0001

Where computation stops: Church’s unsolvable problem

Alonzo Church’s 1936 boundary on uniform mechanical decision, rebuilt through lambda conversion, Gödel numbering and the diagonal breach.

Published 0002

A machine with no foresight: Turing’s computable numbers

Alan Turing’s 1936 analysis of mechanical procedure, rebuilt through finite control, machine descriptions, universality and the circle-free classifier contradiction.

Published 0003

How much uncertainty can a channel carry?

Claude Shannon’s 1948 theory rebuilt through entropy, typical sequences, equivocation, channel capacity and the exact boundary between low error and zero error.

Published 0004

A judge behind a screen: Turing’s imitation game

Alan Turing’s 1950 proposal rebuilt through the imitation game, digital universality, nine families of objection and the child-machine learning programme.

Published 0005

Prediction without a trusted model: Solomonoff’s universal prior

Ray Solomonoff’s 1964 induction theory rebuilt through universal descriptions, exponential programme weights, Bayesian mixtures and the exact limits of machine choice and approximation.

Published 0006

How many bits make an object simple? Kolmogorov’s three complexities

Kolmogorov’s 1965 separation of combinatorial, probabilistic and algorithmic information, rebuilt through conditional description length, universal simulation and the exact limits of incomputability and runtime.

Published 0007

When a string outruns every short machine: Chaitin’s 1966 measure of patternlessness

Chaitin’s 1966 state-complexity account of finite-string patternlessness, rebuilt through programme-table normalisation, counting bounds, bounded transfers and the impossibility of deciding exact shortest descriptions.

Published 0008

When verification becomes satisfiability: Cook’s universal reduction

Cook’s 1971 completeness theorem rebuilt through polynomial query machines, bounded computation tableaux, local consistency clauses, D3 splitting, subgraph embeddings and the exact boundary of P versus NP.

Published 0009

When a programme knows where it ends: Chaitin’s self-delimiting information

Chaitin’s 1975 programme-size theory rebuilt through prefix-free machines, Kraft allocation, universal simulation, algorithmic probability, the coding theorem and the starred chain rule.

Published 0010

When examples become evidence: Valiant’s theory of the learnable

Valiant’s 1984 learning theory rebuilt through positive examples, partial Boolean vectors, one-sided error, clause elimination, oracle-guided implicants and the boundary between sample sufficiency and feasible computation.

Published 0011

A machine that learns from useful parts: Solomonoff’s inductive inference machine

Solomonoff’s 1957 learning machine rebuilt through spatial prediction n-grams, reusable tuples and structures, five memory banks, consistency testing, utility-guided search and the boundary before universal induction.

Published 0012

How much information can a choice afford? Ortega and Braun’s bounded rationality

Ortega and Braun’s 2010 bounded-rationality framework rebuilt through logarithmic utility, free utility, reference-policy tilting, KL information cost, causal I/O streams and the exact α limits.

Published 0013

When an unknown chance must be inferred: Bayes’s inverse probability

Bayes’s 1763 inverse problem rebuilt through expectation-valued probability, the square-table experiment, binomial count curves, area-ratio conditioning and the boundary between Bayes, Price and later beta notation.

Published 0014

When rejected moves still count: the Metropolis sampling machine

The 1953 Metropolis, Rosenbluth and Teller method rebuilt through Boltzmann-frequency sampling, rejected self-loops, pairwise probability flow, periodic hard disks and Hastings’s asymmetric-proposal correction.

Published 0015

When pairwise dependence builds the best tree: Chow and Liu’s approximation

Chow and Liu’s 1968 dependence-tree approximation rebuilt through mutual information, Kullback divergence, maximum spanning trees, root invariance, higher-order failure and finite-sample uncertainty.

Published 0016

When the proposal has a direction: Hastings’s Markov-chain correction

Hastings’s 1970 sampler rebuilt through asymmetric proposals, detailed-balance repair, normalising-constant cancellation, rejected self-loops, correlated error and proposal-dependent mixing.

Published 0017

When every unit has two futures: Rubin’s causal effects

Rubin’s 1974 causal framework rebuilt through trial-specific potential outcomes, the missing counterfactual, exact randomisation sets, matching, observational adjustment and the separate problem of transport.

Published 0018

When hidden cases become weighted counts: Dempster, Laird and Rubin’s EM algorithm

Dempster, Laird and Rubin’s 1977 EM framework rebuilt through incomplete-data maps, expected sufficient statistics, likelihood ascent, missing-information speed and convergence limits.

Published 0019

When a variable becomes irrelevant: Dawid’s conditional-independence calculus

Dawid’s 1979 framework rebuilt through information irrelevance, symmetry, a small inference calculus, Simpson’s paradox, sufficiency, selection, invariance and causal limits.

Published 0020

When many covariates become one: Rosenbaum and Rubin’s propensity score

Rosenbaum and Rubin’s 1983 framework rebuilt through balancing scores, strong ignorability, overlap, matching, subclassification, covariance adjustment and modern design limits.

Published 0021

When an image becomes an energy landscape: Geman and Geman’s Gibbs restoration

Geman and Geman’s 1984 framework rebuilt through Markov random fields, Gibbs energies, latent line processes, stochastic relaxation, annealing guarantees and finite-time limits.

Published 0022

When evidence travels both ways: Pearl’s belief-propagation network

Pearl’s 1986 framework rebuilt through Bayesian factorisation, path separation, predictive and diagnostic messages, exact polytree inference, loop boundaries and causal-tree structuring.

Published 0023

When missing data become a route to the posterior: Tanner and Wong’s data augmentation

Tanner and Wong’s 1987 framework rebuilt through posterior mixtures, fixed-point operators, imputation, Markov chains, convergence theory, multimodality and augmentation design.

Published 0024

When loops become a tree of cliques: Lauritzen and Spiegelhalter’s local computation

Lauritzen and Spiegelhalter’s 1988 framework rebuilt through moralisation, triangulation, clique trees, separator calibration, evidence propagation and exact-inference complexity.

Published 0025

When probability follows arrows: Charniak’s Bayesian networks without tears

Charniak’s 1991 tutorial rebuilt through the family-out network, d-separation, local factorisation, exact and approximate inference, applications, assessment and causal boundaries.

Published 0026

When observation must answer an intervention: Pearl’s causal diagrams

Pearl’s 1995 framework rebuilt through structural interventions, identification, back-door and front-door criteria, do-calculus, nonidentifying patterns and the published debate over causal assumptions.

Published 0027

When an equation is allowed to mean action: Pearl’s causal calculus

Pearl’s verified 1995 AISTATS paper rebuilt through two conditioning operators, structural-equation semantics, graph-licensed transformations, surrogate experiments and imperfect-compliance limits.

Published 0028

When belief is priced by action: Ramsey’s truth and probability

Ramsey’s 1926 essay rebuilt through partial belief, utility-based measurement, betting coherence, the divide between consistency and truth, and a pragmatic defence of induction.

Published 0029

When the worst case chooses the rule: Wald’s minimax decision function

Wald’s 1945 decision theory rebuilt through loss, statewise risk, randomized rules, worst-case protection, least-favourable priors and the zero-sum game against Nature.

Published 0030

When memory activates itself: Pearl’s Bayesian network for evidential reasoning

Pearl’s 1985 architecture rebuilt through local factorisation, graph separation, source-aware messages, multidirectional evidence, exact polytree propagation and explicit strategies for loops.

Published 0031

When the function answers with noise: Robbins and Monro’s stochastic approximation

Robbins and Monro’s 1951 method rebuilt through noisy root-finding, signed residual updates, reach-and-settle gain schedules, mean-square convergence, quantile estimation and exact finite-run audits.

Published 0032

When two noisy measurements point uphill: Kiefer and Wolfowitz’s stochastic maximum

Kiefer and Wolfowitz’s 1952 method rebuilt through paired experiments, noisy finite differences, coupled gain and probe schedules, probability convergence, boundary-aware design and exact audit trails.

Published 0033

When one sample must answer every set: Vapnik and Chervonenkis on uniform convergence

Vapnik and Chervonenkis’s 1971 theory rebuilt through empirical frequencies, growth functions, shattered samples, distribution-free deviation bounds, almost-sure convergence and a live capacity laboratory.

Published 0034

When only the boundary points speak: Cortes and Vapnik’s support-vector network

Cortes and Vapnik’s 1995 learning machine rebuilt through optimal margins, active constraints, soft errors, implicit feature spaces, capacity control and a live dual-optimisation laboratory.

Published 0035

When a constraint makes coefficients vanish: Tibshirani’s Lasso

Tibshirani’s 1996 estimator rebuilt through L1 corner geometry, soft thresholding, correlated paths, prediction-error tuning, competing signal regimes and a live KKT laboratory.

Published 0036

When weak rules become a strong vote: Freund and Schapire’s AdaBoost

Freund and Schapire’s 1997 construction rebuilt through adversarial allocation, multiplicative weights, adaptive example distributions, error-weighted voting, exponential training bounds and a live AdaBoost laboratory.

Published 0037

When every advantage has a matching cost: Wolpert and Macready’s no-free-lunch theorem

Wolpert and Macready’s 1997 conservation result rebuilt through finite search, uniform function averages, problem–algorithm alignment, benchmark limits, dynamic objectives and exact exhaustive enumeration.

Published 0038

When a prior becomes a complexity budget: McAllester’s PAC-Bayesian theorems

McAllester’s 1999 bridge between informed priors and distribution-free learning rebuilt through weighted failure budgets, conditioned concept mixtures, quantifier reversal, bounded loss and an exact prior-mass laboratory.

Published 0039

When every coordinate remembers its own scale: Kingma and Ba’s Adam optimiser

Kingma and Ba’s 2014 optimiser rebuilt through paired gradient moments, zero-start correction, coordinate-wise scaling, experiment boundaries, later convergence evidence and an exact optimiser laboratory.

Published 0040

Which gradient should move the model? Ruder’s optimisation map

Ruder’s 2016–2017 survey rebuilt through batch granularity, momentum, adaptive coordinate histories, distributed staleness, auxiliary training strategies and an eight-optimiser laboratory.

Published 0041

When regularisation leaves the gradient: Loshchilov and Hutter’s AdamW

Decoupled weight decay rebuilt through SGD equivalence, adaptive-preconditioner mismatch, Adam moment separation, hyperparameter geometry, budget normalisation and an exact decay laboratory.

Published 0042

When every error becomes a direction: Friedman’s gradient boosting machine

Friedman’s 2001 gradient boosting machine rebuilt through function-space descent, pseudo-responses, loss-specific TreeBoost, shrinkage, model interpretation and an exact stagewise stump laboratory.

Published 0043

When every split must scale: Chen and Guestrin’s XGBoost system

XGBoost’s 2016 system rebuilt through regularised second-order tree scores, exact and approximate split search, Hessian-weighted quantile sketches, learnt sparse routing, cache-aware blocks and out-of-core training.

Published 0044

When velocity learns the landscape: Sutskever et al.’s initialization and momentum

Sutskever, Martens, Dahl and Hinton’s 2013 study rebuilt through classical velocity, Nesterov look-ahead, curvature eigen-directions, momentum schedules, initialization scale, recurrent spectra and an exact quadratic laboratory.

Published 0045

When memory becomes a schedule: Chen et al.’s sublinear deep-network training

Chen, Xu, Zhang and Guestrin’s 2016 system rebuilt through activation liveness, segment checkpoints, mirror counts, graph-aware planning, recursive schedules and an exact rematerialization laboratory.

Published 0046

When a neuron becomes a proposition: McCulloch and Pitts’s logical calculus

McCulloch and Pitts’s 1943 calculus rebuilt through binary threshold firing, absolute inhibition, temporal propositions, constructive synthesis, recurrent memory, computability and causal ambiguity.

Published 0047

When random connections learn a class: Rosenblatt’s probabilistic perceptron

Rosenblatt’s 1958 perceptron rebuilt through random association populations, response source-sets, value dynamics, statistical separability, differentiated environments and the boundary at relational abstraction.

Published 0048

When error learns what the hidden layer should mean: Rumelhart, Hinton and Williams’s back-propagation

Rumelhart, Hinton and Williams’s 1986 learning rule rebuilt through output discrepancy, chain-rule credit assignment, hidden-unit sensitivity, emergent representations, momentum and exact gradient checks.

Published 0049

When one hidden layer can approximate any mapping: Hornik, Stinchcombe and White’s universality

Hornik, Stinchcombe and White’s 1989 theorem rebuilt through affine ridge dictionaries, density on compacta, probability-relative approximation, Stone–Weierstrass, the cosine bridge and the boundary between capacity and learning.

Published 0050

When depth, data and GPUs met at scale: Krizhevsky, Sutskever and Hinton’s ImageNet network

The 2012 ImageNet system rebuilt through convolutional image priors, an eight-learnt-layer architecture, ReLUs, two-GPU training, normalisation, pooling, augmentation, dropout and exact result accounting.

Published 0051

When a machine learns the representation: LeCun, Bengio and Hinton’s deep-learning synthesis

The 2015 Review reconstructed through learned hierarchies, supervised objectives, chain-rule credit, convolutional priors, distributed word vectors, recurrent depth, gated memory, attention and evidence boundaries.

Published 0052

What makes a representation good? Bengio, Courville and Vincent’s new perspectives

The 2013 Review reconstructed through generic priors, the limits of local smoothness, distributed codes, deep reuse, invariance, disentangling, probabilistic models, auto-encoders, manifold geometry and unresolved criteria.

Published 0053

When every mini-batch sets the scale: Ioffe and Szegedy’s batch normalisation

The 2015 mechanism reconstructed through mini-batch moments, differentiable statistics, learned scale and shift, cross-example gradients, train–inference separation, convolutional sharing, scale properties and evidence-calibrated ImageNet results.

Published 0054

When identity takes the shortcut: He et al.’s deep residual learning

The 2015 residual-learning paper reconstructed through degradation, identity-reference corrections, projection shortcuts, bottleneck blocks, matched ImageNet comparisons, extreme CIFAR-10 depth and detection transfer.

Published 0055

When every unit must stand on its own: Srivastava et al.’s dropout

The 2014 journal paper reconstructed through Bernoulli masks, thinned shared-weight subnetworks, approximate test-time averaging, co-adaptation, optimisation interactions, multi-domain evidence and exact marginalisation boundaries.

Published 0056

When a continuous signal chooses a code: van den Oord et al.’s VQ-VAE

The 2017 discrete-representation model reconstructed through nearest-code quantisation, straight-through credit, stop-gradient loss routing, commitment, separated autoregressive priors, multi-modal evidence and an exact codebook laboratory.

Published 0057

When depth must share one update: Zhang et al.’s Fixup initialization

The 2019 normalisation-free residual parameterisation reconstructed through functional-update scale, the depth exponent, zero final layers, scalar controls, extreme-depth tests, explicit regularisation, translation transfer and an exact update-budget laboratory.

Published 0058

When each layer invents data for the next: Hinton et al.’s deep belief net

The 2006 deep belief net reconstructed through explaining away, complementary priors, the infinite tied-network equivalence, restricted Boltzmann machines, contrastive divergence, greedy layer-wise learning, up–down fine-tuning and an exact RBM laboratory.

Published 0059

When one feature can count how strongly it is present: Nair and Hinton’s rectified RBM

The 2010 rectified RBM reconstructed through tied offset-binary copies, softplus expectation, noisy rectification, contrastive divergence, intensity equivariance, cosine invariance, matched NORB and LFW comparisons and a 225-state exact unit laboratory.

Published 0060

When signals must survive every layer: Glorot and Bengio’s normalised initialisation

The 2010 diagnostic study reconstructed through sigmoid saturation, activation tails, loss plateaux, forward and backward variance products, the fan-in/fan-out compromise, Jacobian transport, condition-labelled evidence and a 972-state exact initialisation laboratory.

Published 0061

When faster correction starts to dance: Maxwell’s dynamical theory of governors

Maxwell’s 1868 paper reconstructed through moderator offset, persistent regulator state, four disturbance modes, characteristic roots, Jenkin’s cubic, damping, three governor classes, disturbance isolation and a 3,125-state exact stability laboratory.

Published 0062

When purpose becomes a feedback class: Rosenblueth, Wiener and Bigelow’s behavioural taxonomy

The 1943 essay reconstructed through input–output analysis, energy source, intrinsic purpose, feedback timing, damped correction, prediction orders, cross-cutting categories, machine–organism comparison and a 768-state exact classifier.

Published 0063

When the whole plan begins with one choice: Bellman’s dynamic programming equation

Bellman’s 1952 paper reconstructed through the first-operation decomposition, sufficient state, a contractive fixed point, successive approximation, hidden-ball search, acquisition indices, resource allocation and a 1,944-state exact laboratory.

Published 0064

When every measurement corrects a prediction: Kalman’s recursive linear filter

Kalman’s 1960 paper reconstructed through state-transition models, orthogonal projection, innovations, prediction and correction, covariance recursion, filter–regulator duality, evidence boundaries and a 256-state exact scalar laboratory.

Published 0065

When uncertainty flows continuously: Kalman and Bucy’s variance equation

Kalman and Bucy’s 1961 paper reconstructed through continuous innovation feedback, covariance intensity, the Riccati flow, finite and infinite starts, Hamiltonian lifting, stability conditions and a 1,024-state exact scalar laboratory.

Published 0066

When the first move is the only move: García, Prett and Morari’s model-predictive control

The 1989 survey reconstructed through explicit models, moving horizons, DMC and MAC, unconstrained feedback equivalence, LQC, active constraints, competing norms, nonlinear routes, robustness and a 1,024-state exact laboratory.

Published 0067

When a loop teaches its own settings: Ziegler and Nichols’s controller tuning

The 1942 tuning paper reconstructed through proportional sensitivity, amplitude ratio, automatic reset, pre-act, the ultimate cycle, reaction-curve tangent, process classification, convention safety and a 768-state exact laboratory.

Published 0068

When an action must learn what it controls: Feldbaum’s dual control

Feldbaum’s 1960 Part I reconstructed through active study, directing action, full hypothesis distributions, action and learning risk, layered adaptive architectures, finite-horizon Bayesian synthesis and a 1,024-state exact laboratory.

Published 0069

When every hidden direction must be reachable and visible: Kalman’s control theory

Kalman’s 1960 general theory reconstructed through minimal state, controllability, dead-beat control, cancellation, quadratic regulation, observability, fastest observation, algebraic duality and a 1,024-state exact laboratory.

Published 0070

When observation quality becomes a control: Zois, Levorato and Mitra’s active classification

Active classification reconstructed through one-hot Markov states, controlled Gaussian likelihoods, innovations, a Kalman-like approximate filter, posterior covariance cost, belief-state dynamic programming, smoothing and a 1,024-state exact laboratory.

Published 0071

When the shortest frontier choice becomes permanent: Dijkstra’s two graph procedures

Dijkstra’s 1959 note reconstructed through minimum spanning trees, shortest routes, permanent and tentative sets, candidate replacement, cut exchange, optimal prefixes, non-negative weights, storage discipline and a 1,024-graph exact laboratory.

Published 0072

When contradiction becomes computation: Davis and Putnam’s quantification procedure

Davis and Putnam’s 1960 proof procedure reconstructed through refutation, witness functions, ordered ground lines, conjunctive normal form, unit reduction, pure literals, resolution, layered termination and a 1,024-formula exact laboratory.

Published 0073

When the road not taken goes to tape: Davis, Logemann and Loveland’s theorem prover

The 1962 IBM 704 prover reconstructed through quantifier-free-line generation, resolution growth, the splitting rule, tape-backed depth-first search, linked occurrence tables, ready-list propagation, order sensitivity and a 1,024-formula exact laboratory.

Published 0074

When substitution and inference become one move: Robinson’s resolution principle

Robinson’s 1965 machine-oriented logic reconstructed through refutation, Herbrand saturation, disagreement repair, most general unification, standardisation apart, semi-commutativity, completeness, search principles and a 1,024-configuration exact laboratory.

Published 0075

When an estimate earns the next expansion: Hart, Nilsson and Raphael’s A*

The 1968 formal theory of heuristic path search reconstructed through implicit graphs, open and closed nodes, f̂=ĝ+ĥ, admissibility, positive-cost termination, information sets, consistency, reopening, expansion optimality, the 1972 correction and a 1,024-configuration exact laboratory.

Published 0076

When a bound makes a branch irrelevant: Knuth and Moore’s alpha-beta analysis

The 1975 analysis reconstructed through negamax, decision windows, shallow and deep cutoffs, move ordering, critical positions, the independent-leaf lower bound, random uniform trees, dependent values and a 1,024-order exact laboratory.

Published 0077

When local impossibility is removed before search: Mackworth’s network consistency

Mackworth’s 1977 framework reconstructed through backtracking thrash, node, arc and path consistency, NC-1, REVISE, AC-1 to AC-3, relation composition, minimality, PC-1 and PC-2, OR problem graphs and a 1,024-network exact laboratory.

Published 0078

When repetition costs less than memory: Korf’s depth-first iterative deepening

Korf’s 1985 search architecture reconstructed through breadth-first and depth-first limits, the geometric restart proof, asymptotic optimality, bidirectional search, IDA*, pathmax, two-person games and a 1,024-goal exact laboratory.

Published 0079

When a complete mistake knows how to repair itself: Minton et al.’s min-conflicts

The 1992 heuristic-repair method reconstructed through complete inconsistent assignments, the GDS network, greedy initialisation, n-queens, Hubble scheduling, sparse graph failure, probabilistic limits and a 1,024-case exact laboratory.

Published 0080

When pruning first makes search harder: Sabin and Freuder’s constraint-satisfaction reversal

The 1994 challenge to solver folklore reconstructed through arc consistency, dynamic minimum-domain ordering, the preprocessing reversal, forward checking, MAC, total-cost relevance and a 240-instance exact laboratory.

Published 0081

When failure writes the next rule: Marques-Silva and Sakallah’s GRASP

The 1999 conflict-driven SAT architecture reconstructed through CNF, Boolean constraint propagation, decision levels, implication graphs, resolution, UIPs, learned clauses, non-chronological backtracking and a three-scenario exact laboratory.

Published 0082

When two watched literals let a clause sleep: Moskewicz et al.’s Chaff

The 2001 SAT solver reconstructed through transitive BCP, two watched literals, no-undo watch maintenance, cache-aware access, VSIDS, lazy clause deletion, restarts, bounded benchmark evidence and a three-trace exact laboratory.

Published 0083

When a complete guess climbs one flip at a time: Selman, Levesque and Mitchell’s GSAT

The 1992 local-search method reconstructed through complete assignments, one-variable gain, sideways moves, random restarts, hard-random density, structured encodings, deceptive formulae and an exhaustive four-variable landscape laboratory.

Published 0084

When the best move needs a little noise: Selman, Kautz and Cohen’s local search

The 1994 noise-strategy comparison reconstructed through annealing, global noise, defect-focused random walk, clause-first WalkSAT, hard random SAT, structured encodings, MAX-SAT and an exact executable laboratory.

Published 0085

When one plan must work in every possible world: Bryce, Kambhampati and Smith’s belief-space heuristics

The 2006 planning framework reconstructed through belief-state distance, positive interaction, independence, overlap, planning graphs, labelled uncertainty, mutexes, CAltAlt, POND and an exact aggregation laboratory.

Published 0086

When failure names only what caused it: Kambhampati’s dynamic-CSP Graphplan

The 2000 planning-graph study reconstructed through dynamic constraint activation, explanation-based learning, dependency-directed backtracking, memo utility, search enhancements, random restarts and an exact conflict-set laboratory.

Published 0087

When time is assigned after structure: Cavrel, Pellier and Fiorino’s temporal-event planner

The 2023 HTEP proposal reconstructed through HTN refinement, snap-event compilation, causal-link flaws, non-temporal guidance, timestamp constraints, concurrency boundaries and an exact scheduling laboratory.

Published 0088

When a plan is a proof guided by distance: Soutchanski and Young’s lifted theorem prover

The 2023 TPLH proposal reconstructed through situation histories, Basic Action Theories, bounded entailment, run-time grounding, progression, delete-relaxation guidance, benchmark trade-offs and an exact proof-search laboratory.

Published 0089

When a problem chooses its own rule: Moreno-Scott et al.’s heuristic map

The 2016 CSP study reconstructed through six dynamic orderings, density–tightness geometry, phase-transition hardness, regional dominance, static and dynamic selectors, structured transfer and an exact nearest-rule laboratory.

Published 0090

When a plan repairs its own flaws: Gerevini, Saetti and Serina’s LPG

The LPG planner reconstructed through linear and temporal action graphs, causal and mutex orderings, stochastic flaw repair, relaxed-plan evaluation, numerical state, anytime improvement and an exact repair laboratory.

Published 0091

When failure becomes evidence: Clark’s negation as failure

Clark’s 1978 account reconstructed through finite failure, ground negative queries, predicate completion, free equality, soundness, incompleteness, hierarchical databases and an exact query-tree laboratory.

Published 0092

When one meaning can run many ways: Kowalski’s algorithm = logic + control

Kowalski’s 1979 analysis reconstructed through Horn clauses, reasoning direction, procedural invocation, representation, data abstraction, scheduling, alternative search, path finding and an exact control laboratory.

Published 0093

When a rule may fire unless contradicted: Reiter’s default logic

Reiter’s 1980 default logic reconstructed through consistency-conditioned rules, fixed-point extensions, normal theories, proof supports, belief revision, open defaults and an exact extension laboratory.

Published 0094

When what is not forced disappears: McCarthy’s circumscription

McCarthy’s 1980 circumscription reconstructed through qualification, predicate minimisation, minimal models, puzzle closure, induction, domain circumscription, representation choice and an exact finite laboratory.

Published 0095

When belief must survive its own assumptions: McDermott and Doyle’s non-monotonic logic

McDermott and Doyle’s 1980 logic reconstructed through modal consistency assumptions, fixed-point theories, plural belief states, linked tableaux, truth maintenance and an exact finite laboratory.

Published 0096

When knowledge predicts action: Newell’s knowledge level

Newell’s 1982 hypothesis reconstructed through computer system levels, rational action, radical incompleteness, symbolic realisation, access costs, mixed models and an exact prediction laboratory.

Published 0097

When every rule becomes a weighted feature: Richardson and Domingos’s Markov logic

Richardson and Domingos’s 2006 framework reconstructed through weighted first-order formulas, ground Markov networks, soft constraints, query-directed inference, pseudo-likelihood learning, relational tasks and an exact possible-world laboratory.

Published 0098

When descriptions find their own place: Brachman and Schmolze’s KL-ONE

The 1985 overview reconstructed through structured descriptions, primitive and defined concepts, first-class roles, monotonic inheritance, classifier-driven subsumption, contextual assertions and an exact placement laboratory.

Published 0099

When a belief must regenerate itself: Gelfond and Lifschitz’s stable models

The 1988 semantics reconstructed through grounding, candidate-dependent reduction, positive closure, minimal Herbrand models, non-stratified programmes, autoepistemic correspondence and an exact reduct laboratory.

Published 0100

When telling becomes programming: McCarthy’s Advice Taker

McCarthy’s 1959 proposal reconstructed through declarative instruction, representationally simple improvement, local deduction, property-list relevance, composed action and an exact airport-advice laboratory.

Published 0101

When change needs a place to stand: McCarthy and Hayes’s situation calculus

The 1969 programme reconstructed through representational adequacy, counterfactual ability, situations, fluents, strategy results, epistemic feasibility, persistence and an exact frame laboratory.

Published 0102

When one sequence must undo every start: Faber, Morak and Chrpa’s reversible actions

The 2021 STRIPS analysis reconstructed through exact restoration, uniform quantifiers, state-set variants, plasp translation, ASP guessing, ELP world views, saturation and an exhaustive reversal laboratory.

Published 0103

When the planner acts on what it knows: Eiter et al.’s language K

The 2001 knowledge-state planner reconstructed through typed causal rules, stable-model transition semantics, optimistic and secure plans, possible-world and partial-state encodings, conditional inertia, complexity and an exhaustive bomb-domain laboratory.

Published 0104

When one theory answers many questions: De Cat et al.’s IDP system

The modelling architecture reconstructed through FO(ID,AGG,PF,T), partial structures, well-founded definitions, reusable inference, procedural integration, ground-and-solve, hybrid propagation and an exhaustive graph laboratory.

Published 0105

When a failed state remembers its budget: Zhou, Barták and Dovier’s Picat planner

The 2015 planner reconstructed through mode-directed tabling, resource-sensitive failure, iterative reuse, structured state keys, domain commitment, admissible heuristics and an exact cyclic-graph laboratory.

Published 0106

When time flows inside a logical state: Balduccini, Magazzeni and Maratea’s PDDL+–CASP translation

The 2016 hybrid planner reconstructed through timed logical states, answer-set choices, numerical constraints, compulsory processes and events, validation-guided refinement and an exact trajectory laboratory.

Published 0107

When an open world must fit a closed planner: Borgwardt et al.’s Horn-DL compilation

The 2022 expressivity result reconstructed through eKAB semantics, ontology-mediated queries, stratified Datalog rewriting, derived predicates, guarded compilation, the Horn-ALCHOIQ/Horn-SROIQ frontier and an exact inference laboratory.

Published 0108

When a hierarchy must share one clock: Pellier et al.’s HDDL 2.1 proposal

The 2023 temporal-HTN baseline reconstructed through task identifiers, durative boundaries, four constraint families, concurrent happenings, method decomposition, state-at-event semantics and an exact schedule laboratory.

Published 0109

When actions and fluents become intervals: Ramirez et al.’s temporal-planning theory

The 2024 interval-logic planner reconstructed through motion skills, half-open time, temporally qualified assertions, three axiom families, CP-SAT flow constraints, required concurrency and an exact satisfiability laboratory.

Published