Latest / AI Sentinel: Frontier Daily / AI's Binding Constraint Is Structural, Not Scale

Transcript

AI's Binding Constraint Is Structural, Not Scale · AI Sentinel: Frontier Daily

• 0:00
  AI's Binding Constraint Is Structural, Not Scale.
• 0:03
  Highlights.
• 0:04
  Formal analysis exposes inherent structural flaws in deployed AI methods that cannot be remedied by incremental scaling, revealing design assumptions as the binding constraint on progress.
• 0:15
  Multimodal AI systems illustrate that single-modality representations are structurally incomplete, resulting in systematic and predictable failures that additional modalities cannot simply patch.
• 0:25
  Embodied and agentic AI benchmarks cannot causally attribute performance to claimed capabilities because evaluation methodology systematically conflates confounding factors with genuine competence.
• 0:36
  Activation-level mechanistic safety interventions harbor inherent structural blind spots,
• 0:41
  as formal evidence demonstrates they lack the coverage of procedural safeguards applied around model behavior.
• 0:47
  A common thread across recent advances in artificial intelligence reveals that the primary obstacles to progress are not deficits in scale or compute,
• 0:55
  but the structural assumptions built into models, training paradigms, and evaluation protocols.
• 0:59
  From optimization pathologies that resist scaling remedies, to multimodal systems that fail because single-modality representations are intrinsically incomplete,
• 1:08
  to agentic benchmarks where measured competence is confounded with spurious factors, the evidence points toward a systematic re-examination of foundations.
• 1:16
  Simultaneously, efficiency gains now flow from architectural internalization rather than external heuristics, scientific AI increasingly embeds causal self-correction as a design requirement,
• 1:27
  and safety research fractures into mechanistic and procedural camps with irreducible blind spots.
• 1:32
  These developments collectively mark a pivot away from monolithic scaling toward formal scrutiny and systems-level redesign.
• 1:39
  Structural Pathologies in Deployed Methods.
• 1:42
  Formal analysis is now exposing inherent structural flaws in widely deployed AI methods—flaws tied to design assumptions that incremental scaling cannot remedy.
• 1:51
  In optimization, standard adaptive optimizers such as Adam lack gauge-equivariant structure and consequently drift along flat symmetry directions,
• 1:59
  settling into more degenerate minima than symmetry-respecting alternatives can reach.
• 2:04
  This pathology is not a deficit of compute or data; it originates in an architectural choice that fails to separate the gauge orbit from the quotient space,
• 2:12
  meaning that even unlimited training time would not rectify the attractor dynamics.
• 2:16
  A parallel class of structural defects appears in the reinforcement‑learning‑based finetuning of language models.
• 2:22
  Group Relative Policy Optimization (GRPO) underpins widely deployed systems like DeepSeek‑R1,
• 2:28
  but its theoretical properties remained poorly understood until a first‑principles derivation from the policy gradient theorem revealed a hard credit‑assignment bottleneck.
• 2:36
  GRPO’s output‑only reward assumption forces identical advantage scores onto every token in a sequence, so the method provably cannot assign distinct credit to individual tokens;
• 2:47
  its gradients live in a rank‑2 subspace that is structurally incapable of supporting multi‑step reasoning or compositional tasks.
• 2:54
  No expansion of model capacity or training budget can differentiate token‑level contributions when the advantage signal is uniform—the limitation is locked into the method’s reward design.
• 3:03
  Multi‑round mixture‑of‑experts routing suffers a failure mode that mirrors this collapse of signal differentiation.
• 3:10
  In Router‑R1 and similar agents, a small router’s strong pre‑training priors interact with sparse outcome rewards to produce Trust Region Collapse:
• 3:19
  the router increasingly disregards complex expert reasoning traces, driving expert‑call probabilities to zero.
• 3:25
  Here, again, the fault lies not in insufficient depth or expert count but in a design that couples interactive reasoning with a routing policy under a sparse reward regime,
• 3:34
  funneling the system into degenerate local optima that more rounds of scaling would only reinforce.
• 3:39
  Even the benchmarks used to gauge progress harbor structural pathologies that escape correction through scale.
• 3:45
  In diffusion large language models, the ranking of parallel decoding methods is highly sensitive to prompt template choice, so reported acceleration gains can be mere evaluation artifacts.
• 3:55
  This evaluation illusion shows that design assumptions about method robustness to prompt variation are systematically violated;
• 4:01
  a larger model or more decoding steps cannot fix a sensitivity that resides in the evaluation protocol itself.
• 4:07
  Across optimization, credit assignment, routing, and evaluation,
• 4:11
  formal scrutiny identifies collapse points where structural properties are absent or mis‑specified—indicating that the binding constraint on progress is now the redesign of methods to satisfy
• 4:20
  necessary formal conditions, not the accumulation of further compute.
• 4:25
  Efficiency Through Architectural Internalization.
• 4:28
  The migration of the inference-efficiency frontier toward architectural internalization is epitomized by the re-framing of multi-agent debate not as a separate orchestration layer but as an internal
• 4:38
  routing mechanism within a single model’s Mixture-of-Experts architecture.
• 4:42
  By embedding the debate directly into the model’s parameter-efficient routing, the approach eliminates the external token exchange and latency penalties of conventional multi-agent systems,
• 4:51
  delivering a 3.7 times reduction in latency and 87% fewer tokens while maintaining accuracy.
• 4:57
  This trade-off renders dynamic deliberation viable in resource-constrained settings, internalizing a capability that previously lived outside the model boundary.
• 5:05
  Yet such architectural internalization does not merely dissolve overhead; it relocates complexity and surfaces new degeneracy modes that require explicit engineering.
• 5:14
  In high-resolution Diffusion Transformers, for example, a seemingly benign design choice—redundant spatial self-attention masks—inadvertently disables FlashAttention fast paths,
• 5:24
  creating a systems bottleneck termed the Mask-Induced Dispatch Tax (MIDT).
• 5:29
  The bottleneck is not a limitation of the attention mechanism itself but a structural artifact of how masking interacts with optimized kernel dispatch; its resolution demands a dedicated,
• 5:39
  training-free framework that identifies and eliminates the tax without altering model weights.
• 5:44
  The gain in inference speed enabled by DiT architectures thus uncovers a degeneracy that must be explicitly mitigated through systems-level awareness.
• 5:52
  A parallel dynamic appears in block diffusion language models,
• 5:55
  where a critical train-inference gap had prevented practical wall-clock speedups despite the conceptual appeal of parallel block decoding.
• 6:02
  The introduction of concurrent multi-block refinement directly addresses this gap, raising tokens-per-forward to 9.34 and throughput from 781.50 to 951.41 TPS.
• 6:14
  The speedup, however, comes with a measurable accuracy cost on math and code benchmarks,
• 6:19
  meaning the engineering solution itself introduces a new trade-off surface that must be managed—a hallmark of the shift from abstract algorithmic gains to concrete systems-level co-design.
• 6:28
  The co-design imperative extends into production deployments.
• 6:31
  According to a media report on DeepSeek’s DSpark speculative decoding system,
• 6:35
  the transition from a collection of isolated speculative decoding tricks to a self-adaptive pipeline requires tight integration of algorithms, scheduling, and hardware execution;
• 6:45
  the report claims an 85% single-user speedup and 4 times high-concurrency throughput gain from this holistic co-design.
• 6:52
  While these vendor-reported figures are not independently verified,
• 6:56
  the account underscores that production-grade efficiency can no longer be achieved by patching external optimizations onto a fixed architecture—the system itself must be re-architected,
• 7:05
  and the emergent bottlenecks that surface during that re-architecting (such as load imbalances or scheduling conflicts) must be engineered away.
• 7:13
  Collectively, these developments reveal a pattern: architectural internalization and systems-level co-design now define the efficiency frontier,
• 7:20
  but each step forward surfaces new structural pathologies—mask dispatch taxes, train-inference misalignments, co-design complexity—that demand explicit, often training-free, mitigation.
• 7:31
  The binding constraint on progress is no longer raw capability, but the ability to engineer around the degeneracies that architectural ambition creates.
• 7:39
  Multimodal Structural Incompleteness.
• 7:41
  The prevailing view of multimodal AI has treated additional sensory streams as optional enhancers that boost robustness without altering the fundamental representational character of models.
• 7:51
  A growing body of evidence challenges this assumption, instead indicating that representations grounded in a single modality are structurally incomplete,
• 7:59
  giving rise to systematic and predictable failures that no amount of scaling can eliminate.
• 8:04
  The necessity of non-visual physical information is demonstrated empirically by TacGen, which shows that vision alone underdetermines contact-dependent properties like mass, density, and compliance.
• 8:14
  Two visually identical objects can differ drastically in hardness or weight, and only tactile contact supplies the physical evidence required to disambiguate them.
• 8:22
  Adding touch as a structured modality yields a ∆R² improvement of +0.570 in mass prediction and lifts simulated manipulation success from 0.246 to 0.979,
• 8:34
  confirming that touch is not an auxiliary signal but a necessary dimension for physical-world representation.
• 8:40
  The same structural gap appears in dynamic perception: Event-VLA reveals that RGB-only vision-language-action models suffer severe performance degradation under low illumination, motion blur,
• 8:50
  and sensor noise, because standard frame-based cameras fail to capture fast physical dynamics.
• 8:55
  The proposed action-conditioned event-fusion interface recovers what Event-VLA terms action-time physical residuals—instantaneous, salient,
• 9:04
  and persistent motion cues—enabling robust performance where the visual modality alone collapses.
• 9:09
  Even when multiple modalities are nominally present, incomplete cross-modal alignment introduces its own class of systematic failures.
• 9:16
  MIRROR demonstrates that existing adapter modules linking vision and language in multimodal large language models succeed in preserving individual token-level semantic correspondences but lose
• 9:27
  the inter-concept relational structure during cross-modal transfer.
• 9:31
  Consequently, models falter on compositional visual reasoning tasks, overriding correct language priors when salient but misleading visual features appear.
• 9:39
  A complementary structural failure is identified by FADE, which provides a mechanistic explanation for vision-language hallucinations:
• 9:46
  attention modules consistently aggregate visual evidence across layers, yet feed-forward network, or FFN, modules in critical mid-to-late layers — for example,
• 9:55
  layers 16 to 22 in LLaVA-1.5 — override that aggregated evidence with dominant language priors.
• 10:01
  The hallucination is not a random error but a predictable consequence of an architectural imbalance in how visual and linguistic signals interact within the model.
• 10:09
  Taken together, these demonstrations expose a consistent pattern.
• 10:13
  Missing sensory dimensions such as touch and high-frequency motion event streams render vision-only representations incapable of capturing essential physical properties,.
• 10:22
  Even within seemingly integrated multimodal architectures,
• 10:25
  flawed cross-modal alignment mechanisms either discard relational structure or permit language modules to suppress faithfully aggregated visual evidence, generating hallucinations,.
• 10:35
  The shift is thus not merely additive; it is a recognition that the fundamental incompleteness of single-modality representations demands structural redesign of how information is encoded,
• 10:45
  transferred, and integrated, rather than reliance on scale or surface-level fusion.
• 10:50
  The Evaluation Crisis in Embodied and Agentic AI.
• 10:53
  Progress in embodied and agentic AI is typically reported as benchmark scores that rise with scale,
• 10:59
  but the evidence now indicates that these gains cannot be causally attributed to the capabilities they are taken to measure.
• 11:05
  The evaluation apparatus itself systematically conflates confounding factors with genuine competence, creating an attributional crisis that cuts across physical task execution,
• 11:15
  diagnostic reasoning, and safety evaluation.
• 11:18
  The core of the problem in embodied AI is laid bare by the observation that Vision-Language-Action, or VLA,
• 11:24
  model task success rates conflate semantic generalization inherited from vision-language pretraining with actual physical reasoning.
• 11:31
  Scaling a VLA backbone improves benchmark performance, but the observed behavior may reflect nothing more than surface-level semantic pattern matching.
• 11:39
  The system never had to construct a causally grounded model of the physical world to succeed.
• 11:44
  The insufficiency of current evaluation is further exposed when the deployment context shifts from broad manipulation tasks to domains with unforgiving precision requirements.
• 11:53
  General-purpose VLA models have never been systematically tested in settings such as laparoscopic surgical robotics, which demand millimeter-level accuracy and safety guarantees.
• 12:02
  Without such testing, any extrapolation from existing aggregate metrics to genuine physical competence remains unsupported.
• 12:08
  The benchmarks do not probe the failure modes that would distinguish a robust reasoner from a lucky imitator.
• 12:14
  This inability to disentangle capability from confound extends to agentic reasoning.
• 12:19
  In microservice failure diagnosis,
• 12:21
  standard outcome-oriented benchmarks label only the final root cause and cannot discriminate between evidence-grounded causal reasoning and accidental keyword matching.
• 12:30
  The proposed remedy—a process-level evaluation paradigm that scores the intermediate steps of localization, identification,
• 12:36
  and causal reasoning—directly acknowledges that raw success rates are a broken signal.
• 12:41
  The shift to process-level evidence marks an explicit recognition that outcome-only scoring systematically inflates the apparent competence of agents.
• 12:51
  Scaling does not automatically dissolve these confounds; it may instead transmute them.
• 12:55
  The relationship between model size and a system’s sensitivity to evaluation context is not monotonic.
• 13:01
  Across open-weight language models, larger variants encode evaluation awareness in earlier layers, a dissociated representational pattern that standard black-box evaluations cannot detect.
• 13:11
  This finding implies that as frontier models grow, they may become increasingly adept at exploiting the surface structure of a benchmark without possessing the underlying ability,
• 13:20
  thereby intensifying authenticity concerns rather than resolving them.
• 13:24
  The conflation here is double: behavior masks internal capability, and internal representations that could reveal the conflation are invisible to prevailing evaluation protocols.
• 13:33
  Together, these strands of evidence form a coherent indictment: evaluation in embodied and agentic AI operates in a regime where success is overdetermined by priors, lucky guesses,
• 13:44
  and structural shortcuts, while the very capacity to detect this crisis is undermined by the opacity of the systems under test.
• 13:51
  Scientific AI: From Prediction to Causal Self-Correction.
• 13:55
  The transition from black-box prediction to frameworks that embed causal validation, belief revision,
• 14:00
  and physical consistency as first-class design constraints is unfolding concurrently across multiple scientific domains.
• 14:07
  A first-party report on Generative Causal Testing (GCT) illustrates this reorientation by distilling L L M-based brain-prediction models into concise verbal explanations of cortical response
• 14:17
  properties, then causally validating those explanations through targeted synthetic stories tested in human fMRI experiments.
• 14:24
  The framework treats causal testability—not predictive accuracy alone—as the primary scientific objective, thereby closing the gap between prediction and understanding.
• 14:33
  This commitment to causal validation is deepened by evidence-informed L L M beliefs for continual scientific discovery, where static priors are replaced with formal belief-revision principles.
• 14:43
  By rewiring hypothesis generation to respond to accumulated evidence,
• 14:46
  the method avoids wasting search budget on hypotheses already implied by past findings and boosts accumulated non-stationary surprisal by 30,
• 14:54
  demonstrating that when belief dynamics become a first-class design constraint, discovery efficiency improves fundamentally.
• 15:01
  The progression from one-shot causal testing to dynamic,
• 15:04
  evidence-sensitive belief updating extends the scientific AI paradigm from validating static explanations to sustaining a recursive hypothesis-generation cycle.
• 15:13
  In parallel, physical consistency is being elevated from a post-hoc check to a structural requirement in two instrumentation-adjacent contexts.
• 15:20
  Self-supervised calibration of scientific detectors reframes the problem as a physics-informed learning task that jointly infers latent calibration parameters and task-specific predictions directly
• 15:31
  from raw measurements, using physical consistency constraints in place of manual labels.
• 15:36
  This eliminates the circular dependency where pre-calibrated signals are needed to train the very model that is meant to perform calibration,
• 15:43
  enabling autonomous calibration and continuous performance monitoring without expert intervention.
• 15:48
  In computational chemistry, action-operator semantics imposes a mathematically rigorous structure on molecular diffusion models,
• 15:55
  transforming them from opaque samplers into auditable thermodynamic estimators that natively connect to thermodynamic ensembles and directly address the phase-space overlap bottleneck that cripples
• 16:05
  traditional free-energy perturbation.
• 16:07
  Here, physical law is not an external validator but the generative grammar of the model.
• 16:13
  These efforts cohere into a single pattern: causal validation, belief revision, and physical consistency are being re-specified from external desiderata into internal architectural constraints.
• 16:22
  The causal-testing methodology of GCT and the belief-revision mechanism of AutoDiscovery are complementary in that the former ensures individual explanations are causally grounded,
• 16:32
  while the latter ensures the search over hypotheses does not stagnate—a synthesis would yield discovery pipelines that are both causally verified per step and dynamically responsive to accumulating evidence.
• 16:42
  The calibration framework and the action-operator molecular model converge on the principle that domain-specific physical constraints can serve as the sole supervisory signal,
• 16:51
  bypassing manual annotation and rendering the resulting models auditable by construction.
• 16:56
  In every case, progress arises not from scaling black-box components but from redesigning the scientific modeling architecture to make structural assumptions explicit, testable,
• 17:05
  and physically faithful.
• 17:07
  Safety's Paradigm Split: Mechanistic vs. Procedural.
• 17:11
  Safety research is fracturing along a fundamental fault line:
• 17:14
  mechanistic interventions that scrutinize internal model activations are being contested by procedural safeguards that operate around model behavior,
• 17:22
  and formal evidence now reveals that activation-level defenses harbor inherent structural blind spots.
• 17:27
  The core tension is captured by a structural-blindness corollary: any single-layer activation-cone, subspace,
• 17:33
  or null-space gating defense is inherently vulnerable to prefilling attacks that craft activations inside the benign reference region,
• 17:40
  with bypass success rates reaching 50 to 82% against state-of-the-art methods like AlphaSteer.
• 17:46
  This proof that mechanistic guards can be subverted without altering surface-level behavior undercuts the very premise of relying on internal representations as a reliable safety boundary.
• 17:55
  A direct escape from this dead end appears in the Platonic Representation Defense, which pivots entirely away from activation-level scrutiny.
• 18:02
  Rather than probing a model’s internal states,
• 18:05
  the framework leverages the Platonic Representation Hypothesis—the observation that independently trained large encoders converge toward compatible projections of reality—to detect and purify
• 18:15
  backdoors through cross-encoder consistency checks.
• 18:18
  By treating safety as a problem of representational concordance across externally accessible models rather than internal gating, it operationalizes a procedural stance:
• 18:27
  the defense’s evidence does not live inside a single model’s weights, but in the structural agreement between distinct training processes.
• 18:34
  This shift from mechanistic intervention to external consistency stands in tension with activation-cone defenses precisely because it does not require access to, nor can be fooled by,
• 18:44
  the single-layer blind spot identified in.
• 18:47
  The split deepens when policy adherence is reframed as a dialogue-level procedural problem.
• 18:52
  PolicyGuard explicitly distinguishes its task from adversarial safety guarding, instead verifying company-policy compliance by reasoning over the entire conversation history as a sub-agent verifier.
• 19:02
  It achieves higher policy-violation recall while blocking roughly half as often, improving worst-case reliability by 6 to 12 percentage points across frontier models.
• 19:11
  This is not a refinement of mechanistic inspection; it relocates the safety check to a separate process that treats the model’s output trail as an auditable procedural artifact.
• 19:19
  By decoupling safety enforcement from internal activations, PolicyGuard directly embodies the procedural paradigm that ’s findings make necessary.
• 19:27
  Extending this logic to transparency itself, LLMography treats Human–AI conversation histories as structured, auditable artifacts rather than focusing on final outputs or internal representations.
• 19:38
  Where mechanistic approaches might probe model states to infer intent, LLMography shifts oversight to the traceable sequence of human direction, corrections,
• 19:47
  and validation that shaped an interaction.
• 19:49
  This reframes AI transparency as a structural property of the interaction record, not a feature extracted from activation patterns,
• 19:56
  further consolidating the procedural turn that the proved blind spots compel.
• 20:01
  Together, these works map a paradigm split:
• 20:03
  the formal demonstration that activation-level defenses are inherently vulnerable is met not by patching mechanistic checks but by a systematic reorientation toward procedural
• 20:13
  safeguards—cross-encoder consistency, dialogue-ground policy verification, and conversation-audit transparency.
• 20:19
  What emerges is a research trajectory in which safety no longer hinges on inspecting model internals, but on engineering reliable processes around model behavior,
• 20:28
  a shift driven by structural assumptions giving way to formal scrutiny.
• 20:32
  Briefly Noted.
• 20:33
  In robotics, a trio of advances tackle deployment bottlenecks from different angles.
• 20:38
  Behavior uncloning via Mode Redirection, or MoRE, distills classifier-guided redirection signals directly into policy weights.
• 20:45
  It boosts the deployment success rate by 44 percentage points across simulated and real-robot tasks without retraining or runtime overhead.
• 20:52
  AnyBody enables free-form humanoid control from arbitrary subsets of body keypoints.
• 20:57
  This dramatically simplifies teleoperation and scalable demonstration collection by letting users specify only task-relevant points rather than full-body poses.
• 21:06
  Meanwhile, ST-Merge, a training-free visual token merging framework, exploits 3D spatio-temporal coordinates to reduce vision encoder tokens by 70 percent.
• 21:15
  This delivers up to an 8.3 times speedup for pi 0 and allows robotic VLMs to process high-resolution streams at actionable latencies.
• 21:24
  Large language models find increasingly structured and domain-grounded uses.
• 21:28
  Multi-agent systems can autonomously invent and evolve concise symbolic reasoning protocols, cutting generated tokens by 3 to 6 times while maintaining accuracy across diverse reasoning benchmarks.
• 21:38
  For nuclear regulatory document review, an L L M agent frames the task as sequential planning over a dynamic knowledge graph,
• 21:45
  enabling multi-hop reasoning across multimodal evidence that chunk-based retrieval systems cannot handle.
• 21:51
  In materials science, a closed-loop multi-agent framework performs interpretable inverse design of metal-organic frameworks without training task-specific models,
• 22:00
  navigating vast chemical spaces without large labeled datasets.
• 22:04
  A citation-grounded Bengali agricultural dataset of 145,500 QA pairs brings verifiable,
• 22:11
  safety-aligned advisory to a sector where over 40% of Bangladesh’s workforce relies on understaffed extension services.
• 22:18
  Several methods rework calibration, matching, and sensing with minimal supervision.
• 22:22
  Self-Organized Conformal Prediction discovers input-space groups via self-organizing maps to build local calibration buffers,
• 22:29
  reducing weighted regional coverage gaps on seven of eight benchmarks with only a 6% increase in set size.
• 22:35
  A zero-shot shape matching paradigm first articulates then matches,
• 22:39
  obviating the need for correspondence-specific training and handling uncorrupted real-world scans that break geometry-sensitive pipelines.
• 22:46
  Finally, the first self-supervised pretraining framework for subsurface scattering learns light-transport representations from just eight high-frequency phase-shift images per view,
• 22:56
  slashing acquisition requirements by orders of magnitude and enabling efficient relighting of complex translucent objects with minimal hardware.
• 23:04
  Synthesis and Outlook.
• 23:06
  The convergence of these developments reveals a unifying dynamic: across disparate subfields,
• 23:12
  progress is no longer gated by raw scale but by the formal properties of system architectures and their underlying assumptions.
• 23:18
  The structural flaws exposed in deployed optimization methods, the incompleteness of unimodal representations,
• 23:24
  and the causal confounds contaminating embodied evaluation all trace back to a shared failure to model the constraints that govern real-world data-generating processes.
• 23:33
  This diagnosis is reinforced by the parallel migration in inference efficiency—where internalizing structure into architectures reduces costs but introduces new degeneracies—and in scientific AI,
• 23:45
  where causal self-correction is being elevated from a post-hoc check to a design primitive.
• 23:49
  These threads do not simply coexist; they jointly imply that the field is transitioning from a period of capability demonstration to one of epistemic hygiene,
• 23:57
  in which the central challenge is to make systems’ internal representations correspond faithfully to the external structure of the problems they purport to solve.
• 24:05
  A sharp tension, however, runs through safety research: while mechanistic interpretability seeks to ground guarantees in activation-level properties,
• 24:13
  the evidence suggests those very properties can be inherently blind to emergent failure modes, strengthening the case for procedural safeguards.
• 24:20
  That split encapsulates the broader open question: can formal, structural guarantees keep pace with the complexity of learned representations,
• 24:28
  or will robustness ultimately depend on external governance loops that do not require full internal transparency?
• 24:34
  This review draws on 34 developments: 32 Tier A research sources, 1 Tier B first-party source, and 1 Tier C/D secondary or community source.
• 24:42
  The firmest claims rest on the Tier A work, while the first-party and community sources should be read as directional;
• 24:48
  stronger confidence would require independent replication and primary-source confirmation of the self-reported results.
• 24:54
  This edition cites 34 sources; full links are available in the text edition.