sinapse-ai 9.4.0 → 9.5.0

package/squads/squad-brand/knowledge-base/ai-visual-generation-canon.md

@@ -0,0 +1,234 @@
+# Knowledge Base: AI Visual Generation Canon
+
+> **Status:** Canon — canonical source of truth for how `squad-brand` (Meridian) produces brand-consistent visual output at scale using AI image/video models.
+> **Version:** 1.0 (shipped with squad-brand v2.0.0 framework-upgrade / fw-v2.3)
+> **Depends on:** `ai-api-capabilities.md` (capability catalog), `prompt-engineering-visual.md` (prompt mechanics)
+> **Supersedes:** ad-hoc generation workflows
+
+## Why this doc exists
+
+`ai-api-capabilities.md` is a **catalog** — "which API does what". `prompt-engineering-visual.md` is a **mechanics reference** — "how to write a prompt". Neither is a **canon** — "the rules that must hold for every brand asset we ship". In 2026, ~80% of brand visual production runs through AI models, and the gap between "I generated something that looks cool" and "I generated something that is on-brand and legally safe to ship" is the single biggest risk factor for brand drift.
+
+This KB is the canonical playbook. Every agent in `squad-brand` that produces visual assets (brand-collateral-designer, brand-identity-designer, brand-motion-vfx, brand-creative-engineer) MUST comply with this canon for any asset that will leave the studio.
+
+---
+
+## 1. The 5 brand-consistency dimensions (lock these, always)
+
+Every batch of AI-generated visual output is judged on 5 dimensions. Each dimension has a **lock protocol** — the concrete steps that enforce consistency across a batch, across sessions, across months.
+
+### 1.1 Style lock
+
+**What it means:** The "visual vocabulary" — illustration style, render style, graphic sensibility. A batch must NOT mix photorealistic shots with flat illustrations unless the brand canon explicitly allows both.
+
+**Lock protocol:**
+1. Declare the style in the brand canon (e.g. "editorial photography, natural light, medium grain").
+2. For each generation session, prepend the exact canon style phrase to every prompt.
+3. Pin 1-3 reference images in the model session (every modern model accepts reference input — Midjourney `--cref`, Stable Diffusion IP-Adapter, DALL-E upload, Flux img2img).
+4. If the model drifts mid-batch, STOP and re-seed with the reference — never "work around" drift in post.
+
+### 1.2 Lighting lock
+
+**What it means:** Direction, hardness, color temperature, time-of-day. Lighting is the highest-signal brand marker after color and is often the first thing that drifts.
+
+**Lock protocol:**
+1. Codify lighting in the canon: direction (N-SW 45°), hardness (soft/medium/hard), color temp (K), intensity (flat/dimensional/dramatic).
+2. Use precise vocabulary in prompts — "soft north-light 5500K, subtle rim" beats "nice lighting".
+3. Reference-image lock handles lighting implicitly when the reference is on-brand.
+4. For video: lock AFTER first frame — subsequent frames must inherit the anchor frame's lighting.
+
+### 1.3 Palette lock
+
+**What it means:** Every color in the output must either be in the brand palette OR be a naturalistic color that the brand palette allows in context (e.g. skin tones, sky).
+
+**Lock protocol:**
+1. Brand canon declares: primary palette (3-5 colors), secondary palette (5-10), neutrals (3-5), allowed environmental colors (context-dependent).
+2. Prompts reference specific named colors ("cerulean #1E40AF, warm ivory #FAF8F1").
+3. Post-generation palette extraction — run a script (or visual check) on the output against the declared palette. Any color > 5% area that is NOT in the allowed set is a FAIL.
+4. For illustrations/vectors: use a color-palette LUT or Figma variable binding to hard-lock at the design-tool layer.
+
+### 1.4 Subject lock
+
+**What it means:** The specific entity (person, product, space) must be recognizably the same across all shots in a batch. This is the most technically demanding dimension and the one where naive prompting fails hardest.
+
+**Lock protocol:**
+1. For **real people/products** the brand owns: use a trained LoRA / DreamBooth / textual inversion model of the subject (or Midjourney `--cref` / Flux reference).
+2. For **synthetic characters** that recur: lock via seed + prompt signature + reference image pinning. Treat the first accepted generation as the canonical reference and build the LoRA from the batch.
+3. For **products**: the reference image is the source of truth — every shot is inpainting/compositing, not free generation. Never trust a pure text-to-image generation of a real product.
+4. When the subject is **generic** (a crowd, a landscape): no subject lock required but the other 4 dimensions still apply.
+
+### 1.5 Composition lock
+
+**What it means:** Framing, aspect, focal points, negative space behavior. Composition is the dimension most often "drifted" by AI models because they default to centered hero compositions.
+
+**Lock protocol:**
+1. Specify aspect explicitly in every prompt (aspect params, `--ar`, output resolution).
+2. Specify framing vocabulary: close-up / medium / wide / establishing, eye-level / low / high angle.
+3. For batches that need composition variety: define the variety upfront (3 framings × 2 angles = 6-cell grid) and generate per cell, not free.
+4. For templates (social media slots, email headers): use ControlNet/reference with the canonical layout wireframe.
+
+---
+
+## 2. Model selection matrix (when to use which)
+
+| Output type | First choice | Second choice | When to use which |
+|-------------|-------------|---------------|-------------------|
+| Hero editorial photography | Flux / Midjourney v6+ | Stable Diffusion + SDXL refiner | Flux for realism + control, MJ for aesthetic polish |
+| Product shots on white | Stable Diffusion + ControlNet | DALL-E 3 edit | SD for compositing precision, DALL-E when product reference isn't available |
+| Illustration series | Midjourney + style LoRA | Flux + reference image | MJ has the best style coherence for non-photo |
+| Text-in-image (posters, quote cards) | Ideogram | DALL-E 3 | Ideogram is the only model with reliable text rendering in 2026 |
+| Character consistency across batch | LoRA-trained Flux / SD | Midjourney `--cref` | Train once, reuse everywhere |
+| Product in lifestyle scene | SD img2img with product reference | Photoroom API → AI background | SD if you need the product to exist; Photoroom if you have the product photo and need the context |
+| Video (short, 2-5 sec) | Runway Gen-3 / Kling 2.0 | Pika Labs | Runway for cinematography, Kling for physical realism |
+| Motion graphics | Figma → Rive → Lottie | After Effects + plugin generation | Stay vector whenever possible — AI video is still expensive per second and drift-prone |
+| Brand animation (logo, intro) | Manual (After Effects / Rive) | — | Brand logo motion is canon — NEVER free-generate, always hand-authored |
+
+**Rule:** If the output will carry the **wordmark**, the **logomark**, or a **recognizable brand character**, the AI model is assisting a human motion designer, not driving the work. Free-generation is banned for identity-critical assets.
+
+---
+
+## 3. Style-lock techniques (detailed reference)
+
+### 3.1 Reference images
+- Always the first line of defense. Modern models all support it.
+- Pin 1-3 references — more than 3 causes the model to average and lose specificity.
+- References should be EXACT matches for the dimensions you're locking (style reference for style, subject reference for subject, composition reference for composition).
+
+### 3.2 LoRA / DreamBooth / textual inversion
+- For recurring subjects or recurring style, train a LoRA (recommended: 15-50 images, 1500-3000 steps on SDXL or Flux).
+- LoRA is the most reliable lock for subject + style. It's an hour of setup that saves weeks of re-prompting.
+- Maintain a LoRA registry per brand: `brands/{brand}/loras/{subject}/v{n}.safetensors` with changelog.
+
+### 3.3 Seed control
+- Locking a seed pins one axis of randomness. Useful for A/B style tests (same seed, different prompts).
+- NOT sufficient alone — same seed + different prompt still drifts.
+
+### 3.4 ControlNet / Canny / depth / pose
+- Pin composition and structure via a control map derived from a wireframe/sketch.
+- Essential for templated outputs (social media slot layouts, email header banners).
+- Control + LoRA + seed = maximum lock.
+
+### 3.5 Negative prompts
+- Every brand canon should include a banned vocabulary list. Hard-code the list into every prompt as negatives.
+- Example bans: "watermark", "logo overlay", "low quality", "deformed", "stock photo look", plus brand-specific bans.
+
+---
+
+## 4. Batch consistency QA checklist (mandatory per batch)
+
+Before any batch ships out of the studio, it MUST pass this checklist:
+
+| # | Check | Threshold | Action if FAIL |
+|---|-------|-----------|---------------|
+| 1 | Style dimension matches canon reference | Visual inspection | Regenerate with style reference |
+| 2 | Lighting dimension matches canon | Visual inspection | Regenerate with lighting reference |
+| 3 | Palette extraction — all > 5% colors in allowed set | Script / visual | Color-correct or regenerate |
+| 4 | Subject dimension — same entity across all shots where required | Visual + reference comparison | Regenerate with stronger LoRA/reference |
+| 5 | Composition dimension — matches declared framing grid | Visual | Regenerate per cell |
+| 6 | Text accuracy (if text-in-image) — spelling, hierarchy | Proofread | Regenerate with Ideogram or DALL-E |
+| 7 | Face/hand anatomy sanity check | Visual | Inpaint fixes or regenerate |
+| 8 | Watermark / model artifact absence | Visual | Remove via inpaint |
+| 9 | Resolution meets delivery spec | Metadata check | Upscale via Topaz / Flux refiner |
+| 10 | No unintended text/logos visible (hallucinated signage, off-brand logos) | Visual | Inpaint away |
+
+**A batch that fails any check does not ship. Period.** The batch goes back to regeneration. Partial passes are not acceptable — AI generation is cheap enough that "re-run the batch" is always the right call.
+
+---
+
+## 5. Legal / IP guardrails (non-negotiable)
+
+### 5.1 Training data provenance
+- Prefer models with documented training data (Adobe Firefly, Getty Generative, Flux-with-license).
+- For SD/MJ/DALL-E: check the brand's legal policy on generative use. Some brands forbid models trained on scraped web data.
+- If the model is used for backgrounds or textures where likeness doesn't matter, provenance is lower-risk.
+- If the model is used for faces, characters, or recognizable art styles, provenance is high-risk — prefer trained-LoRA approach with brand-owned training data.
+
+### 5.2 Commercial-use flags
+- Every delivered asset must have a logged "commercial use cleared" status.
+- Required metadata per asset: model used, model license, prompt, seed, reference images used, LoRA(s) used, date.
+- Store metadata in sidecar JSON or EXIF — never "in memory".
+
+### 5.3 Right of publicity (people's likenesses)
+- NEVER generate outputs depicting a recognizable real person without signed release.
+- Synthetic people generated by AI are OK if the output does not resemble any real celebrity (run a reverse-image check against known-persons databases before ship).
+- Stock-model LoRAs (models who have signed releases for AI training) are the safest source when you need real-looking faces.
+
+### 5.4 Trademark avoidance
+- Hallucinated logos, branded signage, branded products in scenes → inpaint away before ship.
+- Crowd scenes and urban scenes are the #1 source of accidental trademark content.
+
+### 5.5 Copyright on style
+- "In the style of [living artist]" → legally risky in many jurisdictions.
+- "In the style of [public-domain / historical movement]" → safe.
+- When in doubt, describe the style by its attributes (palette, brushwork, composition) rather than by attribution.
+
+---
+
+## 6. Generation session hygiene (operational discipline)
+
+Every session of AI generation MUST produce:
+
+1. **Session log** — `brands/{brand}/generations/{yyyy-mm-dd}-{session-slug}/`
+2. **Per-asset metadata sidecar** — prompt, seed, model, LoRA(s), references, negative prompts.
+3. **The canonical brand references used** — copied into the session folder (no "floating" references).
+4. **The pass/fail log** per QA check (the 10 items in §4).
+5. **The final selects** — separated from rejects.
+6. **Commercial-use status** — signed off by the brand owner before ship.
+
+Rejects are kept for ≥ 30 days (often useful as training data for the next LoRA) then archived or deleted per the brand's retention policy.
+
+---
+
+## 7. Batch orchestration patterns
+
+| Pattern | When to use | Squad agents involved |
+|---------|------------|----------------------|
+| **Editorial set** (single concept, 5-15 variants) | Campaign hero + secondary shots | brand-creative-engineer drives, brand-identity-designer reviews |
+| **Template grid** (N slots × M variants) | Social media calendar, email series | brand-collateral-designer owns grid definition, brand-creative-engineer generates |
+| **Character series** (same subject, multiple scenes) | Recurring mascot, executive portraits | LoRA-first workflow, brand-identity-designer trains, brand-creative-engineer generates |
+| **Mockup set** (product on N surfaces) | Brand presentation, pitch deck | Photoroom API / SD compositing, brand-collateral-designer owns |
+| **Motion sequence** (logo intro, 3-6 sec) | Brand identity handoff to motion | Hand-authored by brand-motion-vfx, AI used only for assist (upscaling, cleanup) |
+
+---
+
+## 8. Anti-patterns (forbidden)
+
+- **Generating a product shot from pure text-to-image** when a product reference photo exists. Always img2img / compositing when the product is real.
+- **Mixing styles mid-batch without a canon rule allowing it.** If the brand canon says "editorial photography", a single flat illustration mid-batch is a FAIL.
+- **Generating faces resembling real celebrities "for inspiration".** This is a legal landmine, not a creative shortcut.
+- **"Fixing" drift in post.** If a batch drifted, regenerate. Don't color-correct your way out of a style mismatch.
+- **Shipping without session metadata.** If you can't reproduce the asset later, you don't own it.
+- **Using "in the style of [living artist]" prompts for brand-owned deliverables.** Legally and ethically risky.
+- **Treating AI generation as "done" after the first acceptable output.** Every brand asset benefits from N:1 generate-to-ship ratio ≥ 5. Cheap to generate means pick the best, not pick the first.
+- **Free-generating brand identity elements (logomark, wordmark, brand character).** Brand identity is a human craft, AI assists but doesn't drive.
+
+---
+
+## 9. Owner matrix (which squad-brand agent owns what)
+
+| Canon element | Owner agent | Rationale |
+|--------------|------------|-----------|
+| Style lock + reference curation | brand-identity-designer | Owns the visual vocabulary |
+| Lighting + palette lock | brand-creative-engineer | Owns the generation craft |
+| Subject lock + LoRA training | brand-creative-engineer + brand-identity-designer | Joint: one trains, one approves |
+| Composition lock + batch templates | brand-collateral-designer | Owns templated output |
+| Batch QA (§4 checklist) | brand-auditor | Independent review, not the generator |
+| Legal / IP guardrails | brand-legal-ip reference KB + brand-strategist sign-off | Strategic, not creative |
+| Session hygiene + metadata | brand-creative-engineer | Operational discipline |
+| Motion / video canon | brand-motion-vfx | Owns motion craft |
+| Sonic (if multimodal) | brand-sonic-designer | Owns sonic craft — see `sonic-branding-principles.md` |
+
+---
+
+## 10. Reference index
+
+- `squads/squad-brand/knowledge-base/ai-api-capabilities.md` — capability catalog (this squad)
+- `squads/squad-brand/knowledge-base/prompt-engineering-visual.md` — prompt mechanics (this squad)
+- `squads/squad-brand/knowledge-base/brand-legal-ip.md` — legal reference (this squad)
+- `squads/squad-brand/knowledge-base/color-psychology.md` — palette theory (this squad)
+- `squads/squad-brand/knowledge-base/brandbook-structure.md` — where the generated assets land
+- `squads/squad-design/knowledge-base/cross-surface-token-canon.md` — how brand tokens propagate to product surfaces (design squad canon)
+
+## Change log
+
+- **2026-04-12 (v1.0)** — Created as part of squad-brand v3.3 / fw-v2.3. Raises `ai-api-capabilities.md` and `prompt-engineering-visual.md` from "capability notes" to a canonical playbook covering the 5 brand-consistency dimensions with lock protocols, model selection matrix, batch QA checklist, legal/IP guardrails, and agent ownership matrix.

package/squads/squad-brand/squad.yaml

@@ -1,13 +1,13 @@
 # ══════════════════════════════════════════════════════════════════════════════
-# SQUAD MANIFEST — squad-brand v3.2
+# SQUAD MANIFEST — squad-brand v3.3
 # ══════════════════════════════════════════════════════════════════════════════
 # The world-class brand system squad for creating, managing and evolving
 # complete brand systems — from strategic positioning to compiled brandbook.
-# 15 agents | 97 tasks | 4 workflows | 19 KBs | 6 templates | Full coverage
+# 15 agents | 97 tasks | 4 workflows | 30 KBs | 6 templates | Full coverage
 # ══════════════════════════════════════════════════════════════════════════════
 
 name: squad-brand
-version: 3.2.0
+version: 3.3.0
 description: >
   Squad de classe mundial para criacao, gestao e evolucao de sistemas de marca
   completos. 15 agentes especialistas cobrindo 100% de um brand system world-class:
@@ -24,17 +24,31 @@ slashPrefix: SINAPSE
 
 metadata:
   created_at: "2026-03-11"
-  updated_at: "2026-03-13"
+  updated_at: "2026-04-12"
   created_by: "squad-creator (Craft)"
   blueprint: "squads/.designs/squad-brand-design.yaml"
   quality_target: "5.0/5.0 em todas as dimensoes"
   agents_count: 15
   tasks_count: 97
-  knowledge_bases_count: 19
+  knowledge_bases_count: 30
   checklists_count: 3
   templates_count: 6
   workflows_count: 4
-  total_files: 143
+  total_files: 155
+
+changelog:
+  - version: "3.3.0"
+    date: "2026-04-12"
+    story: fw-v2.3
+    epic: EPIC-framework-upgrade-v2
+    highlights:
+      - "Shipped ai-visual-generation-canon.md KB — raises ai-api-capabilities.md + prompt-engineering-visual.md from capability notes to canonical playbook. Covers the 5 brand-consistency dimensions with lock protocols (style/lighting/palette/subject/composition), model selection matrix, batch QA checklist, legal/IP guardrails, session hygiene, batch orchestration patterns, agent ownership matrix."
+      - "Reconciled knowledge_bases_count 19 -> 30 to match disk reality (drift accumulated over multiple minor releases)."
+      - "Minor version bump (additive-only, no breaking changes, no persona edits)."
+    breaking_changes: []
+    deferred_to_follow_up:
+      - "AC2 — brand-to-platform-token-bridge.md (partially superseded: squads/squad-design/knowledge-base/cross-surface-token-canon.md already defines the Nexus<->Atlas contract that includes Meridian upstream; a brand-side mirror doc remains useful for brand-originated handoffs but is lower priority)."
+      - "AC3 — brand-ops-playbook.md (separate canon doc, worthy but scoped out to avoid diluting AI visual canon shipment quality)."
 
 # ══════════════════════════════════════════════════════════════════════════════
 # QUALITY STANDARDS — Regras de Robustez (v3.2)

package/squads/squad-claude/knowledge-base/context-window-optimization.md

@@ -152,7 +152,7 @@ Only loaded when editing API files.
 - Historical context (why decisions were made)
 - Duplicate information
 
-### 5. Document Sharding (BMAD Pattern)
+### 5. Document Sharding
 
 Large documents split into focused shards:
 

package/squads/squad-claude/knowledge-base/swarm-orchestration-patterns.md

@@ -321,9 +321,9 @@ Workers request authorization for dangerous operations via a queue:
 
 ---
 
-## BMAD Method Patterns
+## Spec-Driven Agile-AI Patterns
 
-BMAD (Breakthrough Method for Agile AI-Driven Development) v6 provides patterns SINAPSE can adopt.
+External agile-AI methodologies focused on spec-first development with strong human-in-the-loop checkpoints provide patterns SINAPSE can adopt.
 
 ### Document Sharding
 

package/squads/squad-content/knowledge-base/ai-native-content-loop.md

@@ -0,0 +1,220 @@
+# AI-Native Content Loop — Canon Playbook
+
+> **Status:** Canon (squad-content v2.0)
+> **Owner:** content-orqx (Nexus) with cross-agent collaboration
+> **Created:** 2026-04-12 for Story fw-v2.4
+> **Relationship to existing KBs:** Builds on `content-spine-engine.md`, `ai-leverage-content-framework.md`, `content-measurement-framework.md` and `ai-content-production.md`. This KB is the **canonical end-to-end pipeline** — the other KBs are mechanisms inside phases of this loop.
+
+---
+
+## 1. Why This Exists
+
+In 2026, the content pipeline for any mature SINAPSE user is no longer `brief → writer → editor → publish → measure`. AI assistance is now a load-bearing layer at every step, and the real bottleneck has shifted to **where humans intervene and why**.
+
+This playbook canonizes the 6-phase loop that squad-content runs for every content program, regardless of client, language or platform. It makes the human-in-the-loop decision explicit so that AI-native work does not mean commodity work.
+
+## 2. The 6-Phase Loop
+
+```
+┌───────────────────────────────────────────────────────────────────────┐
+│                                                                       │
+│   1. SIGNAL → 2. AI DRAFT → 3. HUMAN EDIT → 4. PLATFORM ADAPT →       │
+│         ↑                                                  ↓          │
+│         └────────── 6. RETROFEED ←── 5. MEASURE ←──────────┘          │
+│                                                                       │
+└───────────────────────────────────────────────────────────────────────┘
+```
+
+The loop is **closed** on purpose. Phase 6 (Retrofeed) is what turns a one-shot pipeline into a compounding asset.
+
+---
+
+## 3. Phase-by-Phase Specification
+
+### Phase 1 — SIGNAL
+
+| Field | Value |
+|-------|-------|
+| **Lead agent** | `signal-intelligence` (Radar) |
+| **Inputs** | Editorial pillars, client brief, calendar gaps, opportunity windows |
+| **Outputs** | Classified signal (type, temperature, pillar, opportunity window), signal brief (1 paragraph + evidence) |
+| **Tool stack** | Google Trends, X/Twitter API v2, TikTok Research API, Reddit API, Apify actors, Substack RSS, YouTube Data API, Exploding Topics, SparkToro |
+| **Quality gate** | Signal-quality score ≥ 3/5 on (relevance, novelty, evidence, actionability, pillar fit) |
+| **Failure modes** | Low-evidence hype, pillar drift, redundant signal (already covered), stale signal (window closed), competitor-exclusive signal (not adaptable) |
+| **Hands off to** | Phase 2 only if quality gate passes; otherwise returns to source scan or archives signal with notes |
+
+### Phase 2 — AI DRAFT
+
+| Field | Value |
+|-------|-------|
+| **Lead agent** | `content-engineer` (Arc) with structured prompt template |
+| **Inputs** | Signal brief, Espinha Dorsal template (tese/mecanismo/prova/direcao), client voice profile, format target |
+| **Outputs** | First-draft artifact (carousel/post/thread/script/article) grounded on the Espinha Dorsal; draft-confidence score (self-reported by the AI), list of claims needing verification |
+| **Tool stack** | Claude / GPT / in-house prompts (see `ai-content-production.md`), retrieval over client KBs for voice lock, `content-spine-engine.md` as structural contract |
+| **Quality gate** | Draft has: (a) one thesis explicit, (b) mechanism explained, (c) at least 1 proof item, (d) direction/CTA, (e) no invented facts flagged, (f) voice within ±1 on Halvorson Voice & Tone scale |
+| **Failure modes** | Drift from Espinha Dorsal, invented facts ("hallucination"), commodity phrasing, voice drift, format-agnostic output that ignores target platform |
+| **Hands off to** | Phase 3 with structured annotations from the AI (what it is confident about, what it is guessing) |
+
+### Phase 3 — HUMAN EDIT
+
+| Field | Value |
+|-------|-------|
+| **Lead agent** | `editorial-strategist` (North) for high-stakes content; `content-governor` (Index) for compliance/voice enforcement; human editor in the client-side loop |
+| **Inputs** | AI draft + AI annotations, client voice profile, brand guidelines from squad-brand, editorial preferences log |
+| **Outputs** | Edited master version (format-agnostic), edit-reason log (what was changed and why), voice-lock confirmation |
+| **Tool stack** | Markdown diff, `content-quality-checklist.md`, `editorial-governance.md`, preference tracking (`preferences/`) |
+| **Quality gate** | Zero invented facts remaining, voice within ±0.5, E-E-A-T signals preserved or added, claims either removed or cited, hook/pattern-interrupt verified per `viral-potential-scoring.md` |
+| **Failure modes** | Over-editing (losing AI's useful phrasing), under-editing (shipping invented facts), voice-scrubbing (removing brand specificity), missing human-in-the-loop on high-stakes pieces |
+| **Hands off to** | Phase 4 after the master version is signed off |
+
+### Phase 4 — PLATFORM ADAPT
+
+| Field | Value |
+|-------|-------|
+| **Lead agent** | `platform-specialist` (Morph) |
+| **Inputs** | Edited master version, target platform set, `template-contract-system.md` contracts |
+| **Outputs** | Platform-native variants (carousel for IG, thread for X, reel script, LinkedIn post, newsletter section, etc.), each bound to a Template Contract |
+| **Tool stack** | `platform-algorithm-intelligence.md`, `content-atomization-repurposing.md`, `template-contract-system.md`, platform previewers, character counters |
+| **Quality gate** | Each variant respects its Template Contract 100% (character limits, field counts, formatting); algorithmic hooks verified per `verify-algorithmic-hooks` task; no variant loses the core thesis during atomization |
+| **Failure modes** | Contract violations, hook dilution, algorithm mismatch (wrong format for platform), lost thesis in atomization, copy-paste across platforms without nativization |
+| **Hands off to** | Phase 5 via client publishing pipeline (outside squad authority) |
+
+### Phase 5 — MEASURE
+
+| Field | Value |
+|-------|-------|
+| **Lead agent** | `content-analyst` (Lens) |
+| **Inputs** | Published variants across platforms, analytics windows (24h, 7d, 30d), client KPIs |
+| **Outputs** | Performance report per variant + cross-platform rollup; top-performer identification; underperformer diagnosis; retrofeed dataset |
+| **Tool stack** | Parse.ly, Semrush, BuzzSumo, native analytics APIs, `content-scoring-models.md`, `content-measurement-framework.md`, `viral-potential-scoring.md` |
+| **Quality gate** | Metrics collected for every variant (no missing variants), attribution model applied, benchmarks set, insights ranked by actionability |
+| **Failure modes** | Vanity metrics only, no attribution, missing variants, retroactive benchmark inflation, platform-silo analysis with no cross-platform rollup |
+| **Hands off to** | Phase 6 with structured retrofeed payload |
+
+### Phase 6 — RETROFEED
+
+| Field | Value |
+|-------|-------|
+| **Lead agent** | `content-orqx` (Nexus) as coordinator; updates flow to Radar (signal tuning), North (pillar/calendar adjustment), Arc (prompt library update), Morph (template contract refinement), Index (preference log) |
+| **Inputs** | Phase 5 performance dataset, hypotheses that produced each piece, prompt versions used |
+| **Outputs** | Updated signal-quality scoring weights, refined editorial pillars, refreshed AI prompt library, updated template contracts, updated preferences/ log for the client |
+| **Tool stack** | `generate-retrofeed-insights` task, preference log template, `ai-content-production.md` prompt library versioning |
+| **Quality gate** | Every update is traceable to a phase-5 insight (no invented tuning); previous prompt/pillar/contract versions archived, not overwritten |
+| **Failure modes** | Retrofeed ignored ("one-shot pipeline"), overfitting to single-post noise, loss of previous version history, no communication of changes to the other phases |
+| **Hands off to** | Phase 1 of the next cycle (loop closes) |
+
+---
+
+## 4. Human-in-the-Loop Decision Tree
+
+The biggest decision of the AI-native loop is **when AI draft ships with light touch vs. when a human must intervene heavily**. This decision happens at the Phase 2 → Phase 3 transition.
+
+```
+                     ┌──────────────────┐
+                     │  AI draft ready  │
+                     └────────┬─────────┘
+                              │
+                    Has any of these?
+       ┌──────────────────────┴─────────────────────┐
+       │ YES                                        │ NO
+       ▼                                            ▼
+  Human REQUIRED                             Evaluate stakes
+  (Heavy intervention)                        and signals
+       │                                            │
+       │                                            ▼
+       │                               ┌────────────────────┐
+       │                               │ High-stakes piece? │
+       │                               │ (Manifesto,        │
+       │                               │ flagship, PR,      │
+       │                               │ crisis response)   │
+       │                               └────┬───────────┬───┘
+       │                                    │ YES       │ NO
+       │                                    ▼           ▼
+       │                              Human heavy   Awareness
+       │                              review        level high?
+       │                                            (L4/L5
+       │                                             Schwartz)
+       │                                                 │
+       │                                        ┌────────┴───────┐
+       │                                        │ YES            │ NO
+       │                                        ▼                ▼
+       │                                   Human light      AI draft
+       │                                   review only      can ship
+       │                                   (craft polish)   with voice
+       │                                                    check only
+       └── "Human REQUIRED" triggers:
+           - Invented fact flagged by AI itself
+           - Legal/compliance topic (health, finance, legal)
+           - Crisis response or reputation-sensitive
+           - Brand new concept not covered in prior content
+           - Voice drift score >1.0 on client voice-tone scale
+           - Opinion piece or manifesto (voice IS the product)
+```
+
+### Gate Signals Table
+
+| Signal | Weight | Where to check |
+|--------|--------|---------------|
+| Copy/content type (opinion vs informational) | HIGH | Editorial calendar tag |
+| Market awareness level (Schwartz L1-L5) | HIGH | Editorial brief + audience research |
+| Stakes (flagship/manifesto vs daily post) | HIGH | Editorial pillars priority |
+| Brand voice criticality | MEDIUM | Client preferences log |
+| Legal/compliance category | CRITICAL | `validate-content-compliance` task |
+| AI self-reported confidence | MEDIUM | Phase 2 annotations |
+| Novelty of subject matter | MEDIUM | Prior content overlap check |
+
+---
+
+## 5. Tool Stack per Phase (Canonical Stack)
+
+| Phase | Required KBs | Required Tasks | Required Templates |
+|-------|-------------|----------------|-------------------|
+| 1. Signal | `signal-detection-methods.md`, `signal-intelligence-v2.md`, `competitor-analysis-methods.md` | `scan-daily-signals`, `classify-signal-temperature`, `map-signal-to-pillar` | `weekly-signal-briefing-template.md` |
+| 2. AI Draft | `content-spine-engine.md`, `ai-content-production.md`, `narrative-structures-library.md`, `hook-formulas-library.md` | `design-content-spine`, `write-*` family | `content-spine-template.md`, `content-brief-template.md` |
+| 3. Human Edit | `editorial-governance.md`, `content-quality-criteria.md`, `writing-rules-engine.md`, `brazilian-content-context.md` | `validate-content-quality`, `check-brand-consistency` | `content-quality-checklist.md` |
+| 4. Platform Adapt | `template-contract-system.md`, `platform-algorithm-intelligence.md`, `content-atomization-repurposing.md`, `viral-potential-scoring.md` | `adapt-for-*` family, `verify-algorithmic-hooks`, `batch-platform-adaptation` | `template-contract-registry-template.md` |
+| 5. Measure | `content-measurement-framework.md`, `content-scoring-models.md`, `eeat-quality-standards.md` | `analyze-content-performance`, `score-content-quality`, `generate-content-report` | `content-performance-report-template.md` |
+| 6. Retrofeed | `content-strategy-master-reference.md`, `content-governance-principles.md` | `generate-retrofeed-insights`, `update-content-preferences` | `preference-log-template.md` |
+
+---
+
+## 6. Anti-Patterns (Loop Violations)
+
+- **Shipping Phase 2 as final** — skipping human edit for speed destroys the voice differential.
+- **Skipping Phase 6** — without retrofeed the loop becomes a forever-beginner pipeline that never learns from performance.
+- **Phase 4 before Phase 3** — platform adapting an unedited draft wastes work when the master version changes.
+- **Phase 1 as gut feeling** — signals without quality gates produce noisy calendars and pillar drift.
+- **Phase 5 with no attribution** — performance data that can't be traced back to hypotheses cannot retrofeed.
+- **Bypassing the Espinha Dorsal at Phase 2** — AI drafts without thesis/mechanism/proof/direction collapse into commodity content.
+- **Phase 3 voice-scrubbing** — editors removing brand specificity to make content "safer" kills the differentiation.
+- **Cross-platform copy-paste at Phase 4** — ignoring Template Contracts is a contract violation per `template-contract-system.md`.
+
+---
+
+## 7. Success Criteria
+
+The loop is operating at v2 standard when:
+
+1. Every phase has its lead agent identified in the sprint plan
+2. Every Phase 2 output carries AI annotations (confidence + claims to verify)
+3. Every Phase 4 variant passes its Template Contract before publishing
+4. Every cycle closes with a Phase 6 retrofeed payload that updates at least one upstream parameter
+5. No piece ships without passing the Phase 3 human-in-the-loop gate
+6. Cross-phase metrics are cross-referenced (signal type → performance type)
+
+---
+
+## 8. Relationship to Other Canon KBs
+
+- `content-spine-engine.md` — structural contract for Phase 2 (AI draft must be spine-grounded)
+- `template-contract-system.md` — contract enforcement for Phase 4
+- `ai-content-production.md` — prompt library feeding Phase 2
+- `ai-leverage-content-framework.md` — strategic framing for AI-about-AI pieces specifically
+- `signal-intelligence-v2.md` (sibling KB from this story) — source taxonomy for Phase 1
+- `task-ownership-map.md` (sibling KB from this story) — which of the 90 tasks live in which phase
+
+---
+
+## Change Log
+
+- 2026-04-12 — v1.0 authored as part of Story fw-v2.4 (squad-content v2.0 upgrade)