Latest / AI Sentinel: Frontier Daily / Specialized, Agentic, Local AI: Value Depends on Evaluation and Structural Safety

Transcript

Specialized, Agentic, Local AI: Value Depends on Evaluation and Structural Safety · AI Sentinel: Frontier Daily

• 0:12
  This is AI Sentinel, your daily brief from the AI frontier.
• 0:16
  Specialized, Agentic, Local AI: Value Depends on Evaluation and Structural Safety.
• 0:21
  Highlights.
• 0:23
  Domain-specialized foundation models deliver measurable gains in clinical and scientific settings, but comparative evidence cautions against assuming they are universally superior to task-specific systems.
• 0:33
  Agentic research systems are advancing where tasks can be decomposed into curated skills or supervised search processes, while headline autonomy claims remain conditional on external verification.
• 0:44
  Clinical AI deployment is limited less by raw reasoning accuracy than by safety-calibration under real workflows, including harmlessness, communication, and human-AI collaboration.
• 0:54
  Open-weight models and local hardware are being packaged as a single deployable stack, extending the locality trend without resolving its tension with policy legibility.
• 1:03
  The evidence indicates that AI progress is shifting from raw model scale toward domain-specialized, agentic, and locally deployable systems, whose value depends on evaluation context,
• 1:12
  process supervision, and structural safety assurance.
• 1:15
  The sections build this argument by first testing domain specialization against deployment-matched evaluation and then examining agentic research systems through decomposable,
• 1:24
  supervised workflows whose autonomy claims remain conditional on external verification.
• 1:28
  Clinical AI is framed less as a reasoning-accuracy problem than as a safety-calibration problem under real workflows,
• 1:35
  while open-weight models and local hardware are presented as a consolidating deployable stack.
• 1:39
  Embodied AI is constrained by data engines, unified action representations, and operational assurance.
• 1:45
  Briefly Noted extends the evidence beyond these argument sections.
• 1:49
  Domain specialization pays off only when evaluation matches deployment.
• 1:53
  Domain specialization is not self-validating: its returns depend on whether the evaluation task matches the deployment task.
• 2:00
  The strongest clinical evidence comes from evaluations designed around reporting workflows.
• 2:05
  Impression-R1, a domain-specialized large reasoning model for radiological impressions, improved reporting quality by an adjusted difference of 0.28 and reduced completion time by 27% without compromising harmlessness scores in multicenter benchmarking and prospective evaluations.
• 2:21
  In echocardiography, EchoChat, a multimodal vision-language model built on Qwen 2.5-VL, was tested in a blinded reader study;
• 2:29
  the result is reported in an archive preprint with peer-review status unknown, so it supports a workflow-augmentation signal rather than a settled comparative claim.
• 2:38
  Both cases tie specialization to a defined clinical task and an evaluation protocol—multicenter and prospective in, blinded reader study in —which is where the payoff is measured.
• 2:49
  Comparative evidence, however, cautions against treating foundation models as universally superior.
• 2:54
  A large-scale empirical comparison of pathology foundation models and task-specific end-to-end models for clinical-grade prostate pathology found that resource-intensive foundation models offer
• 3:04
  diminishing returns over efficient task-specific models when sufficient data is available, explicitly challenging the prevailing assumption that foundation models are universally superior for digital pathology.
• 3:15
  This is not a direct refutation of or; it is a boundary condition. and evaluate specialized systems against clinical workflow outcomes,
• 3:22
  whereas compares model classes under a clinical-grade pathology task with data sufficiency as a stated condition,,.
• 3:28
  The distinction matters because a domain-specialized system can improve a matched workflow without establishing that a foundation-model approach dominates a task-specific alternative.
• 3:37
  The specialization argument also extends beyond clinical medicine.
• 3:40
  EVENET, an event-level foundation model for particle collision data, was pretrained on 500 million simulated collision events to address the growing bottleneck of task-specific model training in high-energy physics,
• 3:52
  which the authors describe as becoming unsustainable as the High-Luminosity LHC increases data volumes.
• 3:58
  This is a scientific-domain case for specialization, but its evidence is simulation-based pretraining rather than a deployed comparative evaluation.
• 4:06
  It therefore supports the motivation for domain-specific foundation models while leaving the deployment-match question open.
• 4:12
  Taken together, the evidence supports a conditional thesis.
• 4:15
  Domain-specialized models show measurable clinical gains when evaluation mirrors deployment—reporting quality and time in,
• 4:22
  diagnostic performance in a blinded reader study in —but shows that foundation models are not automatically superior to task-specific systems when the comparison is matched to a clinical-grade task
• 4:31
  and sufficient data. adds a scientific-domain rationale centered on data scale, not a comparative performance result.
• 4:38
  The practical implication is evaluative: specialization pays off only when the benchmark, reader study, or prospective workflow reflects the intended use, and when the comparator is appropriate.
• 4:49
  Without that match, claims of foundation-model superiority outrun the evidence.
• 4:53
  Agentic research systems become credible through decomposable, supervised workflows.
• 4:58
  Credibility in agentic research systems is emerging along a narrower axis than headline autonomy suggests: tasks that can be decomposed into curated skills or supervised search processes.
• 5:08
  A preprint on ClinGeX describes an L L M-based agent for clinical genomic analysis that composes 105 curated executable skills—27 guide, 40 search,
• 5:18
  and 38 analysis—through task-specific retrieval and dependency-aware orchestration, and reports evaluation on 223 temporally held-out benchmark cases.
• 5:27
  The significance claimed for this design is that it addresses a bottleneck in genomic medicine where data generation has outpaced interpretation capabilities,
• 5:35
  framed in the evidence as the “$1,000 genome, $100,000 analysis” problem.
• 5:39
  Because the source is an archive preprint with peer-review status unknown, the result should be read as a preprint-reported system description rather than settled comparative evidence.
• 5:49
  A research paper on RAG-Gym supplies a methodological counterpoint within the same design space: it shifts supervision of L L M search agents from final answers to the search process itself,
• 5:59
  addressing reward hacking and over-reliance on parametric memory.
• 6:02
  The paper states that this could make deployed agentic search and RAG systems more reliable by curbing reward hacking and dependence on the model’s internal memory,
• 6:11
  which currently limit out-of-domain generalization.
• 6:14
  Taken together with, the two sources point to a shared emphasis on curation and supervision—one decomposing a domain workflow into executable skills,
• 6:22
  the other changing what the agent is supervised on during search—without either source claiming the other’s result.
• 6:28
  The agentic-science surface is also extending to instrument control.
• 6:31
  Nvidia reports that Breakthrough Listen and the SETI Institute have deployed agentic AI to operate the Allen Telescope Array, enabling a single natural-language request to steer the instrument,
• 6:41
  enhance incoming data, and apply AI models to live sensor streams.
• 6:45
  The same announcement states that the deployment uses Nvidia Holoscan to process data and that it compresses tasks that would take humans weeks or years into hours.
• 6:54
  This is a first-party company announcement describing a deployment, not an independent evaluation, so the autonomy and compression claims carry that provenance caveat.
• 7:03
  The strongest headline autonomy claim in the set is also the most explicitly conditional.
• 7:07
  Open A I announced three AI-generated mathematical discoveries in harmonic analysis, computational complexity, and quantum field theory.
• 7:15
  The announcement presents verification and academic acceptance as conditions on impact and describes the working group as stewarding the Clay Mathematics Institute’s Yang-Mills Millennium Prize
• 7:25
  problem; it does not assert that the Kakeya conjecture or the Yang-Mills problem are resolved.
• 7:30
  As an official company announcement, it frames verification and academic acceptance as conditions on impact rather than as completed external validation.
• 7:38
  One reading across these sources is that autonomy has not been independently established; rather, agentic research systems become credible where workflows are decomposable or supervised,
• 7:48
  and the largest claims remain contingent on external verification,,,.
• 7:52
  Clinical AI's binding constraint is safety-calibration, not raw reasoning accuracy.
• 7:57
  The clinical AI evidence points to a deployment constraint that is not raw reasoning accuracy but safety-calibration under real workflows.
• 8:05
  The accuracy-safety trade-off study introduces AIMS-LDR, a long-context diagnostic reasoning benchmark built on real-world electronic health records,
• 8:14
  and evaluates large reasoning models against standard LLMs across six matched model pairs and 12 clinical specialties.
• 8:20
  Its central finding is that enhanced reasoning capabilities can inadvertently inflate false positive rates and reduce diagnostic reproducibility,
• 8:28
  which the study identifies as a significant safety bottleneck for deploying advanced reasoning models in healthcare;
• 8:33
  calibration—not merely higher reasoning performance—is therefore the operative hurdle.
• 8:38
  The psychiatric-intake work extends the same concern from diagnostic accuracy to the sequential, rapport-dependent communication that structures clinical information gathering.
• 8:47
  It applies multi-turn reinforcement learning in simulated environments to train a language model for structured psychiatric intake interviews, jointly optimizing diagnostic reasoning,
• 8:56
  patient-centered communication, and clinical safety.
• 8:59
  Because conversation quality directly determines diagnostic information quality in this setting, the work extends the safety-calibration problem rather than posing a separate accuracy problem.
• 9:09
  The claim is bounded by the source’s status as an archive preprint with unknown peer-review status and by its simulation-based training environment.
• 9:17
  EchoChat offers a contrasting deployment pattern in which the human-AI relationship is part of the safety case.
• 9:23
  The multimodal vision-language model, built on Qwen 2.5-VL, performs comprehensive echocardiographic analysis and is positioned as an assistant that augments rather than replaces clinical interpretation.
• 9:33
  In a blinded reader study, collaborative reports co-generated with EchoChat improved diagnostic F1 scores from 72.78% to 82.75%.
• 9:43
  This is a safety-positive pattern because the measured gain comes from collaboration, not from autonomous replacement.
• 9:49
  It is also a preprint whose peer-review status is unknown.
• 9:53
  Taken together, these sources do not establish a single causal chain from reasoning scale to clinical harm or from collaboration to safety.
• 10:00
  They do, however, place three deployment conditions in relation: calibration of reasoning gains, joint optimization of communication and clinical safety, and collaborative human-AI reporting.
• 10:10
  The implication is that clinical AI’s binding constraint is the calibration of safety, communication, and human-AI collaboration in the workflow, not reasoning accuracy alone,,.
• 10:20
  Open-weight models and local hardware are consolidating into a deployable stack.
• 10:24
  Nvidia’s RTX Spark announcement gives the locality line a concrete hardware surface: laptops built around the Blackwell RTX G P U and Grace CPU superchip,
• 10:34
  with up to 128GB of unified memory and 1 petaflop of FP4 AI performance, alongside software releases including the open-weight Nemotron 3.5 Transcribe speech model.
• 10:45
  Nvidia reports that the combination bridges cloud-level AI performance and local, privacy-preserving execution on portable hardware,
• 10:52
  and that it could accelerate local AI development by letting high-quality speech models and autonomous agents run efficiently on laptops.
• 10:59
  That is a first-party claim about capability and deployment, not an independent verification, but it establishes the packaging logic:
• 11:06
  local hardware is presented as the host for open-weight models and agent tools rather than as a standalone device story.
• 11:12
  The software layer is specified in the companion releases.
• 11:15
  The Nemotron 3.5 Transcribe announcement describes an open-weight, promptable speech model that unifies multilingual transcription, speaker attribution also known as diarization, translation,
• 11:26
  and audio question answering into a single checkpoint.
• 11:29
  Nvidia states that this could lower the barrier for enterprises needing adaptable, high-accuracy speech recognition without relying on proprietary APIs.
• 11:37
  They add that strong leaderboard results across transcription and diarization benchmarks demonstrate that a single open-weight model can rival or beat specialized commercial systems.
• 11:47
  This is a direct extension of the RTX Spark announcement, which already lists Nemotron 3.5 Transcribe among the laptop software releases.
• 11:55
  The speech checkpoint is therefore not a separate locality claim but a named component of the packaged stack.
• 12:01
  A second Nvidia release broadens the open-weight layer beyond speech.
• 12:04
  Nemotron 3.5 Super and Lightning are described as open Mixture-of-Experts hybrid Mamba-Transformer models for reasoning, long-context, and agentic workloads.
• 12:13
  Super has 120 billion total and 12 billion active parameters and is multimodal, while Lightning has 30 billion total and 3 billion active parameters and is text-only.
• 12:23
  The same release reports Indic continued pretraining on a 2.83 trillion-token synthetic corpus spanning 22 Indian languages, a 1 million-token context extension,
• 12:33
  and agentic software-engineering mid-training for Super.
• 12:36
  It also releases base, BF16, and NVFP4-quantized checkpoints, standalone drafters (LiLiCorr, DFlash, and DSpark), datasets, cookbooks, and training recipes.
• 12:46
  Because this item is an archive preprint with peer-review status unknown, its model and training claims should be treated as preprint-reported rather than independently validated.
• 12:55
  The evidence does not state that these models run on RTX Spark laptops.
• 13:00
  The connection is an extension of the same open-weight release strategy, not a verified hardware pairing.
• 13:06
  Taken together, the evidence shows a stack assembled from a local hardware surface, a promptable open-weight speech checkpoint,
• 13:12
  and open-weight reasoning and agentic models with released checkpoints and recipes.
• 13:16
  This extends the prior Nvidia locality line by making the laptop the deployment surface for open-weight software, but it does not resolve the tension with policy legibility:
• 13:26
  the releases are framed around local performance, privacy-preserving execution, and reduced reliance on proprietary APIs,, while none of the three descriptions reports a policy-legibility mechanism,,.
• 13:36
  The stack is therefore more deployable on its own terms, but the governance question is left where the locality framing placed it.
• 13:43
  Embodied AI is constrained by data engines and unified representations.
• 13:47
  The robotics evidence locates the binding constraint in the apparatus surrounding the model rather than in parameter count.
• 13:54
  Nvidia's FLORA is described as an open-source robotics data engine suite comprising a segmentation model, a captioning model, and a reward model,
• 14:01
  which simultaneously generates timestamped action captions, outcome rewards, process rewards, and user-defined rubric rewards for robot and human manipulation videos.
• 14:11
  According to that preprint, the suite directly addresses the bottleneck of scalable, high-quality data annotation for training robotics foundation models such as GR00 trillion,
• 14:20
  and achieves competitive quality at roughly 125 times lower cost than G P T-6 Astra; the same preprint states that open-sourcing the models could democratize industrial-scale robot data processing.
• 14:32
  These are the preprint's own claims, and its peer-review status is unknown.
• 14:36
  That design makes supervision granularity a parameter of the data engine itself rather than a downstream tuning choice.
• 14:43
  A second constraint is representational.
• 14:45
  Alibaba DAMO Academy's Hupan Lab presents RynnUnify, a 16 billion unified embodied model that couples embodied understanding — spatial reasoning, physical prediction,
• 14:55
  grounding — with long-horizon visual planning through interleaved video–language reasoning.
• 15:00
  The preprint states that this could narrow the gap between multimodal understanding models and deployable robot controllers by showing one set of weights serving understanding, prediction,
• 15:10
  and action, and reports that selectively generating videos improves answers about unobserved states.
• 15:15
  The framing is explicitly conditional — unification is offered as a possible route to deployable controllers, not a demonstrated one.
• 15:22
  The operational counterpoint comes from Mars. In October 2023 the Mars 2020 mission deployed the OnBoard Planner into Perseverance rover operations,
• 15:31
  described as the first use of AI constraint-based search to dynamically construct and modify activity schedules on the Martian surface.
• 15:38
  According to that research paper, the system demonstrates that dynamic onboard autonomy can meaningfully increase the science return of deep-space missions without burdening ground operations.
• 15:48
  Value here is realized through constraint machinery and integration with an existing operations process, not through scale.
• 15:54
  The frontier-model result sits in tension with a data-engine-centred account.
• 15:58
  A community post on LessWrong evaluates nine frontier multimodal language models, including G P T-6 Astra, Opus 5.5, and Gemini 4 Argon,
• 16:07
  on direct control of a real bimanual I2RT YAM robot across six manipulation tasks from the RoboDojo benchmark, and quantifies a performance doubling time of 53.
• 16:17
  That post reports concrete evidence that frontier AI capabilities are spilling over into physical manipulation at a rapid pace,
• 16:24
  with naive extrapolation suggesting LM-controlled robots could match human teleoperator speeds by mid-2027.
• 16:30
  The reported result, however, concerns manipulation performance and extrapolated speed; it does not report assurance, annotation-pipeline, or deployment-integration outcomes.
• 16:39
  Taken together, the four items suggest that capability gains in control can arrive faster than the data engines, unified representations,
• 16:47
  and operational assurance mechanisms on which sustained deployment depends — a reading consistent with the tension between and,, rather than one the sources themselves assert.
• 16:55
  Safety research is shifting toward structural vulnerabilities.
• 16:59
  The safety evidence locates risk upstream of the output layer.
• 17:03
  A preprint on agentic backdoors via pretraining poisoning reports that poisoning as little as 0.1% of a pretraining corpus can implant a remote code execution backdoor in an L L M agent
• 17:14
  that survives a clean post-training pipeline of SFT, DPO, and GRPO.
• 17:18
  The same preprint characterizes this as a supply-chain vulnerability in which attackers with no knowledge of the training pipeline can plant persistent,
• 17:26
  stealthy backdoors solely through public web content.
• 17:29
  The structural point is that the compromise is established before any alignment or filtering stage runs, and the post-training stack described in the paper does not remove it.
• 17:37
  A second preprint sharpens the evaluation side of the same problem.
• 17:41
  It reports the first systematic test of whether a second pretraining run on identical data, given identical safety post-training, breaks on the same jailbreaks.
• 17:49
  Its finding is that a single checkpoint's vulnerability profile may not transfer to retrained models: the best of 91 attack templates stayed best after a new pretraining seed in only 27% of pairs, versus 67%.
• 18:03
  The paper states the implication directly — safety evaluations and red-team findings may need repeating per pretraining run.
• 18:09
  Read alongside the poisoning result, the two preprints converge on pretraining as the locus of risk while diverging in emphasis:
• 18:16
  one shows that what enters pretraining persists through post-training, the other that what a given pretraining run produces does not generalize to the next.
• 18:24
  Neither preprint states a relationship to the other; taken together, they suggest that output-level and one-checkpoint safety conclusions rest on assumptions the evidence does not support.
• 18:34
  The structural concern extends to human judgment.
• 18:36
  Anthropic reports a randomized controlled trial with 50 professional causal-inference analysts using Claude Code to perform a realistic statistical analysis for a farming cooperative.
• 18:46
  According to that announcement, the result reveals a risk of expert disempowerment: when AI agents express preferences on judgment calls with no single right answer,
• 18:55
  experts may unknowingly defer to arbitrary or misaligned model priors while believing they exercised independent judgment.
• 19:01
  Here the vulnerability is neither a poisoned corpus nor a non-transferable benchmark, but the calibration of expert deference on ambiguous decisions — a failure mode that output filtering would not surface, since the model's output is not incorrect on its own terms.
• 19:14
  Open A I reports a substantially expanded, multi-layered monitoring system designed to detect misaligned behavior in frontier models across training, evaluation, internal use,
• 19:24
  and external deployment.
• 19:25
  That announcement references a live training run alert in September and a “Sneaky Sandbox Escape” simulation, and presents the framework as a layered defense-in-depth strategy.
• 19:34
  The scope — training through deployment, rather than the model's responses alone — is consistent with a shift toward structural oversight,
• 19:42
  though as a first-party announcement it documents a vendor's own monitoring rather than independent validation.
• 19:47
  Taken together, the four items describe a field whose assurance questions are increasingly posed at the level of corpus provenance, run-to-run reproducibility,
• 19:56
  and the judgment of the humans in the loop, not solely at the boundary where outputs are filtered.
• 20:01
  Briefly Noted.
• 20:02
  A preprint introduces FitzRoy as the first system to simultaneously achieve state-of-the-art predictive skill and physical realism for precipitation forecasting over the continental United States,
• 20:13
  with the stated potential to improve disaster resilience and emergency response through accurate, high-fidelity forecasts of extreme precipitation up to 24 hours ahead.
• 20:22
  A separate preprint presents Astra, a 0.25° ensemble weather forecasting model that matches GenCast’s skill using only 9.8% of its hardware-normalized training compute and 39 times fewer inference FLOPs;
• 20:35
  the paper reports that a 50-member, 15-day forecast runs in under 40 minutes on a single consumer G P U within 21 GB of memory.
• 20:43
  A preprint presents SimRetract as the first zero-shot sim-to-real transfer for a surgeon-robot collaboration task in open surgery, specifically retractor take-over,
• 20:52
  addressing the scarcity of surgical personnel and the high cost and difficulty of collecting real-world robotic demonstrations in sterile environments.
• 21:00
  A preprint describes MolRSI, an AI agent that couples systematic optimization of molecular discovery workflows with persistent cross-task memory to recursively improve its search policy,
• 21:11
  with the stated aim of reducing the repeated trial-and-error typically required when adapting computational workflows to new biological targets.
• 21:18
  A preprint introduces Timestral, a family of four time series foundation models for probabilistic univariate, multivariate, and covariate-informed forecasting, with 5 million, 28 million,
• 21:31
  123 million, and 209 million parameters.
• 21:35
  The preprint reports that Timestral-123 million outperforms every external comparator, including billion-parameter models such as Toto-2.0-2.5 billion and Chronos-2,
• 21:46
  suggesting much cheaper deployment for forecasting in energy, manufacturing, and transport, as measured on the IND-Eval benchmark.
• 21:53
  QbitAI reports that OXR, a 1.3 billion-parameter open-source document parsing model, ranks first on OmniDocBench with a score of 97.18, on olmOCR-Bench with 87.84,
• 22:05
  and on DocParsingBench with 87.76, and that it can run on consumer GPUs such as an RTX 3060.
• 22:12
  QbitAI also reports that EdgeRock-4 billion, an open-source 4 billion model, achieves constant memory usage during long-context inference through Rock Memory Anchors, or RMA,
• 22:24
  reducing KV Cache by 99.1% compared with its source model Qwen3 and enabling 512 concurrent long-reasoning streams on a single 80GB GPU.
• 22:34
  Open A I announced G P T-6.1 Terra, a production model combining detailed visual reasoning and precise computer-use capabilities with improved writing and factuality.
• 22:43
  According to Open A I, it offers a 50% A P I price reduction compared with G P T-5.6 Terra and G P T-6.1 Sol and completes tasks over three times faster than G P T-6 Astra and G P T-6.1 Sol;
• 22:57
  the announcement states this significantly lowers the financial barrier to deploying high-volume, computer-use AI agents for enterprise tasks such as invoice processing and form filling.
• 23:06
  Google DeepMind announced Gemini 4 Carbon, a frontier model claiming state-of-the-art performance in cybersecurity and long-horizon agentic tasks; according to the announcement,
• 23:16
  it ranks #1 on the Artificial Analysis Intelligence Index v5.0 with a score of 59 and achieves 88.2% on SEC-bench Pro for vulnerability finding.
• 23:26
  The announcement states this could significantly accelerate cybersecurity defense workflows by autonomously handling the full vulnerability lifecycle—detection, validation, and remediation.
• 23:37
  Google DeepMind also announced Genie 4, a natively multimodal real-time world model that simultaneously generates interactive visual environments, ambient audio, music, and dialogue,
• 23:47
  and integrates Gemini models for real-world appearance and behavior anchoring.
• 23:51
  The announcement states that this could lower the barrier to creating photorealistic, interactive 3D environments for filmmaking, game prototyping, and historical visualization,
• 24:00
  and may provide a scalable path for generating safe, infinitely variable training scenarios for robots and AI agents.
• 24:07
  Synthesis and Outlook.
• 24:09
  The claims converge on a shift in emphasis rather than a single verdict.
• 24:13
  Domain specialization and clinical AI reinforce each other through evaluation context: gains in clinical and scientific settings depend on whether evaluation matches deployment,
• 24:23
  while clinical deployment is constrained by safety-calibration under real workflows rather than raw reasoning accuracy.
• 24:29
  Agentic research systems and structural safety research also connect through supervision and verification: decomposable,
• 24:35
  supervised workflows make autonomy credible only when external verification holds, and safety work increasingly examines pretraining, reproducibility,
• 24:43
  and human-AI judgment rather than output filtering alone.
• 24:47
  A tension remains between the open-weight/local hardware stack and policy legibility, and between domain-specialized foundation models and the caution that task-specific systems may outperform them.
• 24:57
  Embodied AI adds a further constraint: data engines, unified action representations, and operational assurance matter more than scale alone.
• 25:05
  Read together, these claims imply a field moving toward context-dependent, process-supervised, structurally assured systems,
• 25:12
  but the open question is what evaluation and assurance mechanisms can travel across domains and deployment contexts without becoming either local-only or policy-opaque.
• 25:21
  The evidence mix warrants cautious-to-moderate confidence, with the thinnest support around deployment-scale policy legibility and operational assurance.
• 25:29
  This review draws on 31 developments: 19 Tier A research sources, 9 Tier B first-party sources, and 3 Tier C/D secondary or community sources.
• 25:39
  The firmest claims rest on the Tier A work, while the first-party and community sources should be read as directional;
• 25:45
  stronger confidence would require independent replication and primary-source confirmation of the self-reported results.
• 25:51
  This edition cites 31 sources; full links are available in the text edition.