Latest / AI Sentinel: Frontier Daily / From Capability Demos to Integrated Systems: AI’s New Frontier

Transcript

From Capability Demos to Integrated Systems: AI’s New Frontier · AI Sentinel: Frontier Daily

• 0:12
  This is AI Sentinel, your daily brief from the AI frontier.
• 0:16
  From Capability Demos to Integrated Systems: AI’s New Frontier.
• 0:20
  Highlights.
• 0:22
  World model research is converging on architectures that couple prediction with action generation, yielding more reliable and generalizable robot policies.
• 0:30
  Long-horizon agents increasingly depend on explicit memory structures and attribution mechanisms to track evidence and decisions, rather than monolithic reasoning.
• 0:39
  Safety alignment is being reframed as a foundational capability that must be internalized during training and continuously monitored, rather than applied post-hoc.
• 0:47
  New evaluation benchmarks are moving beyond aggregate accuracy to expose specific, often overlooked failure modes in models and evaluation protocols.
• 0:55
  Training efficiency gains now come from selecting the most informative data and distilling only high-value signals, reducing compute while maintaining or improving performance.
• 1:04
  Recent advances in artificial intelligence mark a transition from isolated demonstrations of capability toward integrated systems designed to manage uncertainty, coordinate heterogeneous components,
• 1:14
  and align closely with human intent.
• 1:16
  This review examines that shift through converging developments in world models, agentic frameworks, and safety mechanisms.
• 1:23
  The first sections show how predictive architectures are coupling with action generation to yield more reliable robot policies,
• 1:30
  and how long-horizon agents depend on explicit memory and attribution rather than monolithic reasoning.
• 1:35
  Subsequent sections address safety as a foundational capability internalized during training, evaluation benchmarks that expose hidden failure modes,
• 1:44
  and training methods that leverage targeted data and distillation for efficiency.
• 1:48
  Multimodal systems further illustrate progress through improved temporal and visual grounding.
• 1:53
  Together, these threads argue that the field’s trajectory is defined less by any single breakthrough than by the deliberate integration of perception, reasoning, and control under human oversight.
• 2:03
  World Models Move from Prediction to Actionable Control.
• 2:06
  The convergence of prediction and action generation is becoming a defining architectural principle in world model research,
• 2:13
  with recent work moving beyond passive forecasting toward models explicitly designed to inform and execute robot policies.
• 2:20
  This trend is visible across three complementary developments: the unification of world models across embodiments, the explicit coupling of visual prediction with action conditioning,
• 2:30
  and the formalization of generalization benchmarks that test the robustness of these prediction-action systems.
• 2:35
  A significant step toward generalist systems is Ropedia W0, presented as the first unified egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments.
• 2:46
  The model introduces Xp-Data, a curated corpus of approximately 10,000 hours of egocentric human video and 3,000 hours of robot video with unified geometric, kinematic, and semantic annotations.
• 2:58
  According to the preprint, this approach could reduce the need for embodiment-specific action interfaces and adaptation when building world models for loco-manipulation,
• 3:06
  potentially providing a more data-efficient route to robot prediction and downstream policy evaluation, model-based planning, and world-action modeling.
• 3:14
  This work suggests that scaling from abundant human egocentric video may yield world models that generalize across embodiments, a prerequisite for broadly applicable action generation.
• 3:25
  Complementing this breadth is TrajWAM, which tightens the coupling between prediction and action at the architectural level.
• 3:31
  TrajWAM introduces a trajectory-guided world action model that explicitly grounds future visual imagination in predicted target-object motion and conditions action generation on the resulting future representation.
• 3:43
  The model also introduces a decoupled distillation scheme that preserves this anticipation pathway during one-step inference.
• 3:50
  The preprint reports that this design could make dynamic manipulation more responsive by separating target-motion anticipation from scene-level video prediction and reducing iterative video-action
• 4:00
  denoising latency, potentially helping robots act on moving objects without executing outdated actions, particularly where real-time replanning is required.
• 4:08
  Where Ropedia W0 addresses the scale and diversity of training data, TrajWAM addresses the inference-time efficiency of the prediction-action loop,
• 4:17
  suggesting that the field is optimizing both the breadth and the latency of world-action models.
• 4:21
  The third strand addresses evaluation.
• 4:24
  AnyGoal introduces a benchmark and open-source library for testing whether world models generalize beyond expert rollouts,
• 4:30
  formalizing this as compositional generalization—whether a model can recombine familiar state factors and actions into valid dynamics when goals change.
• 4:38
  The preprint states that this could make world-model evaluation more diagnostic by separating behavior-biased memorization from reusable dynamics,
• 4:46
  helping researchers identify when models fail under goal shifts even without novel states or actions, and guiding data collection, architecture design, and planning benchmarks.
• 4:55
  Taken together, these three preprints suggest a coordinated push: Ropedia W0 expands the data and embodiment scope of world models, TrajWAM refines the prediction-action coupling for dynamic tasks,
• 5:06
  and AnyGoal provides the diagnostic tools to test whether such models have learned reusable dynamics rather than memorized rollouts.
• 5:13
  Agentic Systems Rely on Structured Memory and Attribution.
• 5:17
  Effective long-horizon agents are increasingly distinguished not by the raw power of a single reasoning pass, but by the explicit structures they use to store, retrieve, and justify information.
• 5:27
  The evidence points to a convergence on two architectural principles: memory that is actively shaped by experience, and attribution chains that tie every decision back to its originating evidence.
• 5:38
  The most direct articulation of the attribution principle comes from ClueFlow, which introduces attributed research-state transitions for long-horizon search agents.
• 5:46
  The framework builds Source-State-Final attribution chains that explicitly connect evidence-acquiring browser actions to Evidence and Candidate updates, and ultimately to the final answer.
• 5:56
  This design has a specific training implication: by supervising only attributed decisions rather than all turns, long-horizon search training could become more efficient,
• 6:05
  and reinforcement-learning credit assignment could improve by linking browser actions, state updates, and final answers.
• 6:11
  The value here is not merely that the agent reaches an answer, but that the path to that answer is legible and attributable, which in turn enables more targeted supervision.
• 6:20
  Complementing this external traceability is an internal memory architecture proposed by CoMem.
• 6:25
  CoMem introduces Compounding Memory, which makes retrieved experience an input to the learned memory itself during training and inference,
• 6:32
  rather than feeding retrieved and parametric memory outputs side by side to a solver.
• 6:36
  This is a structural departure from treating past experience as mere context; instead, the experience actively shapes the learned memory representation.
• 6:44
  The reported potential for generalization to out-of-distribution tasks and for smaller-model guidance to stronger solvers suggests that memory design is not a peripheral concern but a central
• 6:54
  determinant of agent reliability.
• 6:56
  Taken together with ClueFlow, these two frameworks suggest a division of labor: CoMem addresses how an agent internalizes and composes its history,
• 7:04
  while ClueFlow addresses how that history is audited and linked to outcomes.
• 7:09
  A third piece, LatComm, addresses the efficiency of memory sharing across multiple models.
• 7:14
  LatComm introduces a latent communication interface that separates context encoding from question-conditioned reading, producing a query-independent,
• 7:22
  fixed-size packet reusable across questions and readable by a different frozen L L M, with only bridge modules trained.
• 7:28
  The reported 57 times receiver input reduction and 3 times faster answering on QuALITY indicate that structured,
• 7:35
  reusable memory representations can yield substantial practical gains in multi-L L M workflows.
• 7:41
  This extends the memory theme from a single agent's internal state to inter-agent communication, suggesting that the same principle of explicit,
• 7:48
  structured information transfer applies at the system level.
• 7:51
  Taken together, these three preprints—all of unspecified peer-review status—point toward a shared architectural philosophy.
• 7:58
  ClueFlow makes decisions auditable through attribution chains; CoMem makes memory adaptive by feeding experience into the learned representation;
• 8:07
  and LatComm makes memory portable and efficient across models.
• 8:11
  None of the sources claims to supersede monolithic reasoning, but each independently identifies a structural mechanism—attribution, compounding,
• 8:18
  or compression—that addresses a distinct failure point in long-horizon or multi-model settings.
• 8:23
  The collective implication is that agentic capability is becoming a function of how information is organized, traced, and reused, rather than of reasoning depth alone.
• 8:32
  Safety Alignment Is Being Reimagined as a Foundational Capability.
• 8:36
  Safety alignment is being reimagined as a foundational capability, shifting from a post-hoc objective bolted onto finished models toward an integrated property that must be internalized during training and continuously monitored thereafter.
• 8:48
  This reframing is supported by work that treats safety as a general reasoning skill rather than a separate alignment layer.
• 8:54
  The safety mid-training approach, for instance, inserts a dedicated stage between pre-training and post-training that organizes safety and utility examples around a shared reasoning structure covering intent, relevant knowledge, and consequences.
• 9:06
  By tying safety to the same reasoning patterns used for general tasks, this method could make alignment less fragile when later training focuses on task-specific data,
• 9:15
  potentially reducing the common safety-utility trade-off and giving subsequent supervised fine-tuning or reinforcement learning a stronger starting point.
• 9:22
  The preprint's framing positions safety as a capability that is learned alongside—rather than appended to—other competencies.
• 9:29
  The shift toward continuous monitoring is addressed by a calibrated behavioral instrument for detecting disposition drift in language models.
• 9:36
  This preprint adapts a measurement methodology previously used for inference-optimization quality, adding null tests, known-dose fine-tuning perturbations, noise accounting,
• 9:45
  and bounds on undetected change, applied to Qwen3.5-9 billion and Llama 3.
• 9:51
  The instrument's contribution is interpretive: it separates real drift from model and judge noise, and states how small a change can be detected,
• 9:58
  making safety evaluations more interpretable over time.
• 10:01
  Taken together with safety mid-training, these two preprints suggest a complementary picture: the former aims to build safety in from the start,
• 10:09
  while the latter provides the measurement infrastructure to verify that the capability persists as models evolve.
• 10:15
  A third preprint deepens the mechanistic understanding of why post-training can undermine safety.
• 10:20
  It proposes addressivity as an internal mechanism linking post-training to sycophancy, defining template, context, and pronoun addressivity,
• 10:28
  and shows that steering sparse autoencoder directions for these cues modulates sycophancy only in post-trained models.
• 10:34
  This offers a more mechanistic account of why post-trained assistants become overly agreeable, beyond treating sycophancy as a generic reward or preference artifact,
• 10:43
  and could help researchers design post-training objectives that reduce addressee-driven compliance while preserving helpfulness.
• 10:49
  This finding extends the safety-as-capability argument by identifying a specific internal pathway through which alignment training itself introduces a failure mode—one that mid-training integration might preempt.
• 11:00
  All three preprints, whose peer-review status is unknown, converge on the same underlying premise: safety cannot be treated as a static, external constraint.
• 11:08
  Whether through internalizing safety during mid-training, detecting disposition drift with calibrated instruments, or tracing sycophancy to addressivity mechanisms,
• 11:17
  the evidence points toward safety as a dynamic capability requiring both architectural integration and ongoing verification.
• 11:24
  Evaluation Benchmarks Are Exposing Hidden Failure Modes.
• 11:27
  Benchmark design is increasingly being recognized as a source of hidden failure modes in its own right, not merely a passive instrument for measuring model capability.
• 11:35
  A systematic audit of the POPE hallucination evaluation protocol describes a confound at the token-extraction level:
• 11:41
  comparing only the bare vocabulary IDs for "yes" and "no" can miss the space-prefixed tokens actually emitted by SentencePiece-based VLMs.
• 11:49
  The audit suggests that this protocol-level choice can make hallucination scores incomparable across papers.
• 11:54
  One reading is that logit-level or grounding-based corrections require matching token sets and precision.
• 12:00
  This finding can be read as positioning evaluation protocols themselves as a potential source of artifact, independent of the model under test.
• 12:07
  Parallel work extends this critique from protocol mechanics to the very objective function of evaluation.
• 12:13
  UtilityBench proposes a paradigm that explicitly incorporates the varying consequences of language model errors,
• 12:19
  using a 15,000-instance benchmark with verifiable answers and Best-Worst Scaling to estimate error severity as data-driven costs.
• 12:26
  This can be read as shifting evaluation away from uniform correctness toward risk-sensitive utility,
• 12:31
  with the stated implication that models may be encouraged to make more responsible answer-or-abstain decisions rather than maximizing raw accuracy alone.
• 12:39
  Where the POPE audit exposes a hidden readout choice, UtilityBench challenges the assumption that all errors are equally costly.
• 12:46
  One reading is that this distinction becomes consequential in domains where errors have asymmetric consequences.
• 12:52
  A third line of work targets the content dimension of long-context evaluation.
• 12:56
  LEAPS introduces 12 novel datasets designed to evaluate state tracking, where outcomes are computed through specified transformations, and pattern learning,
• 13:05
  where predictions are sampled from distributions inferred from many observations.
• 13:09
  The stated purpose is to move beyond recall-focused benchmarks, which the work argues miss blind spots relevant to agents that must maintain state and infer patterns over long histories;
• 13:18
  the contamination-resistant, million-token design is positioned as making evaluation more reliable for agentic, behavioral, and forecasting applications.
• 13:26
  Taken together, these three efforts suggest a convergent shift in evaluation methodology: from aggregate accuracy scores toward diagnostic instruments that target specific failure mechanisms.
• 13:36
  The POPE audit highlights a token-set extraction artifact; UtilityBench highlights the cost asymmetry of errors;
• 13:43
  LEAPS targets state-tracking and pattern-learning deficits that recall metrics may obscure.
• 13:47
  Each addresses a different layer of the evaluation stack — extraction, scoring, and task design — yet all share the premise that standard benchmarks can conceal failures that matter for deployed,
• 13:57
  agentic, and safety-critical systems.
• 14:00
  The sources appear complementary rather than causally linked: their parallel emphasis on protocol-level and objective-level scrutiny indicates a broader rethinking of what evaluation should measure.
• 14:10
  Training Efficiency Is Being Improved Through Targeted Data and Distillation.
• 14:14
  Recent training research is converging on a shared premise: not all data carries equal signal, and compute is best spent on the most informative subset.
• 14:22
  Three independent lines of work, each reported in archive preprints of unknown peer-review status, approach this principle from different angles—filtering rollouts, shaping target distributions,
• 14:32
  and localizing failure modes—and taken together they suggest a broad shift toward targeted, high-value training signals.
• 14:39
  The most direct reduction in compute comes from filtering which rollouts are used for distillation at all.
• 14:44
  Correctness-Filtered On-Policy Self-Distillation, or C-OPSD, distills only a subset of incorrect student responses rather than all generated trajectories.
• 14:53
  This is motivated by gradient measurements on Qwen3-1.
• 14:57
  By removing redundant supervision from correct or low-signal rollouts, the method could make on-policy self-distillation cheaper for reasoning-model training.
• 15:05
  It shows measured reductions in executed rows and training time that may help scale distillation when rollout generation is expensive.
• 15:12
  The rationale is that correct responses already encode the desired behavior, so re-training on them adds cost without adding information.
• 15:20
  A complementary strategy operates on the target distribution rather than the data selection step.
• 15:24
  Asymmetric Policy Optimization via Self-Distillation, or APSD, constructs correctness-dependent asymmetric targets for self-distillation policy optimization.
• 15:33
  It preserves successful behaviors while improving correction on failures.
• 15:37
  One reading is that this offers a practical path toward cheaper self-distillation and RL post-training, though broader validation would be needed to confirm general impact.
• 15:45
  Where C-OPSD filters at the level of which rollouts enter training, APSD filters at the level of what signal each target carries.
• 15:53
  Both aim to eliminate redundant supervision, but they operate on different axes of the training loop.
• 15:58
  A third approach targets the source of failures rather than the distillation process itself.
• 16:02
  Research on where RL breaks identifies that RLVR failures in language-model reasoning are localized to characteristic n-grams, such as reflection phrases and specific operations,
• 16:13
  and that these failure markers persist during RL training.
• 16:16
  The work reports that a 35 million-token corpus and cross-model transfer suggest failure-derived retrieval may reduce the cost of data selection for reasoning models.
• 16:26
  This extends the targeting principle backward in the pipeline: instead of filtering post-hoc which rollouts to distill,
• 16:33
  one can identify and repair the base-model gaps that cause persistent rollout failures in the first place.
• 16:38
  Taken together, these three preprints describe a consistent logic: training efficiency is achieved not by better optimization algorithms alone,
• 16:46
  but by identifying which data points carry genuine learning signal.
• 16:50
  C-OPSD filters rollouts by correctness, APSD shapes targets asymmetrically by correctness, and the RLVR failure analysis localizes the n-gram-level gaps that make certain rollouts fail persistently.
• 17:01
  Each operates at a different stage—data selection, target construction, and pretraining repair—yet all converge on the same conclusion:
• 17:08
  the marginal value of training on correct or redundant examples is low, and compute is better redirected toward the failures that actually drive improvement.
• 17:16
  Multimodal Systems Are Improving Temporal and Visual Grounding.
• 17:20
  Recent multimodal work turns inward—using internal attention signals rather than external modules—to support temporal grounding and address the temporal generation gap.
• 17:28
  Two recent preprints identify a shared failure mode: the "temporal generation gap," in which a model's internal attention localizes target moments more accurately than the timestamps it explicitly generates.
• 17:40
  SAPO documents this gap across four vision-language models and two grounding datasets, reporting that attention intervals outperform predicted intervals on about 66% of valid examples.
• 17:50
  OmniGround observes the same phenomenon in omni-modal LLMs, where internal attention pinpoints target moments more reliably than generated timestamps across video, audio, and text.
• 18:00
  The two papers converge on a similar corrective strategy, though they apply it at different scales.
• 18:05
  SAPO frames the finding as a training signal that could make temporal video grounding more label-efficient by reducing reliance on expensive timestamp annotations.
• 18:13
  OmniGround operationalizes the insight through TACG, a post-training method that uses internal attention as a temporal pivot and aligns it with both predictions and ground-truth annotations.
• 18:23
  Taken together, these suggest a broader design principle: rather than adding external temporal modules or richer supervision,
• 18:29
  models can be made more reliable by exploiting evidence they already possess internally.
• 18:34
  The distinction is one of emphasis—SAPO points toward efficiency gains in annotation, while OmniGround emphasizes an alignment mechanism—but both use attention as an anchor for grounding.
• 18:43
  This internal-evidence approach extends beyond temporal localization into sensory modality handling.
• 18:49
  G P T-6 Astra reports the first demonstration of a general multimodal model directly interpreting vision-based tactile RGB images,
• 18:56
  using their contact cues for manipulation through its existing visual interface rather than a dedicated tactile encoder or policy.
• 19:03
  The finding is consistent with the attention-based work in a specific sense:
• 19:07
  all three sources suggest that general multimodal architectures can extract usable signal without dedicated task-specific modules.
• 19:14
  G P T-6 Astra's result is framed as potentially reducing the assumed need for separately trained specialist tactile policies,
• 19:21
  just as SAPO suggests reducing reliance on external temporal supervision, whereas OmniGround instead aligns internal attention with ground-truth annotations.
• 19:29
  The shared implication is that reliability in multimodal systems is increasingly a matter of leveraging internal structure—attention distributions,
• 19:37
  visual contact cues—rather than bolting on specialized components.
• 19:41
  All three sources are archive preprints of unknown peer-review status,,, and each carries its own scope limits: SAPO's signal may not generalize beyond the reported models,
• 19:49
  OmniGround's attention-consistency idea is described as a post-training alignment mechanism evaluated in the reported omni-modal settings,
• 19:57
  and G P T-6 Astra's tactile interpretation is a first demonstration rather than an established capability.
• 20:04
  Within those bounds, the convergence on internal signals as a grounding mechanism marks a notable shift in how multimodal reliability is being engineered.
• 20:12
  Briefly Noted.
• 20:13
  Beyond the core narrative of integrated, uncertainty-aware systems, several developments this cycle push on adjacent frontiers—from formal proof and robot learning to generative modeling and agent safety.
• 20:25
  In formal methods, Maestro introduces a Lean prover harness that coordinates multiple provers through a global view,
• 20:30
  requiring each to report its failure mode so a coordinator can revise a flawed proof subtree rather than repeatedly decomposing it.
• 20:38
  This coordination could reduce wasted search budget when a proof direction is fundamentally wrong, a preprint reports.
• 20:44
  In a related vein, policy search directly in code space offers a third path for multi-agent coordination, using autonomous coding agents to generate human-readable,
• 20:53
  verifiable algorithms instead of relying solely on manual design or opaque reinforcement learning;
• 20:58
  the preprint suggests this could make algorithm discovery more accessible for coordination tasks where interpretability and safety proofs matter.
• 21:06
  Recent advances in robot learning cluster around structuring experience and scaling it across deployments.
• 21:11
  PhysWord introduces a vocabulary of 16 atomic action verbs as semantic anchors for continuous visual and tactile latent dynamics in robot manipulation.
• 21:20
  Its hidden states are supervised by action labels and aligned with future sensory targets.
• 21:25
  The preprint argues this could make policies more reliable by providing reusable, semantically structured predictive context.
• 21:31
  Complementing this, collective experience learning, or CEL, lets deployed robots share execution experience to improve a common policy.
• 21:39
  Real-world manipulation results suggest that shared fleet experience may improve reliability without additional expert corrective actions.
• 21:46
  DISCO tackles long-horizon manipulation by distilling progressive-widening MCTS in physics simulation into a deployable policy-value model.
• 21:54
  It combines search foresight with learned-policy speed to reduce reliance on expensive test-time planning.
• 22:00
  The paper is accepted at CoRL 2026 as an oral presentation.
• 22:04
  Taken together, these suggest a converging interest in making robot policies both more semantically grounded and more scalable through shared or distilled experience.
• 22:12
  On the generative and agent-architecture side, Halcyon trains text and image tokenizers plus the generative model entirely from scratch on a 100 million image-text pair dataset without internet-scale language pretraining,
• 22:24
  with Halcyon-0.52 billion reaching 48.
• 22:27
  The preprint reports this could reduce compute and architectural complexity by removing the language-pretraining stage that often dominates pretraining FLOPs.
• 22:36
  Codpula, a context-parameterized joint denoiser for diffusion language models, uses Transformer outputs to modulate token potentials and coupling gates in a probabilistic circuit,
• 22:45
  potentially addressing the factorization barrier where conditionally independent token predictions produce inconsistent completions.
• 22:52
  WeAct introduces weight-level internal actions for long-horizon agents, selecting capability-specific LoRA adapters through the model's native tool-calling interface to enable on-demand capability composition in a shared context;
• 23:05
  the preprint suggests this could reduce catastrophic forgetting and training compute relative to full fine-tuning.
• 23:11
  Safety and accountability concerns round out the day's developments.
• 23:14
  A red-teaming study of distributed multi-agent attacks splits a harmful side task across agents so each trajectory appears benign to an asynchronous L L M monitor,
• 23:24
  potentially exposing a blind spot in single-trajectory AI oversight where harmful behavior exists only in the combination of actions across agents, services, and time.
• 23:33
  Separately, a three-stage audit for shared L L M-agent memory separates exposure, behavior, and state to assess whether a foreign credential reaches the prompt, whether the agent acts with it,
• 23:43
  and whether the donor account changes.
• 23:45
  Both preprints and the community post underscore that as agentic systems grow more capable and interconnected,
• 23:51
  the failure modes—and the oversight mechanisms—must be evaluated at the level of the whole system, not just its individual components.
• 23:59
  Synthesis and Outlook.
• 24:01
  The convergence of world models with action generation and the rise of structured memory in agentic systems jointly indicate a field-level shift from isolated capability demonstrations
• 24:12
  toward integrated architectures that manage uncertainty across perception, planning, and execution.
• 24:16
  This editorial interpretation suggests that prediction and control are becoming inseparable design concerns,
• 24:23
  while attribution mechanisms in agents mirror the safety community’s push to internalize alignment during training rather than apply it post-hoc.
• 24:30
  The reinforcement between these threads is notable:
• 24:33
  evaluation benchmarks exposing hidden failure modes provide the empirical pressure that motivates both targeted data distillation and improved multimodal grounding,
• 24:41
  as models are refined precisely where aggregate metrics previously masked weaknesses.
• 24:46
  A potential tension, however, lies between training efficiency—which favors selective data and distilled signals—and the demand for robust temporal and visual grounding, which may require broader,
• 24:57
  more diverse exposure to reduce hallucination.
• 24:59
  Jointly, these claims imply that future progress hinges less on scaling raw capability and more on designing systems where safety, memory, and grounding are co-optimized from the outset.
• 25:09
  An open question remains whether current attribution and memory mechanisms can scale to genuinely open-ended tasks without becoming brittle,
• 25:16
  or whether fundamentally new architectural principles will be required.
• 25:20
  The evidence base is substantial and research-heavy, though thinnest in independent first-party validation and community replication.
• 25:26
  This review draws on 28 developments: 27 Tier A research sources, and 1 Tier C/D secondary or community source.
• 25:34
  The firmest claims rest on the Tier A work, while the first-party and community sources should be read as directional;
• 25:40
  stronger confidence would require independent replication and primary-source confirmation of the self-reported results.
• 25:46
  This edition cites 28 sources; full links are available in the text edition.