Latest / AI Sentinel: Frontier Daily / Beyond Scaling: Agentic Systems Unlock Discovery and Expose Safety Gaps

Transcript

Beyond Scaling: Agentic Systems Unlock Discovery and Expose Safety Gaps · AI Sentinel: Frontier Daily

• 0:12
  This is AI Sentinel, your daily brief from the AI frontier.
• 0:16
  Beyond Scaling: Agentic Systems Unlock Discovery and Expose Safety Gaps.
• 0:21
  Highlights.
• 0:22
  Open-weight models such as Meta's Muse Spark Open and Nvidia's Nemotron are directly challenging proprietary dominance in agentic and multimodal tasks, shifting the industry toward decentralized,
• 0:33
  enterprise-deployable AI.
• 0:35
  As L L M agents tackle long-horizon tasks, the field is converging on structured memory, recursive self-improvement,
• 0:42
  and causal failure attribution as essential primitives that move beyond monolithic ReAct loops.
• 0:47
  The deployment of L L M agents in real-world workflows is exposing novel, systemic security vulnerabilities—including tool-result misbinding, skill composition attacks,
• 0:57
  and contextual integrity violations—that existing safety alignments do not address.
• 1:02
  AI is transitioning from a tool for scientific analysis to an autonomous agent of scientific discovery,
• 1:07
  a shift that demands new verification frameworks such as Lean-grounded proof checking and reproducible agentic benchmarks to maintain epistemic integrity.
• 1:16
  As agentic workloads scale, system-level bottlenecks are migrating from model training to inference serving, driving innovations in tokenization, KV-cache management, speculative decoding,
• 1:27
  and multi-model memory coordination.
• 1:29
  The contemporary AI frontier is defined less by raw model scaling than by the rise of agentic systems, open-weight infrastructure,
• 1:36
  and specialized scientific tools—capabilities that simultaneously unlock autonomous discovery and expose deep, systemic safety and reliability gaps.
• 1:45
  Open-weight releases from major industry players are challenging proprietary dominance in agentic and multimodal tasks, shifting the landscape toward decentralized, enterprise-deployable AI.
• 1:55
  As agents take on long-horizon objectives, the field is converging on structured memory, recursive self-improvement, and causal failure attribution as essential architectural primitives.
• 2:05
  Yet this autonomy introduces novel attack surfaces—tool-result misbinding, skill composition exploits, contextual integrity violations—that existing safety alignments do not address.
• 2:16
  Parallel advances in scientific automation, embodied robotics, and inference-serving infrastructure extend these dynamics further:
• 2:23
  AI is becoming an autonomous agent of discovery requiring new verification frameworks, while system bottlenecks migrate from training to agent serving.
• 2:31
  Together, these developments reveal a frontier where capability and vulnerability advance in lockstep.
• 2:37
  Open-Weight Frontier Models Reshape the Competitive Landscape.
• 2:41
  The open-weight frontier has shifted decisively from a research novelty into a direct competitive challenge to proprietary model dominance,
• 2:48
  driven by concurrent releases from Meta and Nvidia that target agentic and multimodal capabilities at frontier scale.
• 2:54
  Meta's Muse Spark Open, an 816-billion-total-parameter multimodal mixture-of-experts model with approximately 42 billion active parameters per token and a context window of up to 1 million tokens,
• 3:05
  was released as an open-weight model on Hugging Face under the Muse Model License.
• 3:10
  According to Meta's official announcements, the model processes text, images, audio, and video simultaneously within its 1 million-token context,
• 3:18
  positioning it as the most capable US open-weight model and specifically targeting coding and agentic tasks.
• 3:24
  The release includes additional safety post-training for standalone robustness and ships with Helion, a pie torch-native kernel DSL delivering 1.03 to 3.95 times kernel-level speedups,
• 3:35
  emphasizing infrastructure-level support for open-weight deployment.
• 3:38
  Media reporting by The Decoder frames Muse Spark Open as Meta's direct answer to leading Chinese open-weight competitors including Qwen, GLM, and Kimi,
• 3:47
  intensifying the global open-weight AI race by providing a US-based alternative.
• 3:52
  This framing positions the release within an explicitly competitive geopolitical context rather than merely a technical milestone.
• 3:59
  Nvidia's concurrent release of Nemotron 3.5 Super VL extends the open-weight challenge to proprietary dominance along a complementary enterprise axis.
• 4:08
  According to The Decoder, Nemotron is a 120-billion-parameter open multimodal mixture-of-experts model with 12 billion active parameters, designed for continuous enterprise agent tasks.
• 4:19
  Unlike Muse Spark Open's frontier-scale emphasis, Nemotron is accompanied by two compressed Bolt variants for local hardware—the 74 billion-A9 billion 24GB model and the upcoming E4 billion 4GB model—targeting self-hosted enterprise deployment.
• 4:34
  The Decoder reports that this combination of high throughput, native multimodality, and open datasets may shift enterprise adoption toward open-weight models optimized for specific hardware ecosystems.
• 4:45
  Taken together, these releases suggest a converging strategy: both Meta and Nvidia are building open-weight multimodal MoE architectures explicitly designed for agentic workloads,
• 4:54
  yet they address different deployment regimes.
• 4:56
  Meta's release targets frontier-scale capability and coding-optimized agentic performance, potentially democratizing advanced agentic AI development by lowering barriers for building such applications.
• 5:07
  Nvidia's release targets continuous enterprise agent operation with hardware-specific compression variants.
• 5:13
  Both sources characterize their respective releases as potentially lowering barriers to adoption—Meta for the open-source community building advanced agentic applications,
• 5:22
  Nvidia for enterprises needing customizable, self-hosted AI agents for continuous document and data processing —indicating that open-weight models are positioned to challenge proprietary alternatives across both research and enterprise
• 5:34
  deployment contexts simultaneously.
• 5:36
  Agentic Systems Demand New Paradigms for Memory, Self-Improvement, and Failure Diagnosis.
• 5:42
  The dominant ReAct-style loop—where an agent reasons, acts, and observes within a single, ephemeral context window—is proving structurally insufficient for long-horizon tasks.
• 5:52
  As agents move from isolated queries to sustained research workflows, the field is converging on three new architectural primitives: structured memory for persistence,
• 6:00
  recursive state inheritance for self-improvement, and causal attribution for failure diagnosis.
• 6:06
  For persistent memory, the core problem is that standard nearest-neighbor retrieval fails when context grows.
• 6:12
  MemCode addresses this by defining atomic memory units as self-contained, source-grounded facts and mapping them into a learnable codebook via isomorphic semantic quantization.
• 6:21
  This discrete semantic representation is designed to reduce context-dependent retrieval failures in long-horizon assistants,
• 6:28
  while source pointers make memory construction and auditing more practical.
• 6:31
  The Agentic Scientist pipeline extends this architectural shift by using cross-agent orchestration, rubric-based evaluation,
• 6:38
  and evidence-generation stages to support reliable research workflows rather than only hyperparameter optimization.
• 6:44
  Taken together, these approaches suggest that persistent, auditable memory is a prerequisite for grounding autonomous discovery—preventing the overfitting and ungrounded generation that arises when agents lack structured historical state.
• 6:57
  Beyond static persistence, agents must learn from experience across iterations.
• 7:01
  SwarmRSI formalizes recursive self-improvement for scientific agent swarms by defining evolving inheritable research states—hypotheses, programs, experimental evidence,
• 7:11
  and unresolved needs—stored in a shared graph.
• 7:14
  This mechanism addresses cases where promising hypotheses are abandoned after inadequate implementations and reuses retained hypotheses and complementary program components across agents.
• 7:24
  This represents a shift from monolithic reasoning loops toward architectures where intermediate research artifacts persist as inheritable structure across agent generations.
• 7:33
  However, as architectures grow more complex with shared memory and inherited states, diagnosing failures becomes harder.
• 7:39
  Current debugging practices that rely on judgment-based first-mistake labels are critically flawed.
• 7:44
  For 66% of such labels, the agent's success probability does not even drop at the labeled step.
• 7:50
  To address this, the Executable Prefix Values framework introduces V of H T—that's the probability that a stochastic agent succeeds from a given intermediate history—to distinguish the true cause of a failure
• 8:01
  from its visible symptoms.
• 8:03
  This directly challenges existing agent evaluation by exposing that human-identified failure points often target symptoms rather than causes.
• 8:10
  Collectively, these developments indicate that agentic reliability requires moving beyond single-pass reasoning.
• 8:16
  Structured memory enables persistence, shared graphs enable inheritance, and prefix-value analysis enables causal diagnosis —each addressing a distinct limitation of monolithic agent loops.
• 8:27
  The Hidden Attack Surface of Autonomous Agents.
• 8:29
  As L L M agents transition from isolated conversational interfaces into stateful, multi-step systems embedded in real-world workflows,
• 8:38
  the attack surface they present diverges sharply from the vulnerabilities addressed by conventional safety alignment.
• 8:43
  The emerging threat landscape is defined not by overt content manipulation but by structural exploitation of the agent's own execution semantics.
• 8:51
  Tool-result misbinding exemplifies this shift.
• 8:54
  TRAP identifies a failure mode in which a real, unmodified tool result is replayed into an incompatible execution context, causing an L L M agent to take unsafe actions.
• 9:04
  This attack surface is distinct from prompt injection or tool poisoning precisely because it requires no content modification,
• 9:11
  exposing a vulnerability inherent to stateful agent systems rather than to the content itself.
• 9:16
  The danger compounds when agents operate through composed skills rather than monolithic prompts.
• 9:21
  CoRedSkill demonstrates that individually audited skills appearing benign can compose into stealthy, unauthorized operations during execution.
• 9:28
  This suggests that the threat is not limited to replayed results but includes emergent behavior of skill compositions that preserve original task utility while executing malicious operations.
• 9:39
  Beyond direct manipulation of tools and skills, agents exhibit systemic failures in maintaining contextual integrity during executable tasks.
• 9:46
  AgentCIScope reveals contextual integrity violation rates ranging from 32.5% to 94.4% across leading L L M agents in real-world application environments,
• 9:56
  exposing a gap between task-completion capability and privacy preservation.
• 10:00
  This leakage occurs through the agent's normal operation rather than through adversarial intervention,
• 10:05
  suggesting that functional competence and privacy preservation can diverge dramatically within the same system.
• 10:11
  TRACER further clarifies why existing alignments fail to catch these vectors.
• 10:16
  It identifies that the same malicious payload has an 18% average attack success rate as a follow-up message but 53% when re-wrapped as retrieved context.
• 10:25
  This framing gap means that defenses trained on one deployment wrapper may fail when untrusted content arrives through RAG, tools, or other agents.
• 10:33
  Taken together, these findings suggest that the deployment of L L M agents introduces vulnerabilities—misbinding, compositional stealth, contextual leakage,
• 10:42
  and wrapper-dependent injection—that operate below the threshold of current content-level safety alignments and demand defenses attuned to agent execution architecture,,,.
• 10:51
  Scientific Discovery and Verification Increasingly Rely on AI-Mediated Automation.
• 10:56
  AI's role in scientific research is shifting from analytical support toward autonomous discovery, but the evidence indicates this transition is outpacing the verification frameworks needed to maintain epistemic integrity.
• 11:07
  Google DeepMind introduced Google Antigravity, an agentic science workbench that compresses fragmented biological database searches into single interactive sessions for antimicrobial discovery,
• 11:17
  according to an official company announcement.
• 11:20
  This positions AI not merely as an analytical instrument but as an active participant in the discovery workflow.
• 11:26
  Yet parallel developments reveal that the outputs of such agentic systems frequently fail to meet domain standards.
• 11:32
  The INTO THE UNKNOWN benchmark, introduced in an archive preprint, tasks AI agents with conducting open-ended scientific research by writing a follow-up paper to a provided study;
• 11:42
  it reveals that frontier agents can produce superficially plausible scientific follow-ups, but none of the 135 submissions met domain scientists' standards.
• 11:51
  The tension between Antigravity's promise of accelerated discovery and the benchmark's exposure of inadequate scientific output underscores that agentic capability and scientific validity
• 12:01
  are not converging at the same rate.
• 12:03
  The verification challenge intensifies as AI systems generate novel mathematical content at scale.
• 12:09
  A structured extraction presented in an archive preprint isolates new mathematical instruments—lemmas, constructions, inequalities,
• 12:16
  and proof mechanisms—from 38 of the 372 problem families in Open A I's October 2026 release of 722 AI-generated mathematical manuscripts,
• 12:26
  serving as a triage tool for the mathematical community confronting a massive, unrefereed corpus.
• 12:31
  Another archive preprint identifies a related bottleneck: as AI systems generate increasingly complex mathematical proofs, checking step-by-step validity becomes critical,
• 12:41
  and existing methods for obtaining step-level labels rely on costly human annotations or final-answer checks that do not apply to proofs.
• 12:48
  The LeanPRM framework responds to this bottleneck by using the Lean proof assistant to automatically generate step-level correctness labels for natural-language mathematical proofs without human
• 12:59
  annotation, producing a corpus of 10,634 labeled proofs with 152,005 step labels plus LeanPRM-Bench, a held-out benchmark of 200 proofs.
• 13:09
  Taken together, these sources suggest an emerging structural pattern:
• 13:13
  as agentic systems accelerate scientific production—whether in biological database triage or mathematical manuscript generation—the field is simultaneously constructing verification mechanisms
• 13:23
  that range from formal proof assistants to reproducible agentic benchmarks.
• 13:28
  The triage approach applied to Open A I's manuscripts and the Lean-grounded step-level verification of LeanPRM represent complementary responses to the same epistemic challenge
• 13:38
  that the INTO THE UNKNOWN benchmark makes visible: autonomous scientific output requires verification infrastructure commensurate with the scale of generation.
• 13:46
  Embodied AI Advances Through World Models, Tactile Integration, and Sim-to-Real Transfer.
• 13:51
  Embodied AI is advancing along three complementary fronts: force-aware manipulation, predictive world models, and scalable simulation pipelines.
• 13:59
  Each one addresses a different bottleneck in the path from perception to physical action.
• 14:04
  The TacClaw-Infinity dataset, described in a preprint, provides a large-scale structured vision-tactile-language-action corpus.
• 14:11
  It comprises over 10,000 hours, 1 million episodes, and 1 billion frames across more than 1,000 task types,
• 14:17
  with the aim of advancing contact-rich robotic manipulation through structured supervision.
• 14:22
  This emphasis on tactile integration reflects a recognition that vision and language alone are insufficient for tasks requiring precise physical contact.
• 14:30
  Complementing the data side, a preprint presents the first systematic, controlled empirical investigation of world modeling integration in Vision-Language-Action, or VLA, models.
• 14:40
  It disentangles confounding variables that previously obscured the true sources of empirical gains, offering actionable design principles for world-model-augmented VLAs.
• 14:50
  Taken together, these two lines of work suggest that progress in embodied manipulation depends jointly on richer multimodal data and principled architectural integration of predictive models.
• 15:01
  Scalable supervision and simulation pipelines further accelerate the field by reducing dependence on costly robot teleoperation.
• 15:07
  The HUMANVERSE dataset and the λ0 humanoid vision-language-action policy, introduced in a preprint,
• 15:14
  draw on 500 hours of open-world egocentric human data pairing video with synchronized body and hand motion;
• 15:20
  the reported alignment-error reductions and monotonic scaling suggest that robot-compatible human data may improve transfer to physical robots.
• 15:27
  This approach of converting everyday human whole-body demonstrations into training signals extends the data pipeline beyond controlled laboratory settings.
• 15:35
  A media report by The Decoder describes Google's Gemini 4 Argon as a single model designed to unify robot control, interactive 3D simulation generation, and training data evaluation,
• 15:46
  generating over 4,500 varied simulation training episodes from a single photo and boosting a real robot's success rate from 56 to 75 percent.
• 15:55
  While the HUMANVERSE and λ0 work leverages human experience as scalable supervision, the Gemini 4 Argon approach, as reported by media,
• 16:03
  replaces fragmented toolchains with a single model for simulation, control, and evaluation;
• 16:08
  both strategies address the sim-to-real transfer problem from different directions—one through human-derived data, the other through simulation generation from minimal inputs.
• 16:17
  Despite these advances, the evidence carries inherent limitations.
• 16:20
  The TacClaw-Infinity, world-model VLA, and HUMANVERSE sources are all preprints with unknown peer-review status,,,
• 16:27
  and the Gemini 4 Argon results are vendor-reported metrics conveyed through a media report rather than independently verified.
• 16:33
  These caveats underscore that, while the integration of force-aware manipulation, predictive world models, and scalable simulation pipelines is accelerating robotics research,
• 16:43
  the field has not yet established whether these approaches reliably achieve long-horizon physical reasoning or cross-embodiment generalization.
• 16:50
  Infrastructure Bottlenecks Shift from Model Training to Agent Serving and Inference Efficiency.
• 16:55
  As agentic workloads scale, the inference serving stack is experiencing a structural reallocation of bottlenecks.
• 17:01
  They are shifting from model training to the preprocessing, memory management, and distributed execution layers that sustain agent execution.
• 17:08
  Tokenization, historically a negligible preprocessing step, has become a meaningful constraint on time-to-first-token within LLM serving pipelines.
• 17:16
  A paper formalizing online tokenization introduces delay as a measure of how far an online tokenizer's output lags the maximum stable prefix.
• 17:25
  It proves that exact zero-delay online tokenization is NP-hard unless P equals NP, with no polynomial-time algorithm capable of beating O of M T delay.
• 17:35
  This theoretical result directly bounds the achievable responsiveness of streaming character-to-token conversion,
• 17:41
  a mechanism that could otherwise reduce time-to-first-token and memory pressure in serving systems.
• 17:46
  The tokenization bottleneck intensifies specifically under KV-cache reuse, a condition central to agent serving.
• 17:52
  Delta Tokenizer addresses this by introducing an exact incremental tokenization front end that treats each agent turn as a delta over the previous request rather than re-encoding the full history.
• 18:03
  This approach targets the growing CPU-side bottleneck where KV-cache reuse elevates tokenization's share of time-to-first-token,
• 18:10
  and it preserves exact reference token IDs to avoid the subtle cache misses or model-input mismatches that approximate tokenizers may introduce.
• 18:18
  Where the online tokenization paper establishes the theoretical hardness floor for streaming conversion,
• 18:23
  Delta Tokenizer extends the practical response into the agent-specific regime of incremental session deltas.
• 18:29
  Beyond preprocessing, multi-model agent architectures introduce a distinct memory management challenge.
• 18:35
  JointServe introduces a unified residency formulation treating model weight shards and KV cache blocks as competing objects within a single G P U memory budget.
• 18:43
  This formulation responds to the breakdown of the assumption that all models fit in G P U memory, a constraint that fails for private and edge deployments as agents evolve toward architectures where specialized capabilities reside in distinct weights.
• 18:56
  Distributed inference offers a complementary path to the memory-constrained approach.
• 19:00
  Apple reports that MLX v0.30.1 introduces the jaccl communication backend supporting distributed L L M inference across multiple Mac Studio computers using a low-latency RDMA driver over Thunderbolt
• 19:13
  5, introduced in macOS 26.2.
• 19:17
  According to Apple, this capability could democratize access to frontier-class LLMs by enabling near-trillion-parameter local inference in private environments without cloud infrastructure,
• 19:27
  with demonstrated energy efficiency of 200W-300W peak power and high decode speeds.
• 19:32
  Taken together, JointServe's unified residency formulation and MLX's distributed RDMA approach represent distinct strategies—memory-internal optimization versus hardware-level distribution—for resolving the serving
• 19:44
  constraints that agentic workloads impose.
• 19:47
  Briefly Noted.
• 19:48
  Microsoft's official company announcement introduces Magnetar, a zero-knowledge virtual machine that proves execution of general-purpose RISC-V programs written in Rust, Go,
• 19:58
  or C without manual circuit rewrites.
• 20:00
  It runs 10 to 39 times faster on average than the fastest comparable existing zkVMs across 15 benchmark programs.
• 20:08
  The announcement suggests this could make cryptographic proof generation a routine output of everyday software execution rather than a specialist activity.
• 20:15
  On the infrastructure front, Nvidia has open-sourced its Run:ai AI workload and G P U orchestration platform under the Apache 2.0 license.
• 20:23
  The release encompasses the full control plane, Kubernetes operator and controllers, C L I, and a local development environment, according to an official company announcement from Nvidia.
• 20:33
  A community post on Hacker News corroborates the release, noting it could lower the barrier for organizations to share and govern G P U resources across teams on Kubernetes.
• 20:42
  The Nvidia announcement further specifies four modular, independently adoptable components: KAI Scheduler, contributed to CNCF as a Sandbox project; G P U Fractioning Operator;
• 20:54
  KAI Resource Management; and Workload Map. These could potentially standardize how shared AI infrastructure is managed on Kubernetes.
• 21:01
  Several releases advanced open-weight and multimodal model development.
• 21:05
  An official company announcement from Hugging Face describes Apertus 1.5, which extends 8 billion and 70 billion models with native image and audio understanding via early fusion,
• 21:15
  a 262 thousand context window, an optional thinking mode, and improved tool use, releasing weights, data, intermediate checkpoints, and training recipes under Apache 2.0.
• 21:25
  The announcement notes that transparent Evaluation Cards mapped to EU AI Act governance requirements could set a new standard for reproducible AI reporting.
• 21:33
  Black Forest Labs released FLUX.
• 21:36
  3 Image Klein, an open image generation model integrated into the Hugging Face Diffusers library from day one that unifies text-to-image, multi-reference image-to-image,
• 21:46
  and bounding-box layout control within a single model, according to an official company announcement from Hugging Face. Together AI reports via a first-party vendor blog post that Meta's Muse Spark Open,
• 21:56
  an open-weight model for agentic coding with a 1 million-token context window designed for long-running agentic software engineering workflows, computer use, web browsing, and multimodal input,
• 22:07
  is now available on its platform as a Day 0 launch partner.
• 22:10
  IBM's official company announcement on Hugging Face introduces Granite Switch 4.1, in 8 billion and 30 billion sizes,
• 22:17
  a modular model system composing a base LLM with multiple Activated LoRAs and a co-located ASR encoder into a single endpoint, including a general-purpose Decision aLoRA,
• 22:27
  a specialized Guardian aLoRA for safety tasks, and Granite-Speech-5, which could reduce GPU and VRAM costs for enterprise multi-task pipelines.
• 22:36
  A community post on Hacker News describes a study systematically analyzing AI coding agent instruction files across 1,866 popular GitHub repositories.
• 22:45
  It finds that 34 percent of projects now ship such files, with AGENTS dot md emerging as the de facto standard at 78 percent adoption among projects with instruction files.
• 22:55
  The post reports that most checkable rules lack named enforcement mechanisms, suggesting a concrete improvement path of attaching linter or CI references directly to rules.
• 23:04
  An official company announcement from Apple introduces Selection-based Structured Reasoning, or SSR,
• 23:09
  a framework that reformulates LLM reasoning from open-ended generation to discrete selection over pre-specified, reusable natural-language candidates.
• 23:17
  This addresses inference cost and latency bottlenecks of small multimodal models that often generate lengthy but unhelpful reasoning traces.
• 23:25
  An official company announcement from PyTorch describes the ExecuTorch Cortex-M backend, which enables PyTorch models to run directly on Arm Cortex-M CPUs using CMSIS-NN acceleration.
• 23:37
  This eliminates the need to convert models to alternative frameworks and removes a major friction point for embedded ML deployment.
• 23:44
  Synthesis and Outlook.
• 23:46
  The current AI frontier is defined less by raw model scaling than by the emergence of agentic systems, open-weight infrastructure,
• 23:54
  and specialized scientific tools—capabilities that simultaneously unlock autonomous discovery and expose deep, systemic safety and reliability gaps.
• 24:02
  Open-weight releases from Meta and Nvidia are decentralizing capabilities that, as an editorial interpretation, directly intensify the security concerns surrounding deployed agents:
• 24:11
  the broader the distribution of agentic models, the wider the attack surface from tool-result misbinding and contextual integrity violations.
• 24:19
  Meanwhile, the architectural convergence toward structured memory and recursive self-improvement in agentic systems creates a tension with infrastructure demands, as persistent,
• 24:27
  self-auditing agents impose inference-serving pressures that training-centric paradigms never anticipated.
• 24:33
  Scientific and embodied AI applications represent the sharpest test of these intersecting pressures—autonomous discovery and physical manipulation require both the safety guarantees that current
• 24:43
  alignments lack and the long-horizon reasoning that today's serving stacks struggle to sustain efficiently.
• 24:48
  Whether verification frameworks such as Lean-grounded proof checking can scale fast enough to match the accelerating deployment of under-secured autonomous agents remains an open question.
• 24:58
  The evidence base warrants moderate confidence regarding architectural and infrastructure trends, drawn from a mix of first-party releases and peer-reviewed research,
• 25:06
  though it is thinnest where safety governance and market dynamics are concerned, relying more heavily on secondary and unclassified sources.
• 25:14
  This review draws on 34 developments: 17 Tier A research sources, 11 Tier B first-party sources, and 6 Tier C/D secondary or community sources.
• 25:23
  The firmest claims rest on the Tier A work, while the first-party and community sources should be read as directional;
• 25:29
  stronger confidence would require independent replication and primary-source confirmation of the self-reported results.
• 25:35
  This edition cites 34 sources; full links are available in the text edition.