feat(training): add blog-014 new_product and blog-015 competitor_analysis

Completes training data coverage for all 8 blog types:
market_alert(2), comparison(1), technology_deep_dive(4), tutorial(3),
hype_cycle(1), buying_guide(1), migration_guide(1), new_product(1),
competitor_analysis(1) — 15 gold-standard articles total

blog-training-data/README.md

@@ -17,6 +17,10 @@ Gold-standard blog posts generated by Claude Sonnet (claude-sonnet-4-20250514) f
 | blog-009-100g-to-400g-migration-what-breaks.md | 100G to 400G Migration: What Actually Breaks and Why | migration_guide | 9/10 |
 | blog-010-qsfp-dd-vs-osfp-form-factor-reality.md | QSFP-DD vs OSFP: The Form Factor War That Already Ended | technology_deep_dive | 9/10 |
 | blog-011-transceiver-procurement-checklist.md | The Transceiver Procurement Checklist Nobody Gave You | tutorial | 9/10 |
+| blog-012-coherent-vs-direct-detect-decision.md | Coherent vs. Direct Detect: The Decision Your Network Will Make for the Next Decade | technology_deep_dive | 9/10 |
+| blog-013-price-drop-timing-when-to-buy.md | When to Buy: Reading the Transceiver Price Cycle Before It Reads You | market_alert | 9/10 |
+| blog-014-800g-new-products-what-ships.md | 800G Is Shipping: What's Actually Available and What You Can Deploy Today | new_product | 9/10 |
+| blog-015-compatible-vendor-comparison-who-to-trust.md | Compatible Transceiver Vendors in 2026: Who Does the Testing and Who Just Says They Do | competitor_analysis | 9/10 |
 
 ## Quality Rules Met (per article)
 

blog-training-data/blog-014-800g-new-products-what-ships.md

@@ -0,0 +1,32 @@
+---
+title: "800G Is Shipping: What's Actually Available and What You Can Deploy Today"
+type: new_product
+target_audience: technical
+score: 9/10
+---
+
+When a new generation of optical modules ships, there's a gap between "shipping" and "available for your environment." 800G is shipping. Understanding the difference between what's on a datasheet and what's production-ready for a specific use case takes more precision than the press releases provide.
+
+The form factors matter first. 800G pluggables come in two physical formats: OSFP (Octal Small Form Factor Pluggable) and QSFP-DD800. These are not interchangeable. OSFP is the larger module, with better thermal headroom — up to 15W versus around 12W for QSFP-DD800. Most 800G switch platforms shipping today use OSFP ports. QSFP-DD800 platforms are arriving in 2026 but are in early deployment. Before ordering 800G modules, verify which form factor your switch requires.
+
+The speed is achieved through different lane architectures. 800GBASE-SR8 uses eight lanes of 100G each over OM5 or OM4 multimode fiber with MPO-16 connectors. 800GBASE-DR8 uses eight lanes of 100G over OS2 single-mode fiber with MPO-16 connectors at up to 500m. 800GBASE-FR8 extends single-mode reach to 2km over eight wavelengths. Each is a different electrical and optical architecture. An 800G SR8 is not a direct replacement for 800G DR8 — the fiber plant, connector type, and switch port configuration are all different.
+
+The switch ASICs driving 800G ports are predominantly NVIDIA Spectrum-4 and Broadcom Tomahawk 5 in current deployments. Both support 800G OSFP natively. Arista's 7060X6 series, Cisco Nexus platforms based on Cisco Silicon One G200, and various white-box platforms using these merchant ASICs are where you'll find actual 800G deployments in production today. If you're evaluating 800G, verify the specific ASIC and software release — not just the platform family.
+
+The application scope is currently narrow. 800G is primarily deployed in two contexts: AI/ML training clusters where GPU-to-GPU bandwidth drives fabric requirements, and hyperscale spine layers where port density and power efficiency at 800G justify the premium. An enterprise data center running SQL databases and web applications has no performance argument for 800G in 2026. The physics works, the economics don't. 400G DR4 at $25-40 per port is the right answer for general enterprise spine until 800G prices reach a comparable level.
+
+For AI/ML fabric, the math is different. A single DGX H100 system has eight 400G or eight 800G HDR network ports depending on the InfiniBand variant, plus multiple 10/25G management paths. Connecting pods of these systems requires the highest bandwidth-per-rack possible. At 800G with 128-port switches, a single 2RU switch handles the entire port count for a 16-node H100 cluster with full bisection bandwidth. At 400G, you need multiple switches plus additional uplinks. The density argument at AI scale is real.
+
+Module pricing in mid-2026 for production 800G OSFP SR8: OEM prices from the major platform vendors run $1,500-2,500 per module. Compatible options from tier-1 vendors are appearing at $600-1,200, though the compatible 800G ecosystem is substantially thinner than the 400G compatible ecosystem. Fewer manufacturers, shorter track records, and platform validation that's still being established. The guidance for 800G compatible at this stage is the same as it was for 400G ZR in 2021: ask for the specific platform validation matrix, verify the test data yourself, and start with a pilot before committing to a large batch.
+
+The fiber infrastructure question is often where 800G hits a wall. 800GBASE-SR8 uses MPO-16 connectors, not MPO-12. If your existing multimode fiber infrastructure uses MPO-12 pre-terminated assemblies — which covers most enterprise MMF installed before 2020 — you cannot plug 800G SR8 modules directly into it. You need either MPO-16 assemblies (expensive, require re-cabling) or 800GBASE-DR8 on single-mode fiber. Many organizations evaluating 800G for existing campus data centers discover this constraint after reviewing the spec sheet, not before. Check your installed fiber and connector inventory before the procurement conversation.
+
+The DOM monitoring profile for 800G differs from 400G. Eight TX lanes instead of four means eight sets of power readings, per-lane bias current, per-lane temperature data. If your network management system pulls DOM data via SNMP or streaming telemetry, verify it handles the expanded 800G DOM structure. Some older NMS platforms have hardcoded assumptions about QSFP28 lane count that fail silently on 800G modules rather than returning an error.
+
+Interoperability between manufacturers at 800G is less established than at 400G. IEEE 802.3df standardizes 800GBASE-SR8 and 800GBASE-DR8, but the initial implementations came from a small number of manufacturers and were deployed in validated pairings before broad third-party testing was possible. By mid-2026, the major platform/module combinations are well-characterized. Mixing OSFP modules from different manufacturers on the same switch is generally fine for SR8 and DR8 (both are defined optical specs with clear compliance testing). The edge cases are in 800G coherent and proprietary higher-reach variants, where firmware interaction matters more.
+
+Power consumption: an 800G OSFP SR8 draws approximately 10-12W per module. A 128-port 800G switch draws 10-12kW just in optics, before the ASIC and platform overhead. Compare to a 32-port 400G switch drawing 1-3kW in direct-detect optics. The power infrastructure requirements for AI fabric are not incremental — they require a different data center PDU density, cooling design, and power delivery architecture. 800G deployments go alongside significant infrastructure investment in power and cooling, not just in the networking layer.
+
+The roadmap beyond 800G is 1.6T, which is in draft standardization and initial sampling from a handful of manufacturers. The lane architecture moves to 200G per lane, which requires a new generation of DSP and SerDes. Practical deployment of 1.6T is realistically 2027-2028 at the earliest. If you're making 800G decisions today, you're not racing ahead of the curve — you're at the beginning of the mainstream adoption slope, not the bleeding edge.
+
+The right questions before a purchase: What ASIC and switch platform? What fiber plant — MMF with MPO-12 or SMF? What software version? Is the module vendor on the platform's tested vendor list for your specific linecard? What's the DOA rate history for 800G OSFP from this vendor? Those questions determine whether an 800G deployment is straightforward or expensive.

blog-training-data/blog-015-compatible-vendor-comparison-who-to-trust.md

@@ -0,0 +1,32 @@
+---
+title: "Compatible Transceiver Vendors in 2026: Who Does the Testing and Who Just Says They Do"
+type: competitor_analysis
+target_audience: sales
+score: 9/10
+---
+
+The compatible transceiver market has a problem that doesn't affect branded markets: every vendor claims the same things. They all say "tested compatible," they all have warranty policies, they all list the same platforms. The difference between a vendor that ships modules that work and a vendor whose modules cause production incidents is invisible in the marketing. It's visible in the data.
+
+Here's what the data actually tells you about the major vendors operating in the European and North American market in 2026.
+
+FS.com is the dominant volume player by price and SKU count. For standard direct-detect optics — 10G SR/LR, 25G SR/LR, 100G SR4/LR4/DR4 — their pricing is consistently at or near the market floor. $18-25 for 100G SR4 puts them at the lowest end of the credible compatible market. Their testing documentation is publicly available, their DOA rates for commodity optics are in line with the industry (0.2-0.5%), and their burn-in process for standard optics is documented. The weakness is coherent: FS.com's DWDM and ZR portfolio is less mature, with fewer platform validations published for coherent line cards. For commodity 100G and 400G direct-detect, they are a legitimate choice.
+
+Flexoptix operates differently from the pure volume players. The business model centers on EEPROM programming as a service — you specify the platform you're deploying on, and the module is programmed with the correct vendor identifier and part number for that switch before it ships. This matters in environments where the switch vendor's software checks EEPROM data and either blocks traffic or logs warnings for unrecognized modules. The price premium over the cheapest market alternatives is typically 20-40% for commodity optics, justified primarily by platform-specific programming and the ability to return modules for reprogramming if the platform changes. For environments running multiple switch platforms that require specific EEPROM data, the programming service has real operational value.
+
+ProLabs is the enterprise-focused compatible vendor that most closely mirrors the sales model of OEM vendors. They have formal account management, volume pricing programs, and a tested vendor list that covers the major enterprise switch platforms in depth. Their pricing is in the middle of the market — not the cheapest, significantly cheaper than OEM list. The strength is in breadth of platform coverage and documentation quality. If you need tested compatibility data for a Cisco Catalyst 9000 with a specific IOS-XE version, ProLabs is likely to have it documented. They also have advance-replacement RMA programs standard, not optional.
+
+ATGBICS occupies the mid-market: lower prices than ProLabs, more complete testing documentation than many grey-market resellers. Their 100G SR4 pricing runs in the $25-35 range. Platform testing focuses on the common enterprise switches. Their documentation for niche platforms is thinner. The DOA rate data they publish is at the high end of acceptable (0.3-0.5%) — not alarming, but worth factoring into procurement decisions for large batches.
+
+10Gtek ships from China directly to buyers worldwide and represents the factory-direct price tier. Their 100G SR4 pricing can go below $15 in volume. The tradeoff is visible in what's absent: minimal platform-specific EEPROM documentation, burn-in testing not standard, DOA rates not published. For non-production environments, lab infrastructure, and test benches where downtime cost is low, the price point is compelling. For production infrastructure where a dead module costs engineering time to replace and diagnose, the savings may not survive the first incident.
+
+The grey-market tier — unbranded modules from AliExpress, Amazon third-party sellers, and small importers — is a different risk profile entirely. These modules are often manufactured on the same production lines as the branded compatible modules, sometimes even the same physical hardware. The risk isn't necessarily higher failure rates in the first week of operation. The risk is in the unknowns: no DOA data, no burn-in testing documentation, no EEPROM consistency guarantee, no warranty with actual replacement terms. For a single test module, this tier is fine. For 500 modules deployed in production, it's a different calculation.
+
+The coherent segment has different competitive dynamics. 400G ZR and ZR+ are platform-validated, not just spec-compliant. The platforms that matter — Arista 7160/7280 series, Cisco ASR 9000, Juniper PTX10000 — have specific firmware revisions and linecard combinations that need to be tested. In this segment, the tier-1 compatible vendors (Flexoptix, ProLabs, Coherent/II-VI compatible lines) have done the validation work. The tier-2 vendors generally have not. Buying a $200 400G ZR from a vendor with no published platform test matrix is buying an untested assumption.
+
+A useful procurement filter: ask for the EEPROM read output for the specific platform you're deploying. A module vendor who can produce a hex dump of the EEPROM as it will be programmed for your platform within 24 hours has the testing infrastructure to support the claim. A vendor who says "we'll just swap it if it doesn't work" has not done the testing. The difference matters at scale.
+
+The warranty comparison is not a useful differentiator until you read the details. "Lifetime warranty" appears on products from FS.com, ATGBICS, 10Gtek, and most others. The operative question is advance replacement versus return-first. A vendor who ships a replacement before you return the failed unit means a production incident lasts hours, not days. A vendor who requires you to ship first means a day-plus of downtime while you wait for return shipping and replacement processing. ProLabs and Flexoptix standard terms include advance replacement. Most grey-market and AliExpress options do not.
+
+Practical decision framework: for commodity 100G and 400G DR4/SR4/LR4 in an environment where switch software doesn't require specific EEPROM data, FS.com and ATGBICS represent the best price-to-documentation ratio. For environments with multiple switch platforms requiring specific EEPROM programming, Flexoptix's programming service has operational value that justifies the premium. For coherent 400G ZR in a production DCI environment, use a vendor who can provide the test matrix for your specific platform and linecard. For everything else: get the DOA rate in writing, verify burn-in testing is standard, and test a pilot batch before committing.
+
+The vendors who publish their DOA rates by SKU, provide batch test reports on request, and have advance replacement as a standard term (not a paid add-on) are the vendors who have built the operational infrastructure to stand behind what they ship. That's not a long list, which is exactly why it's a useful filter.