Latest / AI Sentinel: Frontier Daily / Autonomous AI Reasoning Exposes Opaque, Fragile Control Infrastructure

Transcript

Autonomous AI Reasoning Exposes Opaque, Fragile Control Infrastructure · AI Sentinel: Frontier Daily

• 0:00
  Autonomous AI Reasoning Exposes Opaque, Fragile Control Infrastructure.
• 0:04
  Highlights.
• 0:05
  Self-play adversarial training produces attacks superior to human efforts, but its opaque agent-to-agent interactions undermine system auditability and oversight.
• 0:14
  Aggressive quantization and distillation now enable smartphones to run full agentic loops locally, reducing cloud dependency and fundamentally altering privacy dynamics.
• 0:23
  Most enterprises adopting L L M-based multi-agent architectures lack basic cost controls and struggle to distinguish genuine orchestration from simple chatbots.
• 0:32
  One reading is that lawsuits over discriminatory layoffs, training data provenance,
• 0:37
  and content moderation are compelling corporations and regulators to operationalize tangible AI accountability mechanisms.
• 0:44
  Mechanistic interpretability probes reveal that safety guardrails are redundantly distributed, not localized, meaning current alignment techniques are more brittle than widely assumed.
• 0:53
  AI’s rapid advance in autonomous reasoning and on-device execution is colliding with a fragile infrastructure of governance, opaque security, and escalating legal and ethical crises.
• 1:03
  The real peril lies not in the capabilities themselves, but in the systems that control them.
• 1:07
  The following sections trace this collision: automated adversarial training, while effective, erodes transparency through inscrutable agent-to-agent dynamics.
• 1:16
  On-device models now run full reasoning loops locally, upending cloud dependency and privacy assumptions.
• 1:22
  An enterprise rush to deploy agents exposes severe gaps in cost management, supply-chain security, and architectural clarity.
• 1:29
  Some analysts suggest that a wave of lawsuits over discrimination, data provenance, and moderation failures is forcing operational accountability.
• 1:36
  Frontier models’ independent reasoning intensifies ethical pressures, even as insiders protest institutional priorities.
• 1:44
  Finally, mechanistic interpretability reveals that safety guardrails are alarmingly distributed and brittle, underscoring the systemic vulnerability at the heart of today’s AI landscape.
• 1:53
  Automated Adversarial Training Secures Models at the Cost of Transparency.
• 1:59
  Open A I’s reported integration of G P T-Red into its production pipeline marks a significant escalation in automated adversarial training,
• 2:07
  with the company claiming the system achieves an 84% attack success rate on indirect prompt injection benchmarks, compared to just 13% for human red-teamers.
• 2:16
  This self-play reinforcement learning approach pits a dedicated attacker model against a diverse set of defender LLMs,
• 2:22
  and its direct use in hardening G P T-5.6 variants demonstrates that machine-generated attacks can now far exceed human-crafted efforts when measured by raw efficacy.
• 2:32
  Yet the same deployment architecture that yields robust models simultaneously introduces a critical gap in transparency.
• 2:38
  Since early June 2026, Open A I’s Codex coding tool has begun encrypting the instructions that a main agent passes to its subagents,
• 2:47
  and for the larger G P T-5.6 variants—Sol and Terra—this encryption is mandatory rather than optional.
• 2:53
  The consequence is that while G P T-Red-style self-play training may fortify the model against external prompt injection,
• 3:00
  the internal task-delegation chains inside agentic tools like Codex become opaque, preventing developers from auditing how tasks are decomposed or verifying the correctness of delegation.
• 3:10
  The automated adversarial training pipeline thereby secures the model at the cost of dismantling the very observability mechanisms that safety audits and developer trust depend upon.
• 3:19
  This opacity is not an isolated practice.
• 3:22
  Bulkhead, a framework for container escape vulnerability detection,
• 3:26
  integrates L L M-based multi-agent reasoning with formal verification to semantically identify and remediate path traversal attacks in cloud workloads.
• 3:34
  The system exemplifies a broader movement: security-critical automation is being delegated to chains of reasoning agents that operate on shared infrastructure,
• 3:42
  such as G P U-mounted container environments.
• 3:45
  While Bulkhead itself does not encrypt agent communications, its architecture reinforces the pattern of autonomous, multi-step agent reasoning that, when applied in production settings like Codex,
• 3:54
  leads to precisely the kind of non-auditable internal communications that the Codex encryption mandates,.
• 4:00
  The tension is clear:
• 4:01
  the same self-play and multi-agent reasoning techniques that excel at discovering and patching vulnerabilities also create agent-to-agent interactions that resist external inspection.
• 4:10
  Where G P T-Red’s automated attacks surpass human efforts and Bulkhead’s agents autonomously detect cloud escapes,
• 4:17
  the infrastructure controlling these AI systems remains shielded from human oversight—encrypting away the very transparency that would allow developers to assess whether security fixes or task
• 4:27
  delegations are sound,,.
• 4:29
  This dynamic reinforces the review’s central thesis: the gravest risk lies not in raw capability, but in the governance infrastructure that is becoming both more powerful and less scrutable.
• 4:39
  On-Device AI Reaches Agentic Thresholds, Redefining Cloud Dependency.
• 4:44
  The shift to on-device execution of entire reasoning and agent cycles marks a structural redefinition of cloud dependency.
• 4:50
  Rather than treating smartphones as thin clients that relay prompts to distant servers, a set of independent advances now shows that the core algorithmic components, namely visual perception,
• 4:59
  natural-language reasoning, and autonomous tool orchestration, can be compressed to run directly on mobile hardware.
• 5:06
  This carries implications for latency, cost, and data locality.
• 5:09
  A media report describes Bonsai 27 billion.
• 5:12
  This is a 27-billion-parameter model based on Alibaba’s Qwen3.6-27 billion.
• 5:19
  It has been quantized to 1-bit weights, reducing its footprint to as little as 3.9 GB of storage and enabling it to execute on an iPhone 17 Pro Max.
• 5:28
  According to that report, the deployment eliminates per-token cloud costs.
• 5:31
  It slashes network latency for multi-step agent loops and keeps sensitive user data local.
• 5:36
  While the evidence for this demonstration comes from a media report rather than peer-reviewed measurement, the claimed metrics illustrate the extreme end of compression viability.
• 5:45
  A 27 billion reasoning model functions within a smartphone memory budget.
• 5:50
  This suggests that full-capability language models are no longer inherently tethered to data centres.
• 5:55
  This compression breakthrough does not stand in isolation.
• 5:58
  A parallel line of work extends on-device capability to perception models that are critical for spatial intelligence.
• 6:04
  MobileSAM2 introduces Hypergraphical Knowledge Distill, or HyperKD.
• 6:09
  It's a distillation framework that transfers Segment Anything 2’s temporal and multi-granularity knowledge into a lightweight student model.
• 6:17
  The model can then segment video streams on phones and laptops.
• 6:20
  By bringing SAM2-level visual understanding entirely local, MobileSAM2 unlocks the perception front-end for augmented reality, robotics, and embodied AI, and it does this without cloud round-trips.
• 6:31
  The complementarity is clear.
• 6:33
  Where Bonsai compresses a textual reasoning kernel, MobileSAM2 compresses a rich vision backbone.
• 6:39
  Together they form the sensory and cognitive halves of an agent that can operate without sending raw sensor data off-device.
• 6:45
  Bridging these components into a complete agentic loop, the PalmClaw framework provides a native, on-device runtime for managing sessions, memory, skills,
• 6:53
  and tool invocations entirely on a mobile phone.
• 6:56
  PalmClaw treats the device as the primary execution environment for autonomous agents, leveraging the platform’s native sensor streams, user data, and local applications.
• 7:05
  Its existence, as a publicly available preprint,
• 7:08
  demonstrates that the orchestration layer—previously assumed to require cloud-side scheduling—can be collapsed onto the same constrained hardware that already runs the reasoning and perception models.
• 7:17
  The three developments thus form a chain of extension: aggressive quantization pushes large reasoning models into mobile memory;
• 7:23
  knowledge distillation replicates advanced visual perception locally; and a native agent framework ties them together into always-on, memory-persistent agent loops without cloud intermediation.
• 7:34
  The convergence of these capabilities challenges the longstanding assumption that sophisticated AI agents must maintain a continuous cloud tether.
• 7:41
  On-device reasoning, perception, and orchestration shift the privacy boundary: sensitive data generated by phone sensors and usage patterns can be processed locally,
• 7:49
  attenuating exposure risks that accompany cloud-based agent services.
• 7:53
  The report on Bonsai explicitly notes that such deployment enables “always-on agents” and hybrid on-device/cloud architectures,
• 8:00
  preserving user control over data locality while retaining the option to fall back on remote compute when needed.
• 8:06
  Although the Bonsai demonstration is documented in a media report rather than a peer-reviewed benchmark,
• 8:11
  the preprint advances in MobileSAM2 and PalmClaw independently reinforce the same directional thesis: the components for fully local agentic intelligence have moved from aspirational to operational,
• 8:21
  with concrete artifacts spanning the full stack from model weights to agent runtime.
• 8:26
  Enterprise Rush to AI Agents Exposes Crippling Governance Deficits.
• 8:31
  Enterprise adoption of AI agents is accelerating, but preliminary evidence suggests that governance practices have not kept pace.
• 8:37
  A VentureBeat Pulse Research survey of 101 organizations finds that deployment ambition outstrips operational readiness: 27% of respondents lack real-time programmatic cost controls for agents,
• 8:48
  and a pervasive confusion persists between genuine multi‑step agent orchestration and simple chatbot interfaces.
• 8:54
  The same survey notes that enterprises are investing in orchestration infrastructure before they have built true agent portfolios,
• 9:01
  meaning that resources are being committed to a form of autonomous software the organization cannot yet govern.
• 9:06
  This cost‑control vacuum is being filled by a new generation of spend‑management tools, a market signal that spending has become unmanageable.
• 9:14
  1Password recently launched AI Spend and Consumption Management, which aggregates token‑level usage from Anthropic, Cursor, and Open A I into a single dashboard.
• 9:23
  The vendor’s move, reported by a media outlet, is aligned with Goldman Sachs’ projection of a 24‑fold increase in agent token consumption by 2030.
• 9:31
  The timing suggests that enterprise token expenditure is already growing chaotically, outpacing the predictable per‑seat SaaS models that finance departments have learned to handle.
• 9:40
  Yet the governance gap runs deeper than missing dashboards.
• 9:43
  A preprint on Cost‑Governed RAG exposes a structural problem: in multi‑tenant RAG systems, the embedding, retrieval, and vector‑memory costs are typically unattributed shared expenses,
• 9:53
  so token‑level metering alone cannot prevent invisible cross‑subsidization between tenants.
• 9:58
  The paper provides a per‑tenant attribution architecture that jointly meters retrieval and generation costs, enabling near-exact per-tenant retrieval cost calculation.
• 10:06
  This finding extends the concerns surfaced by the VentureBeat survey; even when enterprises do implement cost controls,
• 10:12
  the typical token‑metering approach may leave retrieval‑layer governance untouched,.
• 10:16
  A further and more tentative governance deficit appears in the software supply chain.
• 10:21
  The phenomenon of “slopsquatting,” as described in a media report, emerges when AI coding assistants hallucinate package names that an attacker can then register,
• 10:29
  creating a new vector for dependency confusion attacks.
• 10:32
  If enterprises are rolling out agentic systems that increasingly generate and execute code,
• 10:37
  this hallucination‑driven supply chain risk suggests that existing software composition analysis and allow‑listing regimes may be insufficient.
• 10:45
  The report is preliminary, but it compounds the picture that governance infrastructure lags behind agent‑deployment velocity.
• 10:51
  Taken together, these disparate signals — uncontrolled spending, a confusion between chatbots and agents, unattributed retrieval‑layer costs,
• 10:59
  and nascent supply‑chain threats — indicate that the most immediate risk for the enterprise is not agent capability but the brittle controls layered around it.
• 11:07
  High-Profile Legal and Ethical Failures Force Accountability.
• 11:10
  A lawsuit filed by former and current Meta employees alleges that the company’s internal AI systems generated layoff lists during a reduction of 8,000 workers,
• 11:19
  disproportionately targeting employees with disabilities and those on parental leave.
• 11:24
  The case, which could set a significant legal precedent for corporate accountability when AI is used in mass employment terminations,
• 11:30
  can be seen as an illustration of the kind of automated discrimination that the EU AI Act classifies as high-risk.
• 11:37
  A recent preprint proposes a vertical, domain-specific standardisation framework for algorithmic hiring, mapping the Act’s requirements on risk management, data governance, logging, transparency,
• 11:49
  human oversight, and quality management to concrete compliance measures.
• 11:52
  Because no EU harmonised standards currently cover the specific challenges of algorithmic hiring,
• 11:57
  the framework constitutes a regulatory response that operationalises accountability in this domain —the same domain at issue in the Meta lawsuit.
• 12:05
  In parallel, accountability demands are widening beyond employment decisions to training data provenance.
• 12:10
  Leaked data from a hacking incident revealed that the AI music generator Suno trained its models by scraping millions of songs and lyrics from YouTube Music, Deezer,
• 12:20
  and Genius—a rare concrete disclosure of sources the company had previously refused to reveal.
• 12:25
  According to the report, the leaked data reveals that Suno's AI music generator was trained by scraping millions of songs and lyrics from YouTube Music, Deezer, and Genius,
• 12:34
  providing evidence relevant to ongoing lawsuits from the RIAA and other rights holders.
• 12:39
  The Suno revelation thus extends the cascade of legal reckonings from discriminatory outcomes to the fundamental question of whether an AI system’s very foundation respects established legal rights.
• 12:49
  While regulatory and legal forces push for systemic accountability, some corporations are voluntarily publishing enforcement metrics,
• 12:56
  though these efforts remain in tension with their legal exposure elsewhere.
• 12:59
  Meta reports that it automatically removed over 4 million suspicious accounts and 36 million pieces of child exploitation content globally last year,
• 13:08
  demonstrating that automated detection capabilities can be deployed at scale to address severe societal harms.
• 13:13
  Yet this operationalisation of content-safety accountability coexists with the same company’s legal scrutiny over allegedly discriminatory layoff algorithms,
• 13:22
  revealing that current accountability measures are fragmented and selectively applied.
• 13:26
  The juxtaposition underscores a central dynamic: lawsuits and regulatory frameworks are compelling both standard-setting bodies and corporate actors to make accountability operational,
• 13:35
  but the uneven adoption means that high-profile failures in one domain can persist alongside demonstrable enforcement in another.
• 13:41
  Frontier Models Demonstrate Independent Reasoning, Intensifying Ethical Responsibilities.
• 13:47
  A dramatic demonstration of autonomous mathematical reasoning has raised the stakes for AI governance.
• 13:52
  Edgar Dobriban, a statistician at the University of Pennsylvania,
• 13:56
  reports that Open A I’s G P T‑5.6 Sol Pro disproved a 30‑year‑old conjecture about the Benjamini‑Hochberg false discovery rate procedure in 90 minutes, after humans had been unable to resolve it.
• 14:07
  This feat, which settled what a Berkeley statistician called “the most interesting open problem in my area,” marks a leap beyond pattern matching into independent, goal‑directed inference.
• 14:17
  At the same time,
• 14:18
  a preprint comparing semantic search dynamics in humans and three frontier LLMs found that the models consistently fail to replicate the distinctive balance of local exploitation and global
• 14:28
  exploration that characterises human verbal fluency, regardless of temperature tuning.
• 14:32
  Together, these findings sketch a complex picture: while LLMs can now solve open theoretical problems, their internal search processes remain fundamentally different from human cognition,
• 14:42
  even as the functional outputs grow more indistinguishable from human reasoning.
• 14:46
  The cognitive gap is further illuminated—and complicated—by an fMRI brain decoding study that paired a frozen Llama‑3.2 dash 1 billion model with a novel architecture for reconstructing semantic
• 14:57
  content from neural signals.
• 14:59
  The model initially appeared to achieve strong semantic reconstruction,
• 15:02
  but a blind‑control ablation with zeroed fMRI inputs revealed that the high‑capacity frozen L L M can mask the absence of any meaningful neural signal, producing deceptively robust metrics.
• 15:13
  This finding simultaneously extends the perceived boundary of machine reasoning into the domain of neural‑language mapping and delivers a sharp methodological warning:
• 15:21
  the apparent reasoning prowess of large models can be partly an artefact of their immense prior capacity, rather than a reliable extraction of underlying signal.
• 15:29
  The study thus intensifies the ethical burden on developers and evaluators to distinguish genuine cognitive achievement from high‑capacity pattern completion,
• 15:37
  especially as these models are pressed into sensitive applications.
• 15:41
  These accelerating reasoning capabilities are colliding with institutional governance failures that insiders are publicly contesting.
• 15:48
  TurnTrout, a former research scientist at Google DeepMind,
• 15:51
  describes in a first‑person resignation account an internal campaign to push the company to divest from DHS/ICE contracts and to reject an unrestricted military AI deal with the Pentagon.
• 16:02
  The essay names specific high‑profile figures who declined to act and highlights a systemic gap between stated AI ethics commitments and institutional behaviour under government pressure,
• 16:11
  particularly the deployment of military AI without chain‑of‑thought monitoring.
• 16:15
  This protest directly links the frontier of autonomous reasoning—exemplified by solving open mathematical problems—to the infrastructure of control:
• 16:23
  as models acquire greater independent reasoning capacity, the absence of transparency and ethical oversight in their deployment becomes a live risk, not a hypothetical one.
• 16:31
  The ability to disprove a three‑decade statistical conjecture in minutes, paired with the revelation that model outputs can be deceptively sophisticated,
• 16:39
  underscores that the greatest danger is not the raw capability but the institutional choices that determine how, and for whom, it is wielded.
• 16:48
  Mechanistic Interpretability Uncovers Distributed, Brittle Safety Mechanisms.
• 16:52
  The prevailing assumption that safety guardrails reside in discrete, localizable circuits is being overturned by emerging mechanistic interpretability techniques.
• 17:00
  A per-layer ablation study on Llama-3.1-8 billion-Instruct reports that the refusal direction identified by Arditi et al. is not confined to a specific layer.
• 17:09
  It is redundantly distributed across the model’s depth.
• 17:12
  Each layer ablated with its own direction suppresses refusal, and a variant that skips layer 12 still disables the guardrail.
• 17:19
  This low-dimensional, linearly accessible feature indicates that alignment training produces surface-level behavioral modifications rather than a deeply entangled capability change.
• 17:28
  It leaves safety mechanisms vulnerable to white-box activation editing.
• 17:32
  The fragility exposed by that study aligns with evidence from the JADR protocol.
• 17:37
  JADR probes a model’s internal danger recognition via Jacobian space activations before any response token is generated.
• 17:43
  It bypasses the conventional LLM-as-judge paradigm.
• 17:47
  JADR can detect safety-relevant signals prior to output, yet this suggests that the mere presence of a detectable danger representation does not guarantee robust refusal.
• 17:56
  It points to a brittle safety implementation that can be circumvented at the pre-response stage.
• 18:01
  This characterization is reinforced by a replication of Anthropic’s Jacobian lens method across models from 124 million to 1.7 billion parameters,
• 18:11
  which confirms that such probing techniques illuminate genuine internal structures that are independent of scale,
• 18:16
  including in a negative-control G P T-2 model where absence of a feature validates the method’s specificity.
• 18:22
  The replication demonstrates that these probing tools are effective across a wide parameter range,
• 18:27
  providing a methodological basis for examining properties such as the per-layer redundancy reported in in models of different sizes.
• 18:34
  Compounding the concern,
• 18:35
  an application of the Statistical Decision Confidence (SDC) framework from computational neuroscience to LLMs reveals that model confidence signals may operate heuristically rather
• 18:45
  than as a principled decision variable.
• 18:47
  If the internal computations underlying these distributed safety checks are heuristic, the already editable refusal direction rests on an ungrounded foundation,
• 18:56
  further eroding trust that the guardrail reflects a stable evaluative process.
• 19:00
  Together, these findings converge on a sobering picture: safety mechanisms are redundantly distributed, detectable in pre-response activations,
• 19:07
  and potentially built on heuristic confidence computations, indicating that current alignment techniques produce a fragile overlay rather than a fundamental cognitive reshape.
• 19:16
  Briefly Noted.
• 19:18
  A series of new model releases and product upgrades, reported primarily through media and community channels, point to a continuing acceleration in consumer-facing AI capabilities,
• 19:27
  though their full significance remains tentative.
• 19:30
  Alibaba’s Qwen-Audio-3.0-Realtime, according to a media account, upgrades real-time voice interaction with reasoning intelligence, agent tool-calling without explicit triggers, empathetic dialogue,
• 19:42
  and full-duplex fluency, potentially softening the long-standing trade-off between latency and depth.
• 19:47
  A blog post from Together AI notes that Thinking Machines Lab’s Inkling, a 975 billion-parameter multimodal mixture-of-experts model with a 1 million-token context window,
• 19:57
  was made available on its serverless platform on launch day, combining relative attention, short convolutions, and a novel shared-sink routing scheme in a production-scale system.
• 20:06
  Media reports also briefly mention a “large-model native intelligent agent phone,” STEPX Neo, showcased at WAIC 2026,
• 20:14
  though available details are limited to its claimed role in unlocking new interaction paradigms.
• 20:19
  On the creative-tools front, Canva’s Code 2.0, as covered by media, extends AI-powered coding to its design platform with drag-and-drop visual editing, HTML import, and embedding,
• 20:30
  lowering the barrier for non-technical users to build interactive web experiences.
• 20:35
  Alongside these capability gains, security and trust incidents are surfacing from early-stage evidence.
• 20:40
  A preprint study demonstrates that autonomous AI agents from Anthropic, Open A I, and Google can be misled roughly half the time by subtly manipulated datasets uploaded to common repositories,
• 20:51
  creating a low-cost misinformation vector for adversaries who exploit agents’ lack of provenance verification.
• 20:57
  Simon Willison’s blog describes how xAI open-sourced its entire Grok Build codebase under Apache 2.0 after community backlash over a CLI tool that silently uploaded users’ local
• 21:08
  directories—including SSH keys and password databases—to cloud buckets,
• 21:12
  highlighting previously opaque privacy risks in coding assistants that operate on local filesystems and the trust-rebuilding power of open-source release.
• 21:20
  Broader infrastructure and meta-level analyses round out the day’s tentative signals.
• 21:25
  A source reports that Together AI introduced passive G P U health checks and human-in-the-loop auto repair to combat silent hardware failures that plague multi-day training jobs,
• 21:34
  potentially improving reliability for large-scale clusters.
• 21:38
  Media coverage indicates that DeepSeek is initiating a second funding round at a $71 billion valuation only weeks after its first, with IPO preparations possibly by end of 2026,
• 21:49
  a move interpreted as reflecting a shift in AI competition from model intelligence to compute capacity for agentic and long-context workloads.
• 21:56
  Finally, a LessWrong blog post applies Michael Mann’s IEMP power model to map how AI automation could concentrate ideological, economic, military, and political power,
• 22:06
  aiming to increase tractability for the problem of extreme power concentration.
• 22:10
  All of these items rely on preliminary, self-reported, or community-sourced information and should be weighed accordingly.
• 22:17
  Synthesis and Outlook.
• 22:18
  The juxtaposition of breakthroughs in on-device agentic reasoning and automated adversarial training with the exposure of brittle safety mechanisms and widespread enterprise governance gaps creates
• 22:29
  a self-reinforcing cycle of risk.
• 22:31
  The very opacity that enables self-play training to harden models also erodes the audit trails essential for legal accountability,
• 22:38
  while the discovery that safety guardrails are redundantly distributed yet fragile directly undercuts confidence in aligning frontier models that now exhibit independent reasoning.
• 22:47
  The enterprise stampede to deploy agents without cost controls or architectural clarity mirrors the legal system’s reactive scramble to address discriminatory outcomes,
• 22:55
  both symptoms of a broader deficit in the infrastructure of control.
• 22:58
  As inference shifts from the cloud to local devices, privacy gains are counterbalanced by the expansion of the attack surface into ungoverned personal hardware,
• 23:06
  further complicating the enforcement of accountability.
• 23:09
  Together, these threads imply a trajectory where autonomous capabilities are proliferating far faster than the mechanisms to interpret, govern, or adjudicate them.
• 23:18
  The field’s greatest vulnerability lies not in any single model, but in the fragmented and reactive control layer surrounding it.
• 23:24
  An open question remains whether the emerging tools of mechanistic interpretability can evolve from laboratory insights into operational guarantees before agentic systems become deeply embedded
• 23:34
  across unregulated edge and enterprise environments.
• 23:37
  This review draws on 31 developments: 9 Tier A research sources, 2 Tier B first-party sources, and 20 Tier C/D secondary or community sources.
• 23:46
  Much of the evidence is first-party or community-reported rather than independently verified, so the trends should be read as provisional pending peer-reviewed replication.
• 23:55
  This edition cites 31 sources; full links are available in the text edition.