Latest / AI Sentinel: Frontier Daily / The Rise of Agentic AI Makes Reliability Engineering as Critical as Model Scaling

Transcript

The Rise of Agentic AI Makes Reliability Engineering as Critical as Model Scaling · AI Sentinel: Frontier Daily

• 0:00
  The Rise of Agentic AI Makes Reliability Engineering as Critical as Model Scaling.
• 0:04
  Highlights.
• 0:06
  Open-weight models like K3 and Inkling match proprietary benchmarks but their uneven capability profiles, high compute demands,
• 0:13
  and reliance on community fine-tuning create new accessibility and safety challenges that licenses alone do not resolve.
• 0:20
  Modular agentic pipelines that decompose reasoning into composable subsystems introduce new failure modes and governance demands, including challenges in turn-level credit assignment.
• 0:29
  Embodied AI systems using world-action models accelerate policy learning but invert the biological grounding hierarchy by prioritizing language over sensorimotor experience,
• 0:39
  limiting true generalization.
• 0:41
  A convergence on verifiability mechanisms—canonical action fingerprints and formal verifier cascades—argues that trust in autonomous agents must be a built-in core property rather than an add-on.
• 0:51
  A growing shift from monolithic models toward composable, agentic, and embodied AI systems has unlocked new capabilities, yet this transition exposes fragility in grounding, trust,
• 1:01
  and governance that threatens large-scale deployment.
• 1:04
  As the review unfolds, these vulnerabilities are examined across multiple dimensions: open-weight models whose uneven capability profiles create accessibility and safety challenges;
• 1:14
  modular agentic architectures whose compositional flexibility breeds new failure modes; embodied systems that prioritize language over sensorimotor experience, undermining generalization;
• 1:24
  and the emerging verifiability stack that treats trust as a built-in property rather than an afterthought.
• 1:29
  Collectively, these sections argue that reliability engineering must now stand alongside model scaling as a central discipline.
• 1:37
  Open-Weight Models Challenge Proprietary Dominance at the Cost of Usability and Safety Oversight.
• 1:43
  The open-weight landscape is shifting.
• 1:45
  A July 2026 media report from The Decoder notes that Kimi's K3, a multimodal mixture-of-experts model with 896 experts and 2.8 trillion total parameters,
• 1:55
  scored 57 on the Artificial Analysis Intelligence Index, a result the report describes as on par with Opus 4.8 and GPT-5.5.
• 2:03
  Separately, a personal blog post by Sebastian Raschka details the release of Inkling, a 975 billion-parameter sparse MoE with 41 billion active parameters,
• 2:13
  which achieves strong scores on IFBench at 79.8 percent and SimpleQA at 43.9 percent.
• 2:20
  These preliminary data points suggest that open-weight models are now competitive with proprietary systems on selected benchmarks.
• 2:27
  However, the same blog post reveals an uneven capability profile.
• 2:31
  Inkling’s performance on reasoning and coding tasks is considerably weaker,
• 2:35
  with scores of 29.7% on HLE and 54.3% on SWE‑Bench Pro that the post frames as lower relative to its instruction-following strengths.
• 2:44
  This mixed profile, Raschka’s analysis implies, makes open models less reliable as direct substitutes without extensive downstream tuning.
• 2:51
  The release strategy compounds the challenge.
• 2:54
  A second personal blog post, this one by Simon Willison, reports that Inkling is distributed under an Apache 2.0 license,
• 3:02
  while Raschka’s post notes that the model is intended for further fine‑tuning and specialization through Thinking Machines Lab’s Tinker platform,.
• 3:09
  Willison’s account frames the license and tooling as positioning Inkling as a base for community customization,
• 3:15
  thereby shifting the burden of safety alignment and domain‑specific performance onto users rather than the developer.
• 3:21
  The scale of these models—Inkling at 975 billion parameters, K3 at 2.8 trillion—further signals that hosting and adapting them demands substantial resources,
• 3:31
  though neither source quantifies these costs.
• 3:34
  Post‑release reliability introduces another gap.
• 3:37
  A second media report from The Decoder documents a stealth update to Google’s Gemma 4 that fixed tool‑calling bugs, reduced truncated responses, and improved agentic reasoning,
• 3:46
  alongside efficiency gains of up to 31% lower time‑to‑first‑token on Nvidia Hopper GPUs.
• 3:53
  The report indicates that even a well‑supported open model shipped with production‑level shortcomings that required a silent patch to meet dependability expectations for agentic workflows.
• 4:02
  The need for such post‑hoc fixes, according to the report, underscores a usability burden that falls on the community adopting the model.
• 4:10
  Taken together, these reports tentatively suggest that while open-weight models like K3 and Inkling now challenge proprietary performance on certain metrics, the uneven capability profiles,
• 4:20
  the delegation of safety to downstream fine‑tuning, and the ongoing necessity of post‑release patching create accessibility and oversight barriers that an open‑source license alone cannot resolve.
• 4:30
  Modular Agentic Architectures Replace Monolithic Reasoning with Composable, Yet Brittle, Subsystems.
• 4:36
  A series of recent contributions reveals a structural turn away from monolithic language model reasoning toward composable agentic subsystems,
• 4:44
  each targeting a specific failure mode inherent in end-to-end pipelines.
• 4:48
  Long-horizon tool use is now treated as a credit assignment problem separable from the language model itself.
• 4:53
  TRACE provides critic-free, dense turn-level rewards at tool-call boundaries, enabling agents to distinguish useful exploration from redundant actions and addressing the sparse,
• 5:03
  high-variance outcome rewards that plague monolithic reinforcement learning setups.
• 5:07
  Error correction, in turn, is extracted from iterative prompt-based reflection loops into a deterministic offline graph matching procedure.
• 5:14
  EMG replaces test-time trial-and-error with a graph computation that reduces latency and API costs while improving robustness for smaller models that struggle with self-reflection.
• 5:24
  Memory access is reformulated as a reinforcement learning problem separate from the core agent.
• 5:28
  MEMCON learns an online contextual-bandit policy with UCB exploration to decide when to retrieve, inject distilled plans, re-retrieve, consolidate, or forget, without any additional LLM calls.
• 5:41
  This replaces static memory heuristics that proved miscalibrated across task regimes — retrieving too aggressively when memory is sparse, too conservatively when reusable plans exist,
• 5:51
  or failing to help stuck agents.
• 5:53
  Finally, SPyCE distills multimodal agent rollout trajectories into a continuously updating hierarchical skill library.
• 6:01
  The library comprises execution skills, which are condition–action–effect triples for local visual operations, and workflow skills, which are high-level tool-orchestration priors.
• 6:10
  It captures recurring patterns that are typically lost as scalar reward signals, thereby avoiding the need for agents to rediscover them on every new task.
• 6:18
  Each of these pieces moves a function that a monolithic L L M agent might subsume—credit estimation, error recovery, memory management, skill reuse—into a dedicated, externally managed component.
• 6:29
  Taken together, they suggest a decisive architectural shift toward modular agentic pipelines where reliability is disaggregated into separately engineered subsystems.
• 6:37
  Yet the very decomposition that yields these gains introduces new fragility.
• 6:41
  MEMCON models memory operations as a Markov Decision Process and learns an online policy that adaptively decides when, what, and how much to retrieve, when to inject a distilled plan,
• 6:51
  and when to consolidate or forget.
• 6:53
  TRACE highlights that without explicit turn-level assignment, training dynamics cannot differentiate useful exploration from redundant or harmful actions,
• 7:01
  leaving the agent vulnerable to high-variance feedback.
• 7:04
  EMG’s design choice arises from the observation that iterative self-reflection is brittle for smaller models, so its offline correction is a targeted remedy for a specific model weakness.
• 7:13
  SPyCE’s closed-loop co-evolution of skills underscores that without a structured library, multimodal agents operate without reusable tool-use foundations, forcing reliance on brittle,
• 7:24
  per-task rediscovery.
• 7:26
  The interplay of these subsystems—reward signals shaping memory retrieval, error graphs referencing skill libraries,
• 7:31
  memory policies affecting credit assignment trajectories—creates a governance surface where each module’s failure mode can cascade,
• 7:38
  demanding explicit management of their co-evolution and calibration.
• 7:42
  Currently, all four contributions are archive preprints whose peer-review status is unknown; their findings are thus provisional,
• 7:49
  but collectively they illustrate that modular agentic architectures are both a response to the limits of monolithic L L M reasoning and a generator of new coordination challenges that will require
• 7:59
  systematic reliability engineering.
• 8:02
  Embodied AI's Push Toward Unified Frameworks Collides with Fundamental Deficiencies in Grounding.
• 8:07
  Recent advances in world-action models and learned simulation loops are pushing embodied policy learning toward practical deployment.
• 8:14
  A preprint on GigaWorld-Policy-0.5 reports that a mixture-of-transformers architecture reduces inference latency to 85 ms on a consumer RTX 4090,
• 8:26
  closing the gap between world-action models and lightweight VLA policies for real-time closed-loop control.
• 8:32
  Separately, QbitAI reports that Dexmal’s DW0.5 learned world model functions as a virtual coach for VLA post-training within the DFOL2.0 framework,
• 8:42
  offering cheap closed-loop feedback that lessens the burden of costly real-world teleoperation.
• 8:47
  These engineering improvements mark tangible momentum toward unified frameworks that integrate dynamics prediction and action generation on a single efficient substrate.
• 8:56
  Yet this push collides with two deeper findings.
• 8:59
  They call the semantic and architectural foundations of current systems into question.
• 9:03
  A preprint on semantic anchoring for robotic action representations diagnoses a systematic degradation.
• 9:09
  Action-only fine-tuning of VLA models progressively erodes the intention-level semantic structure inherited from pretrained vision-language models.
• 9:17
  This erosion correlates strongly with out-of-distribution generalization failure.
• 9:21
  Spearman rho equals 0.964.
• 9:24
  The paper provides both a diagnostic tool and a remedy that mitigates the erosion and yields performance improvements.
• 9:31
  But its central finding exposes a fragility built into standard fine-tuning recipes.
• 9:36
  A separate preprint on grounded world models in biological organisms makes a broader architectural argument.
• 9:42
  It contends that biological intelligence acquires grounded sensorimotor world models through active interaction first, with language attached only later.
• 9:50
  Current embodied AI systems, by contrast, are constructed on a language-first scaffold.
• 9:55
  According to this preprint, passive, language-scaffolded training regimes fail to capture the action-oriented understanding that underlies biological intelligence.
• 10:03
  Taken together, these findings suggest that the accelerating race toward unified embodied frameworks is building on an inverted foundation.
• 10:10
  DW0.5’s virtual coaching accelerates VLA post-training,
• 10:15
  but the semantic anchoring preprint demonstrates that action-oriented fine-tuning of such models systematically degrades the very semantic structure that supports generalization.
• 10:24
  Meanwhile, the pretrained vision-language backbones that commonly underpin VLA policies embody precisely the language-first scaffolding that the biological grounded world models preprint identifies
• 10:36
  as a fundamental architectural limitation.
• 10:38
  GigaWorld-Policy-0.5’s latency breakthrough makes world-model-driven control feasible on edge hardware,
• 10:44
  but it remains unclear whether the underlying representations retain the grounded sensorimotor semantics that robust physical generalization demands.
• 10:52
  The evidence therefore indicates that true generalization may remain elusive as long as embodied AI systems are built on a language-first paradigm that reverses the sensorimotor grounding observed
• 11:02
  in biological organisms, even as new training loops make that paradigm more computationally efficient.
• 11:08
  Enterprise AI Infrastructure Boom Outpaces Cost Visibility and Context Governance, Threatening Returns.
• 11:14
  A VentureBeat survey of 107 enterprises uncovers a stark compute gap in AI infrastructure: 83 percent of respondents operate GPUs at 50 percent utilization or less.
• 11:24
  Yet fewer than half, 44 percent, rigorously track compute costs.
• 11:28
  And 45 percent plan to evaluate specialized AI clouds they barely use today.
• 11:33
  This tentative picture of spending detached from economic discipline suggests that provisioning is racing ahead of the financial instrumentation required to prevent waste.
• 11:41
  The same preliminary pattern appears on the deployment side.
• 11:44
  A separate VentureBeat media report identifies an AI context gap,
• 11:49
  with 57 percent of surveyed organizations reporting that their agents have produced confident but wrong answers because of missing or inconsistent business context.
• 11:57
  The finding positions context governance, not merely retrieval quality, as a central source of authoritative-sounding hallucinations.
• 12:04
  Taken together, these media-reported surveys tentatively signal an emerging double deficit:
• 12:09
  enterprises are rapidly provisioning GPUs and deploying agents while lacking both the cost-visibility tools needed to avoid underutilization and the governed knowledge layers needed to suppress hallucinations.
• 12:19
  This combination threatens to erode the return on infrastructure spend.
• 12:23
  Vendor responses underscore the breadth of the gap.
• 12:26
  An official AWS announcement for Amazon Bedrock Managed Knowledge Base describes a fully managed agentic retrieval solution that handles data ingestion, vector storage, scaling, retrieval,
• 12:36
  and document-level access control without requiring users to provision infrastructure or select models.
• 12:42
  The service’s availability, as described by AWS, indicates that many enterprises lack in-house capability to assemble such governed retrieval pipelines,
• 12:51
  but it remains uncertain whether infrastructure-level retrieval alone can close the context gap that the survey attributes to missing or inconsistent business context.
• 12:59
  Meanwhile, a first-party engineering blog post by Together AI maps inference reliability tiers to failure domains and reports production metrics.
• 13:07
  This adds a further layer of fragility: even provisioned capacity, already reportedly underutilized, may suffer reliability blind spots when inference becomes mission-critical,
• 13:17
  an increasingly common scenario the blog identifies for products like Cursor and Decagon.
• 13:22
  The reliability framework thus extends the compute gap from simple underutilization to a more nuanced concern about whether enterprises understand the failure domains of the infrastructure they are buying.
• 13:31
  These preliminary signals, drawn from vendor disclosures and survey journalism,
• 13:35
  paint a cautionary picture in which the boom in enterprise AI infrastructure is proceeding ahead of the governance and cost-accounting discipline needed to make it reliably productive.
• 13:44
  Legal and Privacy Incidents Are Redefining AI’s Operational Boundaries Across Competition, Content, and IP.
• 13:51
  A wave of reported legal and privacy incidents—spanning antitrust enforcement, liability lawsuits, trade secret disputes,
• 13:57
  and data breaches—is tentatively carving new operational constraints that force AI companies to internalize security, transparency, and fair access as core engineering requirements,
• 14:07
  rather than treating them as afterthoughts.
• 14:09
  According to a media report by The Verge, the European Union has ordered Google to open Android and Search to rival AI assistants under digital antitrust rules,
• 14:18
  mandating search data sharing by January 2027 and Android changes by July 2027.
• 14:24
  The Verge reports that the two decisions could weaken Google's control over Android and Search and open up new opportunities for rivals to gain ground.
• 14:31
  For AI system developers, this order implies that interoperability and data portability must be engineered directly into product architectures, not bolted on as compliance patches.
• 14:41
  In a separate report, The Verge notes that xAI has sued an individual for allegedly using its Grok model to generate CSAM deepfakes.
• 14:49
  The lawsuit tentatively tests the legal boundaries of platform liability for AI-generated user content;
• 14:56
  such litigation could set a precedent that makes content‑safety moderation a fundamental engineering discipline within inference pipelines, rather than a purely legal or policy function.
• 15:05
  Meanwhile, QbitAI reports that Apple has filed a lawsuit against Open A I and two former Apple employees, alleging systematic theft of Apple hardware trade secrets,
• 15:14
  as Open A I pushes into consumer hardware—reportedly a Jony Ive‑designed AI device targeting a 2027 launch.
• 15:21
  The report suggests that AI hardware and model intellectual property are now high‑stakes battlegrounds,
• 15:27
  and the suit underscores that trade‑secret protection is becoming an engineering challenge requiring embedded access controls, model‑weight isolation,
• 15:34
  and secure development practices throughout the AI lifecycle.
• 15:38
  A security incident described by The Decoder exposes how coding‑agent design can turn into an insider threat.
• 15:44
  Users discovered that xAI’s terminal‑based AI coding agent “Grok Build” was silently uploading all files in a directory—including SSH keys, password databases,
• 15:53
  and personal documents—to xAI’s Google Cloud servers.
• 15:57
  The Decoder reports that in response, xAI open‑sourced the tool’s full Rust codebase under the Apache 2.0 license, a move that sets a tentative precedent for transparency‑driven trust recovery.
• 16:08
  The incident highlights that AI agents operating with filesystem access must be engineered with zero‑trust principles and data‑handling guarantees from the outset,
• 16:16
  or they risk becoming powerful vectors for privacy violations.
• 16:20
  Taken together, these reported developments suggest that the legal and privacy environment is beginning to impose constraints that collectively demand a shift in engineering priorities.
• 16:30
  The EU’s antitrust order mandates fair‑access interfaces; the Grok deepfake lawsuit presses for content‑filtering built into generation flows;
• 16:38
  the Apple trade‑secret suit raises the stakes for IP‑protective tooling; and the Grok Build breach illustrates that agent security cannot be an afterthought.
• 16:46
  While none of these incidents alone dictates a binding standard, their convergence hints that operational reliability in AI is increasingly inseparable from legal defensibility.
• 16:56
  Trustworthy Autonomy Requires Built-In Verifiability, from Action Attestation to Hallucination-Resistant Permissions.
• 17:02
  As AI agents transition from conversational interfaces to actuators in real-world workflows,
• 17:07
  the mechanisms that underpin their trustworthiness must evolve from post-hoc content moderation to built-in verifiability across the action lifecycle.
• 17:15
  The works from the review window collectively address a stack that spans canonical action attestation, enforceable permissions, capability verification,
• 17:22
  and rigorous assessment of the verification logic itself.
• 17:25
  At the foundation of this stack lies the problem of making agent actions auditable and non-repudiable.
• 17:31
  CAVA introduces a runtime-semantics layer that converts disparate agent actions—whether shell commands, SDK calls, or browser automation—into deterministic, hashable fingerprints,
• 17:41
  producing reproducible receipts so that what a deployer approves is what actually executes.
• 17:46
  This attestation substrate directly addresses the fragility that arises when governance relies on raw text or first-token matching, which wrappers can easily bypass.
• 17:55
  Without such canonical action traces, any higher-level permission or audit mechanism lacks a reliable enforcement point.
• 18:02
  Building upward, the state of user-level permissions remains uneven, as a survey of 21 proposals from 2024 to 2026 and five commercial agents reveals.
• 18:11
  This survey finds that authorization granularity is inconsistent across systems.
• 18:16
  Taken together with the action-attestation layer, the survey’s findings imply that even where permission interfaces exist, they are often decoupled from a verifiable execution substrate,
• 18:25
  creating a gap between what a user consents to and what the system can be proven to have done.
• 18:30
  Compounding this gap is a failure mode that challenges conventional safety measures.
• 18:34
  Protective Capacity Hallucination, or PCH, is documented in a preprint.
• 18:39
  It describes LLMs cast in protective roles that falsely claim they have performed life-critical actions—calling 911 or dispatching paramedics—when no such capability exists.
• 18:49
  The paper points toward explicit capability verification as a requirement to address this phenomenon.
• 18:54
  In a deployed agent, such verification would need to be grounded not in the model’s self-report but in an attested action log, of the kind that separate research proposes.
• 19:03
  This underscores how hallucination audits and action fingerprints are complementary constraints.
• 19:08
  Yet even explicit verification cascades carry hidden risk.
• 19:11
  A minimal theory of partially correlated verifier cascades shows that assuming conditional independence among verification gates can lead to catastrophic underestimation of failure rates:
• 19:21
  in realistic regimes, failure rates are underestimated by 20 times when five gates are cascaded, and by roughly 3000 times with ten gates.
• 19:29
  A practitioner budgeting for “five nines” of reliability on the basis of naive independence may in fact have only 98%.
• 19:35
  This finding does not undermine the value of verifier chains; it argues that the verifiability stack itself must be engineered with empirically grounded correlation models,
• 19:44
  not optimistic statistical assumptions.
• 19:47
  Collectively, these works trace the contours of trustworthy autonomy not as a single fuse but as a layered system property.
• 19:54
  An action attested by a canonical fingerprint provides the trace for a permission check.
• 19:58
  A hallucination audit constrains what the agent may claim about its own actions.
• 20:02
  And a realistic verifier cascade bounds the confidence one can place in the entire harness.
• 20:07
  The convergence, as documented across these preprints — all from the archive, with peer-review status unknown — suggests that verifiability must be designed into agentic systems from the ground up.
• 20:17
  Briefly Noted.
• 20:19
  A preprint demonstrates that APT-RL, a unified reinforcement learning framework,
• 20:24
  enables a quadrupedal robot to achieve instantaneous peak speeds of 4.25 meters per second over high steps and 6 meters per second during drop-down maneuvers on unstructured terrain.
• 20:34
  Another preprint presents REAL,
• 20:36
  an agentic framework for open-world mobile manipulation that eliminates oracle perceptual APIs and integrates a simulated user for natural‑language instruction resolution.
• 20:46
  According to a media report by QbitAI, RLinf v0.3 upgrades the open‑source embodied‑AI reinforcement learning infrastructure with five major capability leaps covering model ecosystem, algorithms,
• 20:58
  real-robot support, simulation, and system infrastructure, and adds six new embodied models.
• 21:03
  QbitAI also reports that Galbot released WAM‑TTT, a test‑time training framework for embodied AI that uses human video in place of teleoperation data,
• 21:13
  with resistance to catastrophic forgetting of 75.6.
• 21:17
  Elsewhere, a preprint introduces Nexus, a fully diffusion‑based framework for triangle mesh generation that eliminates autoregressive token prediction,
• 21:25
  reducing inference times and cascading errors for film and gaming pipelines.
• 21:29
  Another preprint reframes majority‑vote consensus as dense token‑level supervision for label‑free L L M training,
• 21:36
  achieving up to 12‑point pass@1 gains with roughly one‑seventh the compute of prior reinforcement‑learning approaches.
• 21:42
  OvisOCR2, a compact 0.8 billion‑parameter document parsing model described in a preprint,
• 21:48
  converts page images directly to Markdown with state‑of‑the‑art scores on OmniDocBench v1.6 (96.58 overall) and PureDocBench (75.06 ).
• 21:57
  A preprint introduces generative compilation, a method that provides compiler‑style feedback on partial programs during L L M code generation,
• 22:06
  catching borrow‑check and type errors early in languages like Rust.
• 22:09
  In vector search, a preprint proposes CwA, a clustering approach that decouples database partitioning from query probing,
• 22:17
  achieving up to 4.7 times throughput over state‑of‑the‑art in out‑of‑distribution settings.
• 22:22
  Finally, an official Apple announcement describes interactive proof systems that can verify general distribution properties, extending prior work on label‑invariant properties.
• 22:32
  Synthesis and Outlook.
• 22:34
  The shift toward composable, agentic, and embodied systems documented in this review weaves progress with fragility:
• 22:39
  open-weight models (claim 1) and modular agentic architectures (claim 2) both dismantle monolithic reasoning,
• 22:46
  yet their uneven capability profiles and brittle subsystem interactions directly feed the enterprise context gaps and hallucination risks highlighted in claim 4.
• 22:55
  Embodied AI’s grounding deficiencies (claim 3) compound this brittleness by inverting the sensorimotor foundation required for reliable real-world action,
• 23:03
  even as world-action models accelerate policy learning.
• 23:06
  This review suggests that legal and privacy incidents (claim 5) may externalize the costs of these fragilities, potentially forcing a need for internalization of security, transparency,
• 23:17
  and fair access—challenges that the verifiability stack outlined in claim 6 aims to address.
• 23:22
  Editorially, one might view these dynamics as mutually reinforcing:
• 23:26
  the push for agentic deployment by enterprises (claim 4) without governed knowledge layers could amplify the safety oversights of open-weight models (claim 1) and the failure modes of modular agents
• 23:36
  (claim 2); meanwhile,
• 23:38
  the legal redefinition of operational boundaries (claim 5) may demand the built-in verifiability mechanisms (claim 6) that remain largely absent from current embodied and agentic pipelines.
• 23:49
  A conflict surfaces between the openness that accelerates capability (claim 1) and the governance and trust requirements emerging across claims 2 to 6,
• 23:57
  because community fine-tuning and composable subsystems lack enforceable verifiability or cost discipline.
• 24:02
  The joint implication is that the field is heading toward an inflection where reliability engineering must absorb legal, economic,
• 24:09
  and sensorimotor constraints into core design—transforming trust from an add-on into a system property.
• 24:15
  An open question remains: can verifiability and grounding be layered onto composable agentic stacks without sacrificing the modularity and community-driven innovation that animate them?
• 24:25
  This review draws on 34 developments: 18 Tier A research sources, 2 Tier B first-party sources, and 14 Tier C/D secondary or community sources.
• 24:35
  The firmest claims rest on the Tier A work, while the first-party and community sources should be read as directional;
• 24:40
  stronger confidence would require independent replication and primary-source confirmation of the self-reported results.
• 24:46
  This edition cites 34 sources; full links are available in the text edition.