, implication, negation, and excluded-middle caveats.
• Full and Henkin Second-Order Semantics: formal annex on full second-order quantification, general models, Henkin models, and the categoricity/completeness tradeoff.
• Lean Proof Terms and Tactics: formal annex on Lean proof judgments, direct proof terms, tactic elaboration, inductive objects, and kernel checking.
• Multi-Agent Off-Switch Game: formal annex on group corrigibility, additive composition, and marginal-contribution tests for non-additive off-switch games.
• Multi-Value Responsibility Anticipation: formal annex on multi-value responsibility sets, weak excuses, symbolic regret, and non-dominated strategy selection.
• Finite Probability Counting Rules: formula annex for finite equiprobable sample spaces, binomial coefficients, and conditional counting.
• Floyd Sampling Correctness: formal annex on Floyd’s exact subset-sampling invariant and Fisher-Yates equivalence.
• Segmented Iterator Traits: formal annex on segment/local iterator decomposition and hierarchical single-range algorithm dispatch.
• EXPTIME Complexity Class: formal annex on deterministic exponential time, containment facts, and EXPTIME-hardness vocabulary.
• Structural Intervention Semantics: formal annex on SEM structural assignments and do-style intervention replacement.
• Causal Game Intervention Semantics: formal annex on pre-policy versus post-policy interventions, rational-outcome query sets, and counterfactuals in causal games.
• Dyadic Morality Algebra: formal annex on SCM-style TDM variables, psychological operators, node collapse, sequential dyads, and scoped moral-perception measurement.
• Logistic Sigmoid and Saturating Growth: formal annex on the logistic sigmoid derivative and logistic growth equation.
• Universal Approximation Statement: compact theorem-schema annex for neural-network universal approximation and its caveats.
• Q-Learning Convergence: formal annex on the tabular Q-learning update, convergence conditions, and action-replay-process proof schema.
• Regularized TD Stability and PQN: formal annex on Gallici et al.’s TD stability criterion, LayerNorm regularization argument, and PQN construction.
• Cooperative MARL Baseline Formalisms: formal annex on shared-reward DEC-POMDPs, VDN value factorization, MAPPO/IPPO, and baseline comparison risks.
• MSO Automata Correspondence: formal annex on the finite-word equivalence between regular languages and MSO-definable languages.
• Action-Aligned Interaction Learning: formal annex on AAIL impact rewards, entropy/JSD weighting, and task-impact Q-value blending.
• Bayesian Delegation: formal annex on inverse-planning posteriors over subtask allocations and BRTDP low-level coordination.
• Safety LTL Automata and Games: formal annex on bad/good prefixes, safety automata, safety-game fixpoints, and safety/co-safety complexity.
• Duration Calculus Pacemaker Shields: formal annex on bounded DC requirements, stopwatch automata, safety games, and RL shielding for pacemakers.
• Emerson-Lei Game Solving: formal annex on Emerson-Lei objectives, Zielonka trees, symbolic fixpoints, and Safety-and-EL synthesis.
• Distributional Queries and DVI: formal annex on distributional reward queries, co-safe LTL product construction, DTMC distribution generation, and risk-neutral/risk-sensitive DVI.
• Dynamic Automaton Refinement: formal annex on DARP’s confidence-weighted automaton updates, exploration bonus, and product-MDP maintenance.
• CPO Trust Region Update: formal annex on CPO’s policy-difference bound, surrogate cost constraints, and quadratic trust-region approximation.
• Discounted Safety Bellman Equation: formal annex on the contraction backup and Safety Q-learning update for Hamilton-Jacobi safety analysis.
• Corrupt Reward MDPs: formal annex on true versus observed reward, reward-corruption functions, no-free-lunch limits, and decoupled reward data.
• MACPO Sequential Trust Region: formal annex on multi-agent advantage decomposition and per-agent constrained trust-region updates.
• Safe Mean-Field UCRL: formal annex on conservative mean-field constraints under epistemic transition uncertainty.
• K-Stabilizing Shield Synthesis: formal annex on generic shield recovery windows, fail-safe mode, and safety-game reduction.
• Shield Risk Decision Trees: formal annex on risk-based shield decisions and hierarchical tree explanations.
• Probabilistic Risk-Budget Shields: formal annex on known-model and model-free probabilistic shields that maintain live risk budgets.
• Guaranteed Safe AI Safety Case: formal annex on the world-model, safety-specification, verifier, and probability-bound schema for GS AI.
• Coordinate-Ascent for CMPGs: formal annex on CMPG potential maximization, strong-duality failure, CA-CMPG, and CA-CMPG-E.
• Reward Machine Learning Methods: formal annex on MDPRMs, CRM counterfactual experiences, QRM/HRM, and potential-based reward shaping.
• Hierarchical Reward Machine Semantics: formal annex on HRM hierarchy states, calls, exit conditions, flat equivalence, and exponential blow-up.
• Centralized and Factored MARL Shielding: formal annex on joint-action safety-game shields, minimal per-agent correction, and factored join/leave coordination.
• Cooperative Reward Machine Synthesis: formal annex on synthesizing agent-local reward machines so a cooperative MAS satisfies an ATL objective.
• Distributed Shield Synthesis: formal annex on local shields, projected LTSs, extended shields, and assume-guarantee composition.
• Sound Value Iteration Bounds: formal annex on MC/MDP reachability bounds derived from step-bounded goal probability and undecided-state mass.
• ICRL pMDP Specification Decomposition: formal annex on decomposing probabilistic task specifications into subsystem success thresholds with a high-level pMDP.
• Executability: technical annex on exec(do(a, s)), BAT acceptance, and the DFA correspondence for literal-based situation-calculus theories.
• Situation Calculus Causal Models: formal annex on constructing structural causal models from situation-calculus action theories and histories.
• Temporal Structural Equation Models: formal annex on temporal SEM traces, interventions, finite-support assumptions, and cyclic temporal influence.
• Causal Decision-Making Taxonomy: formal annex on CSL/CEL/CPL, six offline/online data paradigms, and the core causal identification assumptions for policy learning.
• Counterfactual Particle Filtering: formal annex on SCM counterfactual simulation by root-finding, resampling, and sequential Monte Carlo convergence.
• Semi-Public Query Logic: formal annex on resource-bounded dynamic epistemic query actions.
• ATL Model Building: formal annex on rule-changing ATL model-building actions.
• Brier Decomposition: formal annex on the Brier score equations, Murphy decomposition, and skill-score interpretation.
• Invariant Strategy Verification Conditions: technical note on the postdiction strategy schema and the first-order conditions used to verify invariant strategies.
• Counterexample-Guided Strategy Refinement: refinement-loop annex connecting counterexamples, finite game instances, and model-checking-based strategy repair.
• PSTCTL Semantics: formal annex collecting the core syntax and strategic probability semantics for Probabilistic Strategic Timed CTL.
• Branching-Time Semantics: formal annex for Peircean and Ockhamist truth clauses over histories in branching-time temporal logic.
• PCIS Safety Predecessor: formal annex on conservative PCIS predecessor recursion, sample-splitting certification, and safe action maps.

Outputs

• Shielding, Temporal Logic, and Safe MARL: starter filed synthesis showing how query answers can become reusable wiki artifacts.

Raw Intake

• raw/2410.17229v2.pdf: partial-ingested into a source note.
• raw/2024/06/Simulating Counterfactuals.pdf: partial-ingested into a source note.
• raw/2024/06/Towards Guaranteed Safe AI.pdf: partial-ingested into a source note.
• raw/2026/05/Complexity of Safety and coSafety Fragments of Linear Temporal Logic.pdf: alternate arXiv rendering of Artale2023 - Complexity of Safety and coSafety Fragments of Linear Temporal Logic; staged from root raw intake and not ingested separately.
• raw/2026/05/achiam17a.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/2110.02793v2.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/1501.02573v2.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/2306.17052v2.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/NeurIPS-2025-explainably-safe-reinforcement-learning-Paper-Conference.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/ICLR-2025-simplifying-deep-temporal-difference-learning-Paper-Conference.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/1706.05296v1.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/2103.01955v4.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/Latvala.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/1709.07495v2.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/elsynt.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/08719-AAAI23.ArtaleA.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/17211-77299-1-PB.pdf: duplicate PDF of Alshiekh2018 - Safe Reinforcement Learning via Shielding; left untouched and not ingested.
• raw/2026/05/03193-AAAI24.VarricchioneG-SRRAI.pdf: duplicate PDF of Varricchione2024 - Pure-Past Action Masking; left untouched and not ingested.
• raw/2026/05/03193-AAAI24.VarricchioneG-SRRAI-1.pdf: duplicate PDF of Varricchione2024 - Pure-Past Action Masking; left untouched and not ingested.
• raw/2026/05/An Algebraic Exposition of the Theory of Dyadic Morality.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/Aspiration_based_reinforcement_learning.pdf: partial-ingested into a source note after staging from root raw intake.
• raw/2026/05/Abuse of Notation - writings on math, logic, philosophy and art - The case against boolean logic (httpsabuseofnotation.github.ioboolean-thinking).md: partial-ingested into a source note.
• raw/2026/05/Field map (httpsaisafety.commap).md: seeded as a directory-style source note; not expanded into technical concept pages.
• raw/2026/05/multi-agent-off-switch.pdf: partial-ingested into a source note.
• raw/2026/05/2509.24660v1.pdf: partial-ingested into a source note.
• raw/2026/05/2604.02727v1.pdf: partial-ingested into a source note.
• raw/2026/05/Safe Multi-Agent Reinforcement Learning via Shielding.pdf: partial-ingested into a source note.
• raw/2024/12/2019 ICRA - Bridging Hamilton-Jacobi Safety Analysis and Reinforcement Learning.pdf: partial-ingested into a source note.
• raw/2024/10/Tom Everitt Thesis 2018.pdf: partial-ingested into a source note.
• raw/2024/10/Fox_2024_Causality_and_strategic.pdf: partial-ingested into a source note.
• raw/2026/05/Papoudakis2019 - Dealing with Non-Stationarity in Multi-Agent Deep Reinforcement Learning.pdf: partial-ingested into a source note.
• raw/2025/10/Equilibrium in a Stochastic n-Person Game.pdf: partial-ingested into a source note.
• raw/2025/09/Verifiable and Compositional Reinforcement Learning Systems.pdf: partial-ingested into a source note.
• raw/2026/05/Alatur2024 - Provably Learning Nash Policies in Constrained Markov Potential Games.pdf: partial-ingested into a source note.
• raw/2026/05/Toro Icarte2022 - Reward Machines.pdf: partial-ingested into a source note.
• raw/2026/05/Compositional Shielding and Reinforcement Learning for_Multi-Agent Systems.pdf: partial-ingested into a source note.
• raw/2026/05/Varricchione2023 - Synthesising Reward Machines for Cooperative MARL.pdf: partial-ingested into a source note.
• raw/2026/05/Liu2026 - A Counterfactual Cause in Situation Calculus.pdf: partial-ingested into a source note.
• raw/2026/05/Hopkins2005 - Causality and Counterfactuals in the Situation Calculus.pdf: partial-ingested into a source note.
• raw/2026/05/Dolgorukov2024 - Dynamic Epistemic Logic of Resource Bounded Information Mining Agents.pdf: partial-ingested into a source note.
• raw/2026/05/Galimullin2025 - Changing the Rules of the Game.pdf: partial-ingested into a source note.
• raw/2026/05/Gladyshev2026 - Temporal Causal Models as a Model of Computation.pdf: partial-ingested into a source note.
• raw/2026/05/Gladyshev2025 - Temporal Causal Reasoning with Non-Recursive SEMs.pdf: partial-ingested into a source note.
• raw/2026/05/All Exponentials are Eventually S-Curves (httpswww.lesswrong.compostsyLByrXDfjQbKuF4b9all-exponentials-are-eventually-s-curves).md: partial-ingested into a source note.
• raw/2026/05/Sigmoid function (httpsen.wikipedia.orgwikiSigmoid_function).md: partial-ingested into a source note.
• raw/2026/05/Double descent (httpsen.wikipedia.orgwikiDouble_descent).md: partial-ingested into a source note.
• raw/2026/05/Where can I find the proof of the universal approximation theorem (httpsai.stackexchange.comquestions13317where-can-i-find-the-proof-of-the-universal-approximation-theorem).md: partial-ingested into a source note.
• raw/2026/05/Distributional Probabilistic Model Checking.pdf: partial-ingested into a source note.
• raw/2026/05/Furelos-Blanco2023 - Hierarchies of Reward Machines.pdf: partial-ingested into a source note.
• raw/2026/05/Probabilistic Shielding for Safe Reinforcement Learning.pdf: partial-ingested into a source note.
• raw/2025/11/ProSh.pdf: partial-ingested into a source note.
• raw/2026/05/Sound Value Iteration.pdf: partial-ingested into a source note.
• raw/2026/05/ecai_2023.pdf: partial-ingested into a source note.
• raw/2026/05/p2032.pdf: partial-ingested into a source note.
• raw/2026/05/2208.01769v1.pdf: partial-ingested into a source note.
• raw/2026/05/2502.04773v2.pdf: partial-ingested into a source note.
• raw/2026/05/2106.05402v1.pdf: partial-ingested into a source note.
• raw/2026/05/BF00992698.pdf: partial-ingested into a source note.
• raw/2026/05/08450-AAAI23.DoleK-SR.pdf: partial-ingested into a source note.
• raw/2026/05/3-540-48683-6.pdf: CAV 1999 proceedings volume with mostly garbled extracted text; not ingested as a single source, use targeted paper-level ingest if needed.
• raw/2026/05/Neoclassical C++ segmented iterators revisited (1) (httpsboostedcpp.net20260518neoclassical-c-segmented-iterators-revisited-1).md: partial-ingested into a source note.
• raw/2026/05/Second-order and Higher-order Logic (Stanford Encyclopedia of Philosophy) (httpsplato.stanford.eduentrieslogic-higher-order).md: partial-ingested into a source note.
• raw/2026/05/All Lean Books And Where To Find Them (httpslakesare.brick.doall-lean-books-and-where-to-find-them-x2nYwjM3AwBQ).md: partial-ingested into a source note.
• raw/2026/04/Dynamic_Automaton_Refinement_and_Planning_for_Non-Markovian_RL.pdf: partial-ingested into a source note.
• raw/2026/04/A Review of Causal Decision Making.pdf: partial-ingested into a source note.
• raw/2026/04/Open Sourcing Monitorability Evaluations (httpsalignment.openai.commonitorability-evals).md: partial-ingested into a source note.
• raw/2026/04/MIT 6.041 Probability Counting (httpsickma2311.github.ioMathProbabilitycounting.html).md: partial-ingested into a source note.
• raw/2026/04/“Opponent shaping” as a model for manipulation and cooperation – The Dan MacKinlay stable of variably-well-consider’d enterprises (httpsdanmackinlay.namenotebookopponent_shaping.htmlapp-game-theory).md: partial-ingested into a source note.
• raw/2026/04/Thread by @yudapearl (httpsx.comyudapearlstatus2044015287767577048).md: partial-ingested into a source note.
• raw/2026/04/Visualizing Algorithms (httpsbost.ocks.orgmikealgorithms).md: malformed duplicate clip of the quadtree notebook content; re-clip before ingest.
• raw/2026/04/(FLAIROx-JaxMARL) Multi-Agent Reinforcement Learning with JAX (httpsgithub.comFLAIROxJaxMARL).md: partial-ingested into a source note.
• raw/2026/04/Humanity's Last Exam (httpsagi.safe.ai).md: partial-ingested into a source note.
• raw/2026/04/Automated Synthesis of Generalized Invariant Strategies via\nCounterexample-Guided Strategy Refinement.pdf: partial-ingested into a source note.
• raw/2026/04/Brier score - Wikipedia (httpsen.wikipedia.orgwikiBrier_score).md: partial-ingested into a source note.
• raw/2026/04/An interactive intro to quadtrees (httpsgrowingswe.comblogquadtrees).md: partial-ingested into a source note.
• raw/2026/04/k-nearest-neighbor search using D3 quadtrees (httpsgist.github.comllb4ll8709363).md: partial-ingested into a source note.
• raw/2026/04/k-nearest-neighbor search using D3 quadtrees (httpsobservablehq.com@llb4llk-nearest-neighbor-search-using-d3-quadtrees).md: partial-ingested into a source note.
• raw/2026/04/Thread by @predict_addict (httpsx.compredict_addictstatus2042667211618193600).md: partial-ingested into a source note.
• raw/2026/04/Beta distribution - Wikipedia (httpsen.wikipedia.orgwikiBeta_distribution).md: partial-ingested into a source note.
• raw/2026/04/Center for Responsible, Decentralized Intelligence at Berkeley (httpsrdi.berkeley.edublogtrustworthy-benchmarks-cont).md: partial-ingested into a source note.
• raw/2026/04/EXPTIME - Wikipedia (httpsen.wikipedia.orgwikiEXPTIME).md: partial-ingested into a source note.
• raw/2026/04/Executability in the Situation Calculus.pdf: partial-ingested into a source note.
• raw/2026/04/Floyd's Sampling Algorithm (httpsbuttondown.comjaffrayarchivefloyds-sampling-algorithm).md: partial-ingested into a source note.
• raw/2026/04/Floyd's Sampling Algorithm (httpsbuttondown.comjaffrayarchivefloyds-sampling-algorithm) 1.md: empty duplicate clip; left untouched and not ingested.
• raw/2026/04/Introduction to Runtime Verification.pdf: seeded into a source note.
• raw/2026/04/Probabilistic Safety Verification of Neural Policies via Predicate Abstraction.pdf: seeded into a source note.
• raw/2026/04/Pure-Past Action Masking.pdf: seeded into a source note.
• raw/2026/04/Safe Reinforcement Learning via Shielding.pdf: partial-ingested into a source note.
• raw/2026/04/Towards Probabilistic Strategic Timed CTL.pdf: partial-ingested into a source note.
• raw/2026/04/Toward virtuous reinforcement learning, a critique and roadmap.pdf: partial-ingested into a source note.
• raw/2026/04/Zhang-2025-thesis.pdf: seeded into a source note.
• raw/2025/11/Automata Theory and Logic.pdf: partial-ingested into a source note.