Latest / AI Sentinel: Frontier Daily / From Capability to Deployment: AI Progress Meets Structural Safety and Capital Limits

Transcript

From Capability to Deployment: AI Progress Meets Structural Safety and Capital Limits · AI Sentinel: Frontier Daily

• 0:00
  From Capability to Deployment: AI Progress Meets Structural Safety and Capital Limits.
• 0:05
  Highlights.
• 0:06
  Embodied robotics research is shifting from isolated skill demonstrations to systems-level deployment challenges such as body awareness, sim-to-real transfer, and retail-environment robustness.
• 0:17
  L L M security vulnerabilities extend beyond adversarial prompts to structural weaknesses in KV-cache reuse architectures and unlearning mechanisms,
• 0:25
  requiring a move from empirical red-teaming toward formal safety guarantees.
• 0:30
  Physics-informed machine learning is transitioning from heuristic loss penalties to mathematically enforced conservation laws, entropy stability, and distribution-free uncertainty bounds.
• 0:39
  Production AI agent reliability is increasingly constrained by silent failures—behaviors that complete successfully from the system's perspective yet produce incorrect outcomes—now being addressed
• 0:49
  by both platform tooling and academic benchmarks.
• 0:52
  AI infrastructure investment is entering a contested equilibrium between scaling ambition and capital discipline, marked by trillion-parameter models on domestic Chinese hardware,
• 1:02
  open-sourced alternatives to cooda, and the first negative cash-flow quarter at a hyperscaler.
• 1:07
  The day's evidence traces a field advancing simultaneously on three fronts—deployment-ready embodied intelligence, mathematically grounded physics-informed learning,
• 1:15
  and production agent reliability—while exposing structural safety vulnerabilities and financial sustainability pressures that complicate straightforward optimism about scaling.
• 1:25
  The sections that follow build this argument by showing embodied robotics pivoting from capability demonstrations toward systems-level deployment friction,
• 1:32
  L L M security research moving beneath the prompt layer to architectural and formal-provability concerns,
• 1:38
  and physics-informed ML replacing heuristic penalties with enforced conservation laws and distribution-free bounds.
• 1:44
  Production agent reliability is shown converging on silent-failure detection as the binding constraint,
• 1:49
  and infrastructure investment is framed as entering a contested equilibrium between scaling ambition and capital discipline.
• 1:55
  A brief-noted section rounds out the picture with contributions across evaluation, multimodal generation, medical AI, and community safety discourse.
• 2:04
  Taken together—by editorial interpretation—these developments suggest that progress and structural risk are now compounding in the same systems rather than advancing on separate tracks.
• 2:14
  Embodied AI's Frontier Shifts from Capability Demos to Deployment-Readiness Gaps.
• 2:19
  The day's embodied-AI evidence coalesces around a single structural insight: the field's central challenge has migrated from acquiring skills to deploying them.
• 2:27
  This shift is visible across multiple research efforts that, taken together, expose the systems-level friction separating capable models from operational robots.
• 2:36
  At the capability frontier, the HOST framework – that's Human-to-robot One-Shot Skill Acquisition – lets a robot learn new manipulation skills at inference time from a single human video,
• 2:46
  without updating any parameters, while keeping skills it already knows.
• 2:50
  By cutting out the teleoperation-and-fine-tuning loop for each new skill, HOST takes deployment-friction reduction to the extreme for skill acquisition.
• 2:58
  Yet this one-shot capability benchmark also sharpens a question its own preprint scope can't answer:
• 3:03
  whether these methods transfer to unstructured commercial environments where systems-level constraints dominate.
• 3:09
  That gap is quantified directly by DEED, a systems-level framework for deploying VLA-based humanoid robots in real retail settings,
• 3:16
  evaluated on a Unitree G1-Edu performing supermarket chip-restocking with the GR00 trillion N1.6 foundation model.
• 3:24
  The results are stark: naive fine-tuning of this frontier VLA model yields 0% success on the retail task, while the full DEED pipeline reaches 32%,
• 3:33
  and a single RECAP iteration raises performance to 42%—all on a single G P U.
• 3:39
  DEED's inclusion of a latent-space in/out-of-distribution analysis tool for diagnosing policy drift signals that the deployment problem is as much about monitoring and diagnosing failure modes
• 3:48
  as about raw capability.
• 3:50
  The contrast with HOST is instructive: where one-shot methods eliminate fine-tuning entirely, DEED demonstrates that in unstructured retail settings,
• 3:58
  even sophisticated post-training pipelines recover only partial success from a zero-percent baseline,
• 4:03
  directly quantifying a lab-to-deployment gap that one-shot methods do not yet address in commercial contexts,.
• 4:10
  The architectural diagnosis of this gap comes from a different direction.
• 4:13
  According to a media report by QbitAi, iFlytek spin-off Yaofang Intelligence released a technical report on iFLYTEK-Embodied-Omni, a unified multimodal embodied foundation model integrating VLM,
• 4:25
  VGM, and AGM via a shared multimodal representation, which makes body-awareness a training objective rather than a passive input.
• 4:33
  This framing positions the deployment gap as an architectural deficiency—the absence of body-awareness as an explicit optimization target—rather than a data-scaling problem,
• 4:41
  addressing what the report describes as an "executability gap" where visually plausible predictions fail physically.
• 4:47
  This argument extends toward a generalist foundation-model thesis for deployment, echoed by Tastone's AWE 3.5 announcement at WAIC, which, according to another QbitAi media report,
• 4:58
  positions an "embodied native brain" as the scaling-law analogue for robotics.
• 5:02
  Taken together, these sources trace a coherent arc: HOST defines the capability ceiling for frictionless skill acquisition, DEED quantifies the floor of real-world deployment difficulty,
• 5:13
  and the iFLYTEK-Embodied-Omni and AWE 3.5 efforts propose that closing the distance between them requires architectural integration of body-awareness rather than data scaling alone,.
• 5:23
  The convergence suggests that embodied AI's frontier is no longer defined by what robots can learn, but by what they can reliably execute.
• 5:31
  L L M Security Research Is Uncovering Structural Vulnerabilities Beneath the Prompt Layer.
• 5:36
  The dominant framing of L L M security has long centered on adversarial inputs—prompt injections, jailbreaks, and content-level manipulation.
• 5:45
  The day's evidence suggests this framing is insufficient.
• 5:48
  A paper accepted by USENIX Security 2026 formalizes a previously unknown attack surface in position-independent KV cache reuse systems deployed in production serving frameworks such as vLLM and LMCache.
• 5:59
  The vulnerability, termed HijackKV, is structurally distinct from prompt injection: no adversarial tokens need appear in the victim's input.
• 6:07
  Instead, the efficiency-driven optimization of KV cache reuse itself creates the exploitable channel,
• 6:12
  revealing that architectural choices made for throughput can introduce security liabilities invisible to input-level defenses.
• 6:20
  This attack on the serving layer finds a constructive counterpart in work on formal safety certification.
• 6:25
  A preprint introduces a novel framework for formal safety certification of autoregressive language models,
• 6:31
  computing mathematically rigorous lower bounds on the probability that a given L L M generates harmful output for a fixed prompt.
• 6:38
  The framework explicitly positions itself as addressing a gap left by empirical red-teaming and alignment methods, which cannot guarantee the absence of harmful behavior.
• 6:46
  Taken together, HijackKV's discovery of an architectural attack surface and the sound-bounds framework's insistence on mathematical guarantees suggest a converging recognition:
• 6:56
  empirical red-teaming, however extensive, cannot certify structural safety properties that depend on implementation-level and architectural-level guarantees rather than input-output behavior alone.
• 7:06
  The insufficiency of localized defensive interventions reinforces this point.
• 7:10
  A post on LessWrong demonstrates two novel attack routes to recover knowledge "forgotten" by LUNAR, a state-of-the-art unlearning method that edits a single MLP down-projection matrix.
• 7:21
  The work shows that localized unlearning edits in distributed circuits remain vulnerable to both optimization-based and representation-level recovery attacks,
• 7:29
  even without access to original weights.
• 7:31
  This finding extends the structural-vulnerability theme inward:
• 7:35
  even defensive interventions designed to remove hazardous capabilities leave exploitable residues in the model's internal representations, suggesting that unlearning mechanisms,
• 7:44
  like KV cache optimizations, carry security implications that operate beneath the prompt layer.
• 7:49
  The same structural pattern surfaces at the agent layer.
• 7:52
  The Decoder reports that Zenity Labs disclosed a vulnerability dubbed "AgentForger" in Open A I's Workspace Agents, whereby a single manipulated chatgpt.
• 8:02
  com link could autonomously create and publish a rogue AI agent under a victim's identity.
• 8:07
  The disclosure identifies a new class of attack—"agent trust failure"—in which autonomous agents operating under legitimate user identities bypass traditional security tools designed for users and endpoints.
• 8:17
  AgentForger extends the day's broader finding from model internals and serving infrastructure into the deployment ecosystem:
• 8:24
  the attack targets the trust relationships that agents inherit from their creators rather than any specific prompt or input.
• 8:30
  Read collectively, these four findings trace a coherent shift in L L M security research.
• 8:35
  HijackKV exposes an attack surface in serving-system architecture; the LUNAR red-team shows that unlearning edits leave recoverable structural traces;
• 8:44
  AgentForger demonstrates that agent trust relationships constitute a bypassable security boundary; and the sound probabilistic safety framework offers a formal-methods response,
• 8:53
  computing rigorous bounds where empirical methods provide none.
• 8:56
  The shared implication is that L L M vulnerabilities are not confined to adversarial inputs but permeate cache architectures, unlearning mechanisms,
• 9:05
  and agent trust models—domains where only structural guarantees, not additional red-teaming, can establish safety.
• 9:12
  Physics-Informed ML Is Maturing from Proof-of-Concept to Guaranteed-Constraint Methods.
• 9:17
  The day's evidence points to a coherent shift in physics-informed machine learning: rather than adding physics-aware loss penalties and hoping the optimizer respects them,
• 9:25
  new methods enforce conservation laws, stability conditions, and uncertainty bounds through architectural construction.
• 9:32
  This transition is visible across PDE solving, computational fluid dynamics, constitutive modeling, and even model-based reinforcement learning.
• 9:39
  PG-KINN exemplifies the move away from soft penalty formulations.
• 9:43
  It introduces a Petrov–Galerkin weak formulation in which the trial space is a Kolmogorov–Arnold Network with learnable B-spline activations and the test space is an independent, fixed,
• 9:53
  compactly supported piecewise-polynomial space evaluated via Gauss–Legendre quadrature.
• 9:58
  This design directly targets known weaknesses of strong-form PINNs—spectral bias and fragile loss weighting—as well as the restriction of energy forms to self-adjoint operators and the need
• 10:08
  for fundamental solutions in boundary-integral forms.
• 10:10
  By embedding the physics in a principled variational structure rather than a softened constraint,
• 10:15
  PG-KINN replaces heuristic penalty balancing with a formulation whose mathematical properties are inherited from the weak form itself.
• 10:22
  The entropy-stable learned finite volume scheme for the 2D compressible Euler equations pushes this guarantee philosophy further into the hard-constraint regime,
• 10:31
  ensuring positive density and pressure by construction and employing an entropy-stable interior flux in the sense of Tadmor.
• 10:37
  Notably,
• 10:38
  this work provides what the evidence describes as a rare honest accounting of what learning contributes when hard physical guarantees are enforced and comparisons are made at equal computational cost
• 10:48
  rather than equal mesh.
• 10:50
  This introduces a productive tension within the field's trajectory: enforcing guarantees by construction is architecturally feasible,
• 10:57
  but the evidence explicitly frames it as carrying a performance price, suggesting that the transition from penalties to guarantees is not cost-free.
• 11:05
  Interval and fuzzy physics-augmented neural networks extend the guarantee paradigm from deterministic solutions to uncertainty quantification.
• 11:12
  iPANNs learn separate sparse lower, mean, and upper free energy density branches whose automatically differentiated stresses enclose noisy stress observations,
• 11:20
  and fPANNs embed those branches into a fuzzy-set framework.
• 11:24
  The resulting bounds are distribution-free and physics-consistent, avoiding the need to assume probability distributions that are difficult to justify under sparse or noisy mechanical data.
• 11:34
  Taken together with PG-KINN's variational guarantees and the entropy-stable scheme's structural admissibility,
• 11:39
  iPANNs and fPANNs suggest that the guarantee-oriented philosophy is generalizing beyond point estimates to the epistemic and aleatoric uncertainty regime.
• 11:48
  Koopman Dreamer carries this rigor into a different domain: model-based reinforcement learning.
• 11:53
  It replaces the generic neural deterministic latent transition in Dreamer-style world models with a Koopman-inspired spectral backbone composed of two-dimensional rotation-scaling blocks
• 12:03
  with bounded modal radii, explicitly parameterizing damping, persistence, and oscillation of latent modes.
• 12:09
  This provides a tunable, diagnostically interpretable mechanism for balancing error attenuation against long-term information retention in long-horizon latent imagination.
• 12:19
  While the evidence does not establish a direct architectural link between Koopman Dreamer and the PDE-focused methods,
• 12:25
  the shared intellectual commitment—parameterizing networks so that physical or dynamical guarantees hold by construction rather than by penalty—suggests a field-wide pattern in which mathematical
• 12:35
  structure is substituted for optimization-time hope.
• 12:38
  Production Agent Reliability Is Converging on Silent-Failure Detection as the Critical Gap.
• 12:43
  As AI agents transition from prototypes to production systems,
• 12:47
  the binding constraint is shifting from raw capability to the detection of behavioral failures that complete successfully from the system's perspective yet produce incorrect outcomes.
• 12:56
  Amazon Bedrock AgentCore optimization introduces an insights capability that targets precisely this problem—failures in deployed agents where dashboards show green while customers receive wrong outcomes.
• 13:07
  According to the official AWS announcement, this capability could shift agent operations from reactive trace inspection to proactive pattern detection,
• 13:15
  enabling teams to prioritize fixes by scope rather than gut feel.
• 13:19
  The problem framing is notable: silent behavioral failures are described as a critical gap in production AI agent observability, not a peripheral concern.
• 13:28
  This platform-level focus on silent failures is complemented by an end-to-end evaluation pipeline co-developed by Motorway and AWS PACE,
• 13:36
  which combines the open-source strands-agents-evals library for build-time testing with Amazon Bedrock AgentCore Evaluations for production monitoring.
• 13:44
  The blueprint addresses the gap of verifying that agents reliably call the right tools, reason coherently, and produce consistent outputs across repeated runs.
• 13:52
  Taken together, the two AWS-sourced announcements establish a lifecycle spanning pre-deployment and post-deployment detection:
• 13:59
  build-time testing via strands-agents-evals feeds into production monitoring via AgentCore, with the latter's insights capability specifically operationalizing silent-failure detection at runtime,.
• 14:09
  Academic research provides the diagnostic taxonomy that production tools operationalize.
• 14:14
  A paper accepted to the SIGIR 2026 SynthIR Workshop introduces a six-category taxonomy of silent failures in multimodal agentic search—modality shortcut, phantom grounding,
• 14:25
  wrong-evidence-right-answer, over-retrieval laundering, cross-modal contradiction, and provenance hallucination—covering pre-retrieval, during-retrieval, and post-retrieval stages.
• 14:35
  The paper reports that surface answer accuracy consistently overestimates true trajectory-level correctness, with trajectory correctness rate dropping 0.6 to 7.3 percentage points across models.
• 14:46
  Critically, the finding that stronger models and better tools shift rather than eliminate silent failures suggests that capability scaling alone does not resolve the detection gap.
• 14:55
  While the multimodal search taxonomy categorizes failure modes, DocOps – a preprint with unknown peer-review status – makes silent failures measurable through deterministic verification.
• 15:06
  It introduces a verifiable evaluation framework for L L M agents performing end-to-end document manipulation across XLSX, DOCX, PPTX, and PDF formats.
• 15:17
  The benchmark exposes concrete capability boundaries that prior benchmarks miss: even the best configuration, G P T-5.5 with Codex and skills, achieves only 0.671 overall success,
• 15:29
  with Excel workflows plummeting near zero on complex tasks.
• 15:33
  This deterministically verifiable approach complements the pattern-based detection of AgentCore by providing a ground-truth mechanism that makes silent failures quantifiable rather than merely observable,.
• 15:44
  The convergence is striking: platform tooling from a major cloud provider,, an academic taxonomy from a workshop paper,
• 15:50
  and a verifiable benchmark from a preprint all identify the same structural gap—the system's own success signals do not reliably indicate correct outcomes.
• 15:59
  The evidence collectively suggests that production agent reliability now hinges less on what agents can do and more on whether the field can reliably detect when agents succeed incorrectly.
• 16:09
  AI Infrastructure Investment Is Entering a Tension Between Scaling Ambition and Financial Sustainability.
• 16:15
  The day's evidence sketches a preliminary picture in which the supply side of AI compute is diversifying and intensifying even as the demand side begins to show concrete financial strain.
• 16:25
  On the supply side, a preprint reports full-parameter post-training of the 1.6-trillion-parameter DeepSeek-V4-Pro MoE model on an Ascend CloudMatrix384 SuperPOD with 910C NPUs,
• 16:37
  described as among the first publicly documented full-stack studies of trillion-parameter MoE post-training on Ascend NPUs.
• 16:45
  The preprint characterizes this as a meaningful signal for hardware diversification beyond Nvidia ecosystems, suggesting that SIMD-based accelerators can reportedly sustain frontier-scale training.
• 16:56
  This supply-side expansion is not confined to training: Alibaba's Zhenwu M890 super-node has reportedly adapted Qwen3.8, a 2.4-trillion-parameter model,
• 17:06
  and is now live on the Alibaba Cloud Bailian platform for inference services, with QbitAi reporting a claimed 1.5 times inference performance improvement in agentic reasoning scenarios.
• 17:16
  According to the same report, this is claimed to be the first domestic Chinese super-node to run a model exceeding 2 trillion parameters,
• 17:23
  extending the domestic-capability narrative from training into production deployment.
• 17:27
  The software layer is simultaneously under competitive pressure.
• 17:30
  Alibaba's T-Head announced at WAIC 2026 the open-sourcing of its self-developed AI software stack T-Head SAIL, the full-stack software powering its Zhenwu series AI chips, a move that,
• 17:42
  according to QbitAi, directly challenges Nvidia's cooda moat by converting a previously black-box software stack into a transparent, modifiable white-box.
• 17:50
  Taken together, these three developments suggest that the supply side of frontier-scale AI compute is broadening—across training hardware, inference hardware,
• 17:58
  and the software stack —intensifying competitive pressure on incumbent hardware economics.
• 18:03
  The demand-side counterpoint is stark.
• 18:05
  Ars Technica reports that Google posted Q2 2026 total revenue of $119.8 billion—beating analyst expectations—yet experienced its first-ever negative cash-flow quarter,
• 18:16
  driven by sharply increased AI infrastructure capital expenditures, with Search revenue at $63.3 billion and Google Cloud at $24.8 billion.
• 18:26
  The report frames this as a potentially pivotal moment where even the most profitable tech companies are spending so heavily on AI infrastructure that it erodes cash flow despite record revenues.
• 18:36
  The juxtaposition is instructive but should be treated as tentative rather than conclusive, given that the evidence draws on a preprint of unknown peer-review status and media reports.
• 18:45
  The simultaneous emergence of trillion-parameter training on non-G P U hardware, production inference on domestic Chinese super-nodes, open-sourced software stacks challenging cooda,
• 18:54
  and the first negative cash-flow quarter at a hyperscaler suggests—without yet establishing—that the economics of AI scaling are shifting from unbounded optimism toward a contested equilibrium
• 19:04
  between capability ambition and capital discipline.
• 19:08
  Briefly Noted.
• 19:09
  Several developments touch on model architecture and training efficiency.
• 19:13
  A preprint from Upstage describes Solar Open 2,
• 19:16
  a 250 billion-A15 billion mixture-of-experts language model that achieves a 1 million-token context window through a hybrid linear-softmax attention stack,
• 19:25
  reaching roughly one-quarter the memory and computation of all-softmax designs while a proxy-scale ablation reaches MMLU 0.55 with 3.2 times fewer training tokens than baseline.
• 19:36
  Another preprint, DynamicRubric, introduces a response-set-conditioned evaluator–policy co-evolution framework for L L M post-training,
• 19:44
  showing theoretically that relative evaluator score gaps among candidate responses serve as local optimization signals for policy updates and directly addressing the collapse of those gaps
• 19:53
  as policies improve.
• 19:55
  On the infrastructure side, an official pie torch announcement reports that Helion's high-level DSL now includes a TPU backend compiling to Google's Pallas,
• 20:03
  enabling pie torch-style authoring of performant TPU kernels without low-level Pallas expertise and allowing the same kernel set across G P U and TPU.
• 20:12
  A preprint on reliability-aware distillation reveals that standard knowledge distillation on the BanSum Bangla summarization benchmark yields only +0.0003 ROUGE-L over a cross-entropy baseline,
• 20:25
  with 51.3% of training samples producing KD gradients that actively oppose validation-loss improvement.
• 20:31
  In the alignment and self-improvement domain, a preprint challenges the assumption that sycophancy in LLMs is a single behavioral dimension by proposing three distinct modes—Passive Affiliative,
• 20:41
  Strategic Ingratiation, and Defensive Conflict-Avoidant—grounded in HEXACO personality facets,
• 20:47
  arguing that a single sycophancy direction or aggregate score masks mechanistically distinct failure modes requiring different interventions.
• 20:55
  Another preprint studies whether LLMs can extract and apply experiential abstractions from their own solution traces on 7,500 MATH training problems,
• 21:03
  finding that self-extracted abstractions match teacher quality and transfer to other datasets such as OlympiadBench and to other model families,
• 21:10
  potentially reducing dependence on expensive frontier teacher models.
• 21:14
  On the generative and agentic front, The Decoder reports that Black Forest Labs' Flux 3 generates videos with native audio up to 20 seconds long through joint training across image, video,
• 21:25
  and audio, built on a "Self-Flow" approach in which a single model both generates and understands content,
• 21:31
  with early preference evaluations showing 93% preference over Luma Ray 3.2 among eight rivals.
• 21:37
  The Decoder also reports that Poolside's Laguna S 2.1,
• 21:41
  a 118 billion-total-parameter mixture-of-experts coding model with 8 billion active per token and support for up to one-million-token context windows,
• 21:50
  scores 70.2% on Terminal-Bench 2.1 and 40.4% on DeepSWE,
• 21:55
  outperforming models in its weight class and sometimes approaching the performance of systems 10 to 20 times its size by improving agentic behaviors such as persistence, verification,
• 22:04
  and revising failed approaches.
• 22:06
  A preprint from Alibaba Token Foundry presents a unified full-song generation framework combining discrete hierarchical autoregressive planning with continuous flow-matching rendering
• 22:16
  across lyrics-to-song, instrumental music, and cover song tasks, producing full-length songs competitive with leading commercial systems.
• 22:24
  Finally, a preprint presenting a controlled horse race comparing quantum fidelity kernels, projected quantum kernels,
• 22:30
  and a classical RBF kernel for predicting 20-day-ahead cross-sectional stock returns in the Chinese A-share market finds no quantum advantage,
• 22:38
  offering a methodologically rigorous negative result that challenges the growing applied literature reporting quantum kernel advantages in financial prediction.
• 22:46
  Synthesis and Outlook.
• 22:48
  The day's evidence converges on a single tension: AI is simultaneously hardening its technical foundations and exposing the structural fragility beneath them.
• 22:56
  Embodied AI's pivot from capability demos to deployment-readiness gaps and production agents' convergence on silent-failure detection reinforce each other editorially—both reveal that the binding
• 23:06
  constraint has shifted from what models can do to whether their behavior remains trustworthy once deployed.
• 23:11
  Physics-informed ML's transition to guaranteed-constraint methods offers one template for resolving this gap, replacing heuristic approximations with mathematical guarantees,
• 23:21
  and its rigor stands in instructive contrast to L L M security research, which is still uncovering structural vulnerabilities—KV-cache reuse, unlearning failures,
• 23:30
  unprovable safety bounds—that lack analogous formal foundations.
• 23:34
  Meanwhile, infrastructure investment pressures complicate every other claim: if scaling economics are entering a contested equilibrium between ambition and capital discipline,
• 23:43
  the resources available to close deployment-readiness, security, and reliability gaps may contract precisely when those gaps are being recognized as systemic rather than incidental.
• 23:51
  The brief-noted contributions across evaluation, multimodal generation, medical AI, and community safety suggest these tensions are broadening but not yet resolving.
• 24:01
  The open question is whether the field's emerging formal methods—conservation-law enforcement, distribution-free bounds,
• 24:07
  structural security guarantees—can mature fast enough to offset the reliability debt accumulating as deployment outpaces verification,
• 24:14
  or whether financial sustainability pressures will force a narrowing of ambition before that maturity arrives.
• 24:20
  This review draws on 30 developments: 18 Tier A research sources, 3 Tier B first-party sources, and 9 Tier C/D secondary or community sources.
• 24:29
  The firmest claims rest on the Tier A work, while the first-party and community sources should be read as directional;
• 24:35
  stronger confidence would require independent replication and primary-source confirmation of the self-reported results.
• 24:41
  This edition cites 30 sources; full links are available in the text edition.