blog: calibration v4 — technical accuracy + structure limits

- Fix SR4/DR4 fiber count: both use 8 fibers (4TX+4RX), difference is MMF vs SMF
- Fix power per port: 400G=10-15W/port, 800G=15-25W/port (not 1kW/port)
- Fix pricing context: always distinguish OEM ($1-5K) vs compatible ($200-600)
- Add HARD RULES 15-21: fiber count, power, pricing, no markdown headers,
  max 6 sections, no repeated topics, flow over format
- Add QA CALIBRATION FAILS 16-21: same rules enforced at QA step
- Add fiber/power reference tables with correct values
- Strip markdown (##/###/####/**) from all output — plain text only
- Add Style B gold example (10/10 validated prose article)
- Update STEP5 reality injection with correct SR4→DR4 description
- Update STEP8 kill-AI-tone to strip markdown headers + merge duplicates

packages/api/src/llm/fo-blog-pipeline.ts

@@ -56,6 +56,10 @@ STRICTLY FORBIDDEN:
 - Treating 400ZR and DR4 as equivalent — they are completely different: DR4 = DC leaf-spine (500m, 8 parallel fibers), ZR = DCI/coherent (80km, single fiber, 15-20W). Always separate them clearly.
 - Checklist-style "Final Recommendation" sections — they read like AI. Write as a direct statement, not a bullet list of advice.
 - "shiny new toys" or other marketing-speak dismissals at the end — end with something that STICKS
+- "SR4 uses four fibers / DR4 uses two fibers" — THIS IS WRONG. SR4 and DR4 BOTH use 8 fibers (4 TX + 4 RX). The difference is fiber TYPE (SR4=MMF, DR4=SMF), REACH, and LOSS BUDGET — never fiber count.
+- Power numbers like "1kW per port" or "upwards of 1kW" — HARD FAIL. 400G ≈ 10-15W/port, 800G ≈ 15-25W/port. A fully-loaded 32-port 400G switch draws 1-2 kW total, not per port.
+- OEM pricing for compatible optics — "400G DR4 at $2,000-5,000" is OEM pricing. Compatible vendor range (Flexoptix, FS, ProLabs) is typically $200-600. Always specify OEM vs compatible.
+- Markdown headers (##, ###, ####, **bold headers**) anywhere in the article body — write in plain text. No hash symbols, no asterisk headers. Section titles as plain sentence or not at all.
 
 HARD RULES (non-negotiable — article FAILS QA without these):
 1. Start with a BRUTAL hook — not "If you're still..." but "You're about to sign a PO. Stop."
@@ -81,6 +85,13 @@ HARD RULES (non-negotiable — article FAILS QA without these):
 12. NO CHECKLIST endings: If a section ends with 4 tidy bullet points, rewrite as direct prose.
 13. NEVER USE "PoE" in optics context. PoE = Power over Ethernet for endpoints. Use "power consumption per port", "thermal budget", "chassis power envelope" instead.
 14. DR4 vs ZR: Always separate. DR4 = leaf-spine (500m), ZR = coherent DCI (80km+). Never treat them as variants of the same thing.
+15. SR4 vs DR4 FIBER COUNT: BOTH use 8 fibers (4 TX + 4 RX). Never say SR4=4 or DR4=2. The difference is MMF vs SMF, reach, and loss budget.
+16. POWER PER PORT: 400G ≈ 10-15W/port. 800G ≈ 15-25W/port. A switch chassis draws kW total — NOT per port. Writing "1kW per port" is a HARD FAIL.
+17. PRICING CONTEXT: Always distinguish OEM ($1,000-5,000) from compatible ($200-600 for 400G DR4). Mixing them destroys credibility.
+18. NO MARKDOWN HEADERS: The article MUST NOT contain ##, ###, ####, or **bolded header lines**. Plain text only. Sections flow naturally or use a single line break.
+19. MAX 6 SECTIONS: An article with more than 6 named sections reads like a framework, not a human. Combine or cut. Migration guides get 6, opinion pieces get 4.
+20. NO REPEATED TOPICS: Each concept (cleaning, polarity, power) may be explained EXACTLY ONCE. If cleaning appears in "hidden costs" it cannot reappear in "cabling reality" as a separate block.
+21. FLOW OVER FORMAT: Prefer connected narrative paragraphs over modular section-bullet-example patterns. The engine's default output feels "assembled" — it should feel "written".
 
 REFERENCE VALUES:
 - SFP+ SR: Tx -8.2 to +0.5 dBm, Rx sensitivity -18.0 dBm, 1.0W typical
@@ -92,6 +103,21 @@ REFERENCE VALUES:
 - Power budget margin: minimum 3 dB recommended
 - BER: pre-FEC <2.4×10^-4 (KP4), post-FEC <10^-15
 
+FIBER COUNT FACTS (memorize — getting this wrong kills credibility):
+- 100G SR4: 8 fibers total (4 TX + 4 RX), MPO-12, MULTIMODE (OM3/OM4), 100m
+- 100G DR: 2 fibers (1 TX + 1 RX), LC duplex, SINGLE-MODE, 500m
+- 400G DR4: 8 fibers total (4 TX + 4 RX), MPO-12, SINGLE-MODE (OS2), 500m
+- 400G FR4: 2 fibers (LC duplex), CWDM4, SINGLE-MODE, 2km
+- 400G SR4.2 (SWDM4): 8 fibers, MULTIMODE, 100m
+- 400ZR: 2 fibers (LC duplex), coherent, 80-120km
+- KEY: SR4 and DR4 differ in FIBER TYPE (MMF vs SMF), reach, and loss budget. NOT in fiber count.
+
+POWER PER PORT FACTS (not per chassis — per port):
+- 100G QSFP28: ~3.5W typical
+- 400G QSFP-DD: ~10-15W typical (DR4 ~12W, FR4 ~12W, ZR ~15-20W)
+- 800G OSFP: ~15-25W typical
+- A 32-port 400G switch total: ~400-600W chassis power. NOT "1kW per port".
+
 CONTENT MODULES (use 2-3 per article):
 - What breaks in production
 - Migration pain (old → new)
@@ -176,7 +202,7 @@ MANDATORY SECTIONS (article fails without ALL of these):
 5. HIDDEN COSTS NOBODY MENTIONS (MANDATORY):
    - Cleaning effort (MPO requires inspection scope at $2K, not just IPA wipes)
    - Troubleshooting time ($150/hr engineer × 4 hours for a dirty connector = $600 per incident)
-   - Migration costs (SR4→DR4 = different fiber count, new patch panels, re-termination)
+   - Migration costs (SR4→DR4 = same fiber count but completely different fiber TYPE, new patch panels, re-termination — people assume it's just an optic swap, it's not)
    - Cabling redesign (MMF→SMF = $50-200 per drop × number of drops)
 6. WHEN NOT TO USE THIS (MANDATORY):
    - Small DCs (<50 ports)
@@ -201,11 +227,17 @@ STYLE:
 
 STYLE CHOICE:
 Two valid output formats — pick based on the outline's angle and audience:
-- STYLE A (structured): Use headers, specific failure scenarios, bullet lists where needed. Best for troubleshooting guides, migration how-tos, deep dives.
-- STYLE B (prose narrative): No headers after the title, no bullet points, pure flowing paragraphs (1-3 sentences each). Best for roundups, opinion pieces, "state of the technology" takes. The ending MUST be a single reframe sentence, not a list.
+- STYLE A (structured): Max 6 sections, each concept mentioned ONCE. Use plain-text section titles (not markdown headers), failure scenario blocks, minimal bullets. Best for troubleshooting guides, migration how-tos.
+- STYLE B (prose narrative): No section titles, no bullet points, pure flowing paragraphs (1-3 sentences each). Best for roundups, opinion pieces, "state of the technology" takes. Ending MUST be a one-liner reframe, not a list.
 
 Both styles are 10/10 valid. Choose the one that fits the angle selected in the outline.
 
+FORMATTING RULES (both styles — non-negotiable):
+- NO markdown headers: ## / ### / #### / **Bold Title:** are FORBIDDEN in the article body
+- NO repeated topics: if cleaning appears in one section, it does NOT appear in another
+- MAX 6 sections in Style A — more than that reads like a framework, not an article
+- PLAIN TEXT ONLY: write section transitions as prose, not as formatted labels
+
 MINIMUM 2000 words. No placeholders. No TODO markers. Complete article.
 
 Outline:
@@ -232,7 +264,7 @@ MANDATORY ADDITIONS (check that ALL exist, add if missing):
 
 2. CABLING REALITY (add if missing):
    - MPO polarity nightmare: "You ordered Type A cables for a Type B system. Half your links are crossed. That's a $15K re-termination job."
-   - SR4 to DR4 migration: "SR4 uses 8 fibers (4 Tx, 4 Rx). DR4 uses 8 fibers (4 parallel SMF). Different fiber, different patch panels, different everything."
+   - SR4 to DR4 migration: "SR4 and DR4 both use 8 fibers (4 Tx, 4 Rx). What changes is the fiber TYPE: OM3/OM4 multimode becomes OS2 single-mode. That difference alone changes your entire physical layer — new patch panels, tighter loss budgets, different cleaning tolerances."
    - Cleaning: "An MPO-12 connector has 12 fiber end-faces. ONE dirty end-face = entire link degraded. You need an inspection scope, not just wipes."
 
 3. HIDDEN COSTS (add if missing):
@@ -311,6 +343,8 @@ REMOVE:
 - Phrases like "it is important to note", "one should consider"
 - Any section that ends with a tidy 4-item bullet list
 - Perfectly symmetrical sections (same length = AI fingerprint)
+- ALL markdown headers: ##, ###, ####, **Bold Section Title:** — strip every single one. Write plain text only.
+- Repeated explanations: if cleaning, polarity, or power is in two sections, merge them into the first occurrence and remove the second.
 
 REPLACE WITH:
 - Natural, slightly imperfect flow
@@ -377,6 +411,18 @@ CALIBRATION FAILS (auto-reject — fix before returning):
 15. HIDDEN COSTS TOO CLEAN: If the hidden costs section feels like a polished table, roughen it.
     → Bad: "$350 optic → $2,400 troubleshooting cost"
     → Good: "That $350 optic turned into a multi-thousand-dollar problem because someone skipped the connector cleaning."
+16. SR4/DR4 FIBER COUNT ERROR (HARD FAIL): Check for "SR4 uses 4 fibers", "DR4 uses 2 fibers", or any claim that they differ in fiber count.
+    → REPLACE: Both use 8 fibers (4 TX + 4 RX). The difference is MMF (SR4) vs SMF (DR4), reach, and loss budget. This is a credibility-destroying technical error.
+17. POWER PER PORT ERROR (HARD FAIL): Check for "1kW per port", "upwards of 1kW per port", or any per-port wattage claim over 50W for 400G/800G.
+    → REPLACE with: "400G optics draw ~10-15W per port. A fully-loaded 32-port 400G switch total chassis power: 400-800W. Not per port — total."
+18. PRICING CONTEXT MISSING: If pricing is given without specifying OEM vs compatible, flag it.
+    → Compatible 400G DR4 (Flexoptix/FS.com/ProLabs): $200-600. OEM (Cisco/Juniper/Arista branded): $1,000-5,000. ALWAYS specify which.
+19. MARKDOWN HEADERS PRESENT (HARD FAIL): Check for ##, ###, ####, or lines starting with **SomeTitle:** used as section headers.
+    → REMOVE ALL. Replace with plain text transition sentences or nothing at all.
+20. TOO MANY SECTIONS: Count distinct named/headed sections. If more than 6 in Style A, or any in Style B, the article reads like a framework.
+    → COMBINE redundant sections. "Wrong assumptions" + "What engineers miss" = one section.
+21. REPEATED TOPICS: Check if cleaning, polarity, or power budget are each explained more than once across sections.
+    → Each concept gets ONE home. Mention it elsewhere as a single-sentence reference at most.
 
 For each issue:
 - Quote the problematic text
@@ -544,6 +590,57 @@ STYLE B RHYTHM RULES:
 - The ending is a one-liner that reframes everything: not a conclusion, a punch
 - NEVER end Style B with a list or action items — just the thought that sticks
 
+━━━ STYLE B GOLD EXAMPLE (10/10 validated, 2026-03-31) ━━━
+Topic: 400G/800G migration guide as pure prose. This is the TARGET voice.
+
+  "You're about to sign a PO for 400G or even 800G optics.
+
+  On paper, it looks easy. More bandwidth, fewer ports, cleaner design. The vendor tells you it's mature, widely deployed, no surprises.
+
+  They're not lying.
+
+  But they're also not the ones debugging your network at 2AM.
+
+  Because the problem with 400G isn't the technology. The problem is that people treat it like an upgrade. It's not. It's a different game.
+
+  Most teams come from 100G SR4 and assume the jump is incremental. Same idea, just faster. Same cabling, just different optics.
+
+  That assumption is where things start to drift.
+
+  SR4 and DR4 both run parallel optics with eight fibers. On paper, that looks like continuity. In reality, everything around it tightens up. Loss budgets get stricter. Tolerance for dirt drops. What used to 'just work' suddenly doesn't.
+
+  You don't notice that in the lab.
+
+  In a lab, everything is short, clean, controlled. You plug it in, links come up, done.
+
+  Production is where reality kicks in.
+
+  Mixed optics, mixed firmware, longer runs, older patch panels — and suddenly links don't behave the way they did in testing. Not completely broken. Just unstable enough to cost you time.
+
+  That's usually the first surprise.
+
+  [continues — key insight sections flow as connected paragraphs, no headers]
+
+  400G doesn't break things. It exposes them.
+
+  [...]
+
+  None of this means you shouldn't move to 400G.
+
+  You should.
+
+  [...]
+
+  The question is whether everything around them will."
+
+KEY ELEMENTS OF THIS STYLE:
+  - Opens in media res — no setup, no "In today's world"
+  - Each paragraph = one thought, one beat
+  - Technical facts woven into narrative, not listed
+  - No section headers anywhere
+  - Ending is open, not prescriptive
+  - Tone: been there, done that, not afraid to say so
+
 WRONG PATTERNS (both styles — never produce):
 ❌ "Thoroughly Test Your PoE Budget:" (PoE = wrong context, checklist = wrong format)
 ❌ "QSFP-DD DR4 (Direct Attach)" (DR4 ≠ Direct Attach — DAC is Direct Attach Copper)
@@ -553,6 +650,12 @@ WRONG PATTERNS (both styles — never produce):
 ❌ Ending with "consider your options carefully" or any variant of that
 ❌ Starting a new paragraph with "Furthermore", "Additionally", "It's worth noting"
 ❌ Perfectly symmetrical sections (every section same length = AI fingerprint)
+❌ "SR4 uses four fibers" / "DR4 uses two fibers" — BOTH use 8 fibers. Wrong fact, hard credibility kill.
+❌ "1kW per port" for 400G — reality is ~12W/port. Hard technical fail.
+❌ "400G DR4 at $2,000-5,000" without specifying OEM — compatible pricing is $200-600.
+❌ ## or ### section headers inside the article — plain text only, always.
+❌ 8+ sections in one article — looks assembled, not written.
+❌ Cleaning explained in "hidden costs" AND again in "cabling reality" — pick one home.
 
 --- END GOLD STANDARD ---
 `;