Amaze Networks Morning Briefing

---

1. GPT-6 Launches Today — Dual-Tier Reasoning, 2M Context, 87% Agent Task Completion

TL;DR: OpenAI released GPT-6 on April 14 with a 2M-token context window, a dual-tier reasoning architecture that applies internal verification to outputs, and reported agent task completion of 87% — up from 62% on GPT-5.4. That last number is the one that matters for anyone building production agentic workflows.

Key Points:

• 2M token context (double GPT-5.4); 40%+ benchmark improvement across standard evaluations
• HumanEval >95%, MATH ~85%, agent task completion 87% vs. 62% on prior version
• Claimed hallucination rate below 0.1% (unverified by independent third parties yet)
• Architecture: explicit dual-tier reasoning — fast System-1 path for generation, slower System-2 path for internal verification; inference-time compute applied at the model level, not just via chain-of-thought prompting
• Unified interface merges ChatGPT, Codex, and the Atlas browser agent into one endpoint
• Pricing held flat at $2.50/M input, $12/M output

Deep Dive:

The 87% agent task completion rate is the headline buried beneath the headline. GPT-5.4 at 62% was demo-grade — agents completed tasks often enough to be impressive in controlled conditions but broke often enough to be unreliable in production workflows. The gap between 62% and 87% is not a 40% improvement; it's the difference between "occasionally useful tool" and "deployable with appropriate guardrails." Whether that claim holds under real-world conditions outside OpenAI's benchmark suite is the practical question, but if it's even close to accurate, the agentic workload infrastructure buildout just got a much more urgent business case.

The dual-tier reasoning architecture is worth understanding architecturally, not just as a marketing claim. System-1/System-2 framing has been used loosely for years in AI discussions, but what OpenAI appears to have implemented is structural: generation and verification as separate inference passes with different compute budgets, unified behind a single API endpoint. This means token budgets now include verification overhead on some requests, which has direct cost and latency implications for high-volume API consumers. Infrastructure teams sizing API gateway throughput and per-session token limits should revisit those calculations before committing to GPT-6 at scale.

The 2M context window is the other infrastructure concern. Most teams have not stress-tested their API consumption patterns against 2M-token payloads. Token budget policies, per-session cost controls, and memory management in agent harnesses all need revisiting. Context windows that large don't just change what you can do — they change how you have to instrument and monitor what the model is actually doing inside those massive context loads.

So What? Before migrating production agentic workflows to GPT-6, build a benchmark suite against your specific task types — agent task completion benchmarks from OpenAI's test suite won't tell you how it performs on your particular automation workflows. Also: update your API cost models immediately. A 2M context window + dual-tier verification overhead changes the economics for high-frequency agent calls.

SourcesFazm Blog — https://fazm.ai/blog/new-llm-releases-april-2026

---

2. AWS Project Houdini: Factory-Built Datacenters, 15 Weeks Down to Two

TL;DR: Amazon's internal Project Houdini replaces traditional on-site data center construction with factory-prefabricated "skid" modules — semi-trailer-sized units arriving fully wired with racks, power distribution, cabling, lighting, and safety systems pre-installed. Server installation starts in two to three weeks instead of fifteen, eliminating up to 50,000 on-site electrician hours per project.

Key Points:

• Each skid weighs approximately 20,000 lbs, arrives site-ready with complete power distribution and cabling
• Eliminates up to 50,000 on-site electrician labor-hours per data center project
• Three U.S. manufacturing plants planned: Topeka, Houston, Salt Lake City
• Production readiness target: August 2026
• Scale target: more than 100 data center builds per year at full ramp
• This is an internal AWS initiative, not a commercial product offering

Deep Dive:

Project Houdini represents a genuine architectural shift in how hyperscalers think about capacity acquisition. The constraint has evolved: power and land are still the bottleneck categories that attract the most attention, but construction throughput — specifically the availability of skilled electricians and the time required to turn raw facility shell into operational data center — has become the binding constraint in markets where power and land are available. AWS is responding the same way it responded to every other scaling problem: remove the human bottleneck with a manufacturing process.

The factory model also forces standardization that field construction doesn't. When racks, power distribution, and cabling arrive as a module, the network architecture inside that module has to be decided at manufacturing time, not on-site. This upstream pull of infrastructure decisions has direct implications for how AI-era data center networking gets standardized. The flexibility to customize switch placement, cable runs, and power topology on-site disappears — replaced by whatever configuration gets baked into the skid spec. That may look like a constraint, but it's also a forcing function for the kind of standardized, automation-friendly infrastructure designs that make programmable management feasible at scale.

Google and Microsoft will accelerate equivalent internal programs. The AWS disclosure sets a public benchmark: if your hyperscaler competitor is targeting two-week server deployment at 100+ sites per year, matching that speed becomes a competitive requirement. Look for similar prefab programs to be disclosed or accelerated in the next two quarters.

So What? If you're planning major data center deployments and competing for construction crews in markets where AWS is building, expect labor availability to tighten further through late 2026 as hyperscaler manufacturing programs absorb capacity. Prefab modular is no longer a speed optimization — it's becoming structurally necessary for organizations that need predictable deployment timelines.

SourcesDataCenter Dynamics — https://www.datacenterdynamics.com/en/news/aws-launches-project-houdini-to-speed-up-data-center-construction-report/ | StartupNews.fyi, April 14, 2026

---

3. Closed-Loop AIOps Crosses the Sub-60-Second Recovery Threshold

TL;DR: AIOps deployments with Tier 1 automation — where the platform not only detects faults but executes pre-approved remediation autonomously — are achieving sub-60-second fault recovery in production networks. This is not research: these are production deployments using Cisco DNA Center, Juniper Apstra, and Aruba Central. The gap between "AI tells you something is broken" and "AI fixes it in under a minute" is where competitive differentiation lives right now.

Key Points:

• Sub-60-second recovery from common network faults vs. 15–45 minutes for manual response
• Model-Driven Telemetry over gRPC at 30-second poll intervals generates 2–4 GB raw telemetry/day per 500 devices
• MDT pipelines reduce 50,000–200,000 monthly raw alerts to under 500 actionable incidents on a 1,000-device network
• Four AIOps functions in production deployment: telemetry ingestion, ML-based anomaly detection, causal root-cause analysis, autonomous closed-loop remediation
• LLM layer translates natural-language queries into monitoring queries and synthesizes telemetry into human-readable summaries
• Platforms: Cisco DNA Center with AI Assistant, Juniper Apstra + Marvis VNA, Aruba Central AIOps

Deep Dive:

The architectural requirement for closed-loop remediation is often understated in vendor marketing: before the AI can fix anything, you need structured, streaming telemetry. SNMP polling and syslog aggregation are not enough. The minimum viable telemetry stack for closed-loop AIOps is MDT over gRPC (30-second intervals or better) plus BGP Monitoring Protocol for routing state visibility. Without that foundation, the AI only has access to periodic or event-triggered data — and you cannot close a 60-second loop with 5-minute polling.

The alert reduction numbers matter because they reveal what the AI is actually doing most of the time. Reducing 200,000 monthly raw alerts to under 500 actionable incidents is a 99.75% filter rate. That's not anomaly detection doing something sophisticated — that's correlation, deduplication, and baseline normalization working correctly before the ML layer even touches anything complex. The value of the intelligent layer is built on top of a functioning data pipeline, not a substitute for one.

The LLM integration layer is the newest addition and the one least validated in production. Natural language to monitoring query translation works reliably for common queries and starts breaking on complex conditional questions involving multiple device states and time-based context. The useful framing: treat the LLM layer as an acceleration tool for the 80% of queries that are routine, not as a replacement for structured query languages for the 20% that require precision.

So What? Instrument MDT on your highest-traffic 50 devices first — gRPC streaming at 30-second intervals plus BMP for routing state is the minimum viable telemetry stack to feed any AIOps platform. Without structured streaming telemetry in, autonomous remediation is impossible regardless of which platform you buy. Build the data pipeline before evaluating the AI layer.

SourcesThe Network DNA — https://www.thenetworkdna.com/2026/03/ai-driven-autonomous-networking-aiops.html — March 2026

---

UALink v2 Adds gNMI/YANG Management and In-Network Compute — No Silicon Yet

Headline: UALink v2 publishes first open accelerator interconnect with standard network management API

TL;DR: The UALink Consortium released v2 specifications (April 7) including in-network compute capabilities, a separated physical layer spec for faster iteration, and — most interesting for network engineers — a Manageability Specification v1 with native gNMI, YANG, SAI, and Redfish support. No UALink 1.0 products are shipping yet while NVLink-connected products dominate the market.

Key Points:

• UALink 200G DL/PL Specification 2.0: separates physical layer to enable faster silicon iteration without full spec revisions
• Common Specification 2.0: In-Network Compute (INC) allows computation between accelerators during collective operations, targeting all-reduce latency reduction
• Manageability Specification v1 (new): gNMI + YANG + SAI + Redfish — first open accelerator interconnect with a standard network management API
• Chiplet Specification v1: SoC chiplet integration path
• Analyst assessment: UALink "remains desirable" but "quite a way behind NVLink" — no 1.0 silicon shipping
• Published April 7, 2026

So What? The gNMI/YANG Manageability Spec is the detail worth tracking. If UALink silicon ships with standard network management APIs, accelerator interconnect becomes monitorable and configurable using the same toolchain as your switches — blurring the line between fabric management and server interconnect management. When vendors announce UALink-compatible silicon, check whether the Manageability Spec is implemented. That determines whether it fits your existing telemetry stack.

SourcesNetwork World — https://www.networkworld.com/article/4155357/new-v2-ualink-specification-aims-to-catch-up-to-nvlink.html

---

Ultra Ethernet Consortium: Programmable Congestion Management Targets Algorithm Portability

TL;DR: The Ultra Ethernet Consortium's 2026 roadmap centers on Programmable Congestion Management (PCM) — a standard language for congestion control algorithms that allows any algorithm to run portably across any UEC-compliant NIC. Think P4's programmability model, scoped to congestion management.

Key Points:

• PCM: standard language for congestion control algorithms portable across any UEC NIC
• Decouples algorithm development from silicon, allowing congestion control updates without hardware replacement
• UEC 1.0 spec released June 2025, 1.0.1 September 2025; PCM is 2026 roadmap execution
• Context: Keysight/Broadcom LLR + CBFC 800GE interop at OFC 2026 (March) validated transport layer; PCM is the next layer up
• 70% of new AI infrastructure now Ethernet over InfiniBand per Broadcom earnings

So What? RoCEv2 fabrics require DCQCN tuning that is NIC-vendor-specific and notoriously brittle. If PCM lands in production silicon, network teams could swap congestion algorithms via software without hardware replacements. Track which NIC vendors commit to PCM support in their 2026–2027 silicon roadmaps — that's the signal that RoCEv2 fabric management is about to get significantly less painful.

SourcesNetwork World / Ultra Ethernet Consortium — https://www.networkworld.com/article/4113364/ethernet-groups-keep-2026-focus-on-higher-bandwidth-ai-demands.html

---

Nornir Continues Displacing Ansible in Python-Native Pipelines

TL;DR: The trend of Python-native automation engineers choosing Nornir over Ansible for network automation is accelerating in 2026, driven by Nornir's multithreaded execution model, native Python-object return types, and tighter integration with Netmiko, NAPALM, and Scrapli as plugins. Ansible's YAML abstraction layer is increasingly seen as overhead by teams with Python proficiency.

Key Points:

• Key plugins in active use: nornir-netmiko (CLI), nornir-napalm (structured data), nornir-utils (inventory/results), HTTP plugin (REST API targets)
• Nornir returns native Python objects — not YAML/JSON strings — enabling direct integration into validation pipelines, CI test suites, and LLM tool-call schemas
• Thread-pool-based parallelism executes against hundreds of devices concurrently without Ansible fork-per-host overhead
• Hybrid pattern: Nornir for data collection and validation, Ansible for change execution where existing playbooks exist

So What? Pick one use case — config diff collection or structured state validation — and rewrite it in Nornir with nornir-napalm. The Python-object output plugs directly into Batfish or a Pydantic validation model, which is where GitOps-for-networking pipelines need their data.

SourcesOneUptime Blog — https://oneuptime.com/blog/post/2026-03-20-nornir-python-network-automation/view — March 2026

---

Infrahub's Git-Native Source of Truth Has a Six-to-Twelve-Month Adoption Tax

TL;DR: Independent research on NetBox, Nautobot, and Infrahub reveals a clear segmentation: NetBox dominates documentation-first IPAM/DCIM, Nautobot owns enterprise automation workflows, and Infrahub targets relationship-first graph data models — with a 6–12 month productivity ramp that must be scoped into any migration plan.

Key Points:

• NetBox: REST API, extensive plugin ecosystem, consumes 15–25% of engineering time on manual data entry at scale
• Nautobot: commercial licensing $15K–$30K/year for small deployments, built-in job framework, RBAC, approval workflows, GraphQL API, 70% dev time savings vs. manual workflows
• Infrahub: graph-database-native schema, every change tracked through Git versioning, branch-based validation (Batfish/pyATS) before merge — highest operational learning curve
• 25% of teams abandoning GitOps approaches cite workflow complexity as the primary blocker
• The decision framework is clearer than it has been: purpose determines platform

So What? Decision framework: pick NetBox if your #1 pain is IPAM accuracy; Nautobot if you need workflow automation and can absorb the licensing; Infrahub only if your team has strong Git discipline and you want branch-based network change validation as a first-class workflow. Don't adopt Infrahub for speed — its value is in the validation model, not deployment velocity.

SourcesItential Research — https://www.itential.com/research/network-source-of-truth-platforms/ — 2026

---

GLM-5.1: MIT-Licensed 744B MoE Claims Top SWE-Bench Score Over Claude and GPT-5.4

TL;DR: Zhipu AI released GLM-5.1 in early April: 744B total parameters, 40B active via MoE routing, 200K context, MIT license, and a benchmark claim of the top score on SWE-Bench Pro, beating Claude Opus 4.6 and GPT-5.4 on software engineering tasks. Independent third-party verification is pending.

Key Points:

• 744B total / 40B active parameters — similar deployment footprint to Llama 4 Maverick (deployable on 8x H100 configs)
• MIT license — fully commercial, fully self-hostable; removes legal friction Apache 2.0 and custom licenses carry
• 200K token context
• SWE-Bench Pro claim: beats Claude Opus 4.6 (65.3%) and GPT-5.4 (no independent verification yet)
• Pattern: Chinese open-weight labs dropping large MIT-licensed models with competitive benchmark claims is accelerating

So What? If the SWE-Bench claim holds up under independent evaluation, this is the first time a fully open-weight model has credibly contested the top coding slot. Run your own evals on your specific automation tasks before trusting vendor benchmark numbers — SWE-Bench scores on generic software tasks don't translate directly to network automation use cases.

SourcesFazm Blog — https://fazm.ai/blog/new-llm-releases-april-2026

---

Half of All Public MCP Servers Have No Authorization — Seven Thousand Exposed

TL;DR: Researchers catalogued approximately 7,000 internet-exposed MCP servers — roughly half of all known deployments — and found most operate with zero authentication. An independent analysis identified 8 vulnerability classes, with 5.5% showing evidence of tool poisoning: malicious MCP servers returning instructions designed to hijack agent behavior.

Key Points:

• ~7,000 internet-exposed MCP servers of ~14,000 known (roughly 50% exposure rate)
• 8 vulnerability categories; 7.2% show general security flaws, 5.5% show tool poisoning evidence
• Tool poisoning: the attack vector is the server's response payload, not the network layer — malicious responses direct the agent to take unintended actions
• Predates and is architecturally distinct from the MCP Firewall / governance registry pattern (covered April 1–3)

So What? MCP is becoming the new misconfigured S3 bucket: a fast-growing ecosystem where security posture hasn't kept pace with adoption speed. If you're deploying MCP in any capacity, internet exposure without authentication is an agent hijacking surface, not just a configuration oversight. The MCP gateway pattern — treating the MCP proxy as a Zero Trust enforcement choke point — is the direct architectural answer.

SourcesKai Waehner Blog — https://www.kai-waehner.de/blog/2026/04/06/enterprise-agentic-ai-landscape-2026-trust-flexibility-and-vendor-lock-in/ — April 2026

---

Simon Willison: AI Fails at Architecture, Not Implementation

TL;DR: Simon Willison published a detailed post-mortem (April 5) on a project where AI-assisted development produced fast, plausible implementations that collapsed under architectural scrutiny — requiring a full human-led rewrite. The failure mode: AI optimizes for verifiable correctness (does it compile, does it pass tests) but has no signal for whether the design is actually good.

Key Points:

• Used Claude Code to parse 400+ grammar rules — fast, productive at the implementation level
• Repeated architectural dead-ends: designs that "felt productive in the moment but collapsed under scrutiny"
• Final resolution: scrapped AI prototype, rebuilt from scratch with deliberate human-designed architecture
• Core diagnosis: AI lacks objective measures for design quality, only for implementation correctness
• Distinct from Bryan Cantrill's "laziness" critique (April 13): Willison's framing is about verifiability, not verbosity

So What? This maps directly to network automation: AI generates syntactically valid configs and passing CI/CD checks, but structural design of the automation framework — state model, failure handling, rollback logic — still requires human architectural judgment. The useful boundary: implementation and refactoring to AI, architecture and design decisions to human. The line sits wherever "correct" stops having an objective definition.

SourcesSimon Willison — https://simonwillison.net/2026/Apr/5/building-with-ai/ — April 5, 2026

---

Vertiv Acquires BMarko — Vertically Integrates Prefab Manufacturing

TL;DR: Vertiv acquired BMarko Structures, a custom steel-frame prefab fabricator based in South Carolina, to bring structural manufacturing in-house. This closes a supply chain dependency and positions Vertiv as a full-stack physical infrastructure integrator rather than a power/cooling component vendor.

Key Points:

• BMarko founded 2014; recently expanded to 560,000 sq ft fabrication facility in Williamston, South Carolina
• Financial terms undisclosed
• CEO Gio Albertazzi cited customer demand for "time-to-capacity, flexibility, and efficiency across the infrastructure layer"
• Closes in North America Infrastructure Solutions business
• Trend context: this follows Vertiv's MegaMod HDX modular line (April 7) — vertical integration of the structural shell completes the stack

So What? Vertiv is no longer just a power/cooling vendor — expect integrated prefab+power+cooling package quotes with shorter lead times. Evaluate whether the single-vendor convenience is worth the reduced competitive pricing pressure when specifying modular deployments.

SourcesDataCenter Dynamics — https://www.datacenterdynamics.com/en/news/vertiv-acquires-prefab-enclosure-maker-bmarko/ | PRNewswire, April 13, 2026

---

TIA-942 Addendum Codifies AI Data Center Requirements for the First Time

TL;DR: TIA Engineering Committee TR-42.1 has initiated Addendum 1 to ANSI/TIA-942-C specifically covering AI and HPC data center infrastructure — dense GPU cluster cabling, extreme bandwidth requirements, liquid cooling integration, and new power models. Unlike the parallel ISO/IEC 11801-5 Amendment 1 (cabling-only), this addendum covers cooling and electrical alongside cabling, making it the more comprehensive AI DC standard in development.

Key Points:

• Announced March 24, 2026; under development as ANSI/TIA-942-C-1
• Scope: dense GPU cluster cabling, extreme bandwidth, liquid cooling integration, new power models for AI
• Developed alongside new DCE 9000 quality management standard for supply chains and expanded global certifications
• No published finalization date yet — still in active development
• Standards lag deployment reality by 2–4 years in this industry; TIA codifying AI DC requirements signals the experimental phase is ending

So What? Track this addendum actively if you're specifying AI infrastructure today — early alignment avoids future retrofit costs. More immediately: the explicit TIA-942 inclusion of liquid cooling gives procurement and facilities teams a defensible standards reference when making the internal case for direct liquid cooling infrastructure investment. "It's in the ANSI/TIA standard" is a different conversation than "the vendor recommends it."

SourcesTelecom Reseller — https://telecomreseller.com/2026/03/26/tia-advances-ai-ready-data-centers-with-new-ansi-tia-942-addendum-global-certification-leadership-and-expanded-industry-quality-initiative/

---

Microsoft Publishes Zero Trust Reference Architecture for Agentic AI Workloads

TL;DR: Microsoft released a "Zero Trust for AI" framework in March 2026, expanding its Zero Trust Workshop to include a full AI pillar covering 700 security controls across 116 logical groups, with a reference architecture specifically visualizing trust boundaries for agentic systems. The AI assessment pillar goes generally available in summer 2026.

Key Points:

• 700 controls across 116 logical groups covering agentic workloads
• Reference architecture visualizes trust boundaries for agent-to-data and human-to-automated-decision boundaries
• Shifts trust enforcement to every agent-to-tool-call boundary, not just user login events
• Aligns with NIST, CISA, and CIS frameworks — gives security architects an audit-defensible baseline
• AI-specific assessment pillar: general availability summer 2026

So What? Map your agentic AI deployments against the reference architecture's control placement diagram before the summer 2026 assessment tooling goes GA — that gives you two quarters to close gaps before formal audit processes catch them. Prioritize the agent-to-tool-call boundary enforcement controls; that's where traditional Zero Trust architectures have the most coverage gaps.

SourcesMicrosoft Security Blog — https://www.microsoft.com/en-us/security/blog/2026/03/19/new-tools-and-guidance-announcing-zero-trust-for-ai/ — March 19, 2026

---

AI Is Proving Real Mathematical Research — Not Just Olympiad Puzzles

TL;DR: AI systems are now producing publication-quality mathematical results across combinatorics, algebraic geometry, and optimization theory. DeepMind's AlphaEvolve closed previously unknown results in permutation group theory. Ernest Ryu used ChatGPT to close Nesterov's 42-year-old optimization conjecture. In January 2026 alone, fifteen previously open problems moved to solved, with eleven crediting AI involvement.

Key Points:

• AlphaEvolve: discovered previously unknown hypercube structures in permutation group Bruhat intervals
• Ernest Ryu (UCLA): closed Nesterov's 42-year-old optimization conjecture via ChatGPT collaboration
• Javier Gómez-Serrano (Brown/DeepMind) and Ravi Vakil (Stanford) using DeepMind's private FullProof system on algebraic geometry proofs
• UC Berkeley Aletheia agent (preprint Feb 2026): deployed against 700 open Erdős problems, resolved four with Lean-formalized proofs
• Quanta Magazine synthesizing a wave of results across multiple institutions
• Status: mix of peer-reviewed (Ryu conjecture) and preprints (Aletheia)

Why It's Interesting: This is not benchmark performance — it is actual research production. The Lean formalization angle has direct relevance to formal methods in networking: provably correct control plane logic, verified protocol implementations. If AI can close a 42-year-old conjecture with a human-in-the-loop workflow, the pattern applies to any domain with formal representations.

SourcesQuanta Magazine — https://www.quantamagazine.org/the-ai-revolution-in-math-has-arrived-20260413/ — April 13, 2026

---

• Quantum encryption timeline tightens (again): Google published a new Shor's algorithm implementation 10x more efficient than the prior best, reducing the qubit threshold for breaking ECC-based encryption. Combined with the Oratomic findings covered April 7, the practical message is unchanged but the timeline compresses further. Cloudflare has accelerated its post-quantum migration target to 2029. If your organization is on a "done by 2035" timeline, revisit that assumption. [Source: Quanta Magazine, April 3, 2026]

• Oracle AI support quality warning: Independent advisors are warning Oracle customers to watch for support quality degradation and pricing changes as Oracle cuts staff while making large datacenter spending commitments for AI infrastructure. A vendor that's simultaneously cutting support staff and ramping infrastructure spend is a classic service quality inflection point. [Source: The Register, April 13, 2026]

• JAXA rocket failure analysis: Japan's space agency determined that a manufacturing process failure — inadequate temperature compensation in adhesive application — caused the December 2025 satellite launch failure. The adhesive weakened under temperature conditions that weren't properly accounted for in the manufacturing spec. Relevant as a case study in how manufacturing process quality control failures cascade into systems-level failures. [Source: The Register, April 14, 2026]

---

• GPT-6 real-world performance: Independent benchmark runs and practitioner reports will surface throughout the day. Watch for agentic task completion numbers from teams running non-OpenAI benchmark suites.
• Cloudflare Agents Week continues: Day 2 of the five-day product release cadence. Additional announcements expected through April 17.
• AWS Project Houdini details: Following yesterday's report, expect technical briefings and analyst reactions to provide additional specifics on the skid architecture and network configuration standardization approach.
• NANOG 97 CFP closes April 27: If you're submitting, the NEMOPS track and SONiC/AI-fabric topics are explicitly listed as focus areas.

---

Pipeline run: 2026-04-14 | 5 parallel research agents | 15 stories, 3 quick takes | 0 dedup rejections | RSS digest used (4.0 Vertiv/BMarko, 4.0 AWS Houdini, 4.0 TIA-942, 4.0 liquid cooling [sponsored, skipped]) | Quality score: 4/5