# BRIEF EJECUCIÓN V25-CONFIRM-ALL · 100% DETALLE · COPY-READY OTRA VENTANA

**Fecha:** 2026-04-26
**Status:** PAUSADO por Carlos · esperando decisión Opción A/B/C/D
**Owner cuando arranque:** Cowork AI / Claude Code
**Trigger:** Carlos dice "arrancá [opción]"

---

## 0. CONTEXTO COMPLETO PARA CONTINUAR DESDE OTRA VENTANA

**Si abrís nueva ventana Claude · pegá este prompt:**

```
Soy Cowork AI continuando V25-CONFIRM-ALL en Zymplo. Leé en este orden:
1. /Users/cr/Zymplo-HQ/CLAUDE.md (reglas inmutables #1-#12)
2. /Users/cr/Zymplo-HQ/01-comando-central/2026-04-26_BRIEF-EJECUCION-V25-COMPLETO.md (este doc · TODO el detalle)
3. /Users/cr/Zymplo-HQ/01-comando-central/2026-04-25_PLAN-V25-CONFIRM-ALL.md (master plan 197 sub-sims)
4. /Users/cr/Zymplo-HQ/sims/V24-97-HALL-02/V25-MASTER-DASHBOARD.html (status visual)
5. /Users/cr/Zymplo-HQ/sims/V24-97-HALL-02/outputs/scenario-metrics.json (estado real queue v1)
6. /Users/cr/Zymplo-HQ/sims/V24-97-HALL-02/outputs/predictions.json (FAILs y métricas)

Carlos eligió Opción [X]. Arrancá según secuencia waves descrita en sección 5 del brief.
Reglas: silencio + commit + dashboard live + safeguards anti-fantasma activos.
NO toques `.claude/settings.json` ni `CLAUDE.md`.
NO arranques sin verificar pre-checks (sección 4 brief).
```

---

## 1. ESTADO ACTUAL (snapshot 2026-04-26)

### Estado queue v1 (baseline preservado · NUNCA borrar)

| Métrica | Valor | Target | Status |
|---|---|---|---|
| Sims base completas | 71/71 | 71 | ✅ |
| Cases procesados | 1,805 | — | baseline |
| Composite predictability v2 | 84.6% | ≥95% | gap -10.4pp |
| Soft PASS rate | 83.6% | ≥97% | gap -13.4pp |
| **Adversarial resistance** | **64.3%** | ≥85% | **CRÍTICO -20.7pp** |
| Reproducibility IAA | 91.7% | ≥90% | ✅ |
| Cobertura geo BR | 27/27 estados | 100% | ✅ |
| Cobertura ofícios MEI | 28 | ≥25 | ✅ |
| Costo gastado queue v1 | $8.39 | — | claude46 + caching -71.8% |

### Sims paralelas
- HALL-01/HALL-02 sims paralelas: ✅ Sin procesos activos detectados

---

## 2. ARCHIVOS GENERADOS EN ESTA SESIÓN (commits)

### Commit `f124460` (2026-04-25 23:52)
- `01-comando-central/2026-04-25_PLAN-V25-CONFIRM-ALL.md` (plan 197 sub-sims)
- `sims/V24-97-HALL-02/V25-MASTER-DASHBOARD.html` (dashboard master)
- `sims/V24-97-HALL-02/outputs/scenario-metrics.json` (recompute)
- `sims/V24-97-HALL-02/outputs/predictions.json` (recompute)
- `sims/V24-97-HALL-02/outputs/CX-BATTERY-REPORT.html` (regen)
- 22 reports nuevos en `outputs/reports/` (PERS + REG)

### Commit `a7fd648` (2026-04-26 00:18)
- `sims/V24-97-HALL-02/LOGO-OPTIONS.html` (7 logos SVG drafted)
- `sims/V25-BRAND-01-LOGO/sim-v25-brand-01-logo.md` (sim spec completa)
- Updates plan + dashboard

### Commit `e379ef7` (2026-04-26 00:35)
- `01-comando-central/2026-04-26_BRIEF-EJECUCION-V25-COMPLETO.md` (este brief)

---

## 3. LAS 4 OPCIONES EJECUTIVAS (Carlos elige)

### Opción A · FULL ULTRA · 197 sub-sims · $6,000 · 59 noches
- Re-corre las 71 V1 con prompts v7 + adversarial embebido + más agentes
- Ejecuta 125 sub-sims nuevas
- Ejecuta V25-BRAND-01-LOGO + tune Wave 0
- Requiere: budget LLM ≥$6,500

### Opción B · Wave 0 ONLY · GRATIS qwen local · $0
- Solo auto-fix de los 36 cases failing en queue v1
- Modelo: qwen2.5:32b LOCAL (sin gasto API externa)
- Tiempo: ~3-4h
- Output: composite v1 mejorado · soft PASS ≥97% target
- NO ejecuta nada nuevo (V25-BRAND-01-LOGO queda pending)

### Opción C · DENTRO DEL BUDGET ACTUAL ($196.63) · ~$130 · MI RECOMENDACIÓN
**Sub-componentes:**
1. Wave 0 auto-fix qwen local (gratis · $0)
2. V25-BRAND-01-LOGO con LLM externo (1200 agents · 8 rounds · ~$80)
3. Tune adversarial top fails (HALL-02 + ESC-01 · ~$50)
- Total: ~$130 USD
- Margen restante: ~$66 USD
- Tiempo: ~10-12h
- Output: logo ratificado + queue v1 100% PASS + adversarial mejorado

### Opción D · HÍBRIDO escalonado · ~$1,800 · necesita budget $2,000
- Wave 0 (gratis) + Opción C ($130) + 25 sims P0 críticas + 125 nuevas core (sin re-correr 71 V1)
- Tiempo: ~25 noches escalonadas
- Output: 100% PASS + logo + 150 sub-sims core ejecutadas

---

## 4. PRE-CHECKS OBLIGATORIOS (regla #4 anti-fantasma · NO NEGOCIABLE)

```bash
# 1. MiroFish container UP
# (verificar via dashboard health endpoint)

# 2. LLM API balance suficiente para budget de la opción
# (verificar antes de cada wave · safety mínimo $500)

# 3. HALL-01/02 sims paralelas terminadas
ps aux | grep -i "mirofish.*orchestrator" | grep -v grep | wc -l

# 4. Disk space ≥ 10GB
df -h /Users/cr | tail -1 | awk '{print "Disk free:",$4}'

# 5. Sims previas commiteadas
cd /Users/cr/Zymplo-HQ && git status --short | wc -l

# 6. Dashboard server UP localhost:8080
curl -s -m 3 http://localhost:8080/ -o /dev/null && echo "Dashboard UP"
```

---

## 5. SECUENCIA DE EJECUCIÓN POR WAVE

### WAVE 0 · AUTO-FIX QUEUE V1 · GRATIS · ~3-4h

**Objetivo:** llevar queue v1 de 84.6% composite → ≥97% soft PASS · 0 FAILs.

**Cases failing identificados (36):**
```
BRAND-02: 009, 010, 024, 056
CONF-01: 003, 026, 034
ESC-01: 006, 013, 017, 032, 043, 050
INTENT-01: 011, 050, 056, 058
LANG-01: 041, 060
MEM-01: 004, 012, 013, 014
NFSE-01: 004
PERS-CAB: 023, 024, 026, 027
PERS-ELE: 023
PERS-ENC: 003
PERS-MAN: 016
PERS-PED: 004, 008
PERS-PIN: 002, 009, 023
```

**Ejecución:**
```bash
cd /Users/cr/Zymplo-HQ/sims/V24-97-HALL-02

# 1. Backup pre-autofix
cp -r outputs/reports outputs/reports.backup-$(date +%Y%m%d-%H%M)

# 2. Iteración 1 con prompt v6 actual
bash auto-fix-rerun.sh 2>&1 | tee outputs/autofix-iter1-$(date +%Y%m%d-%H%M).log

# 3. Recompute metrics
python3 compute-metrics.py

# 4. Si still FAILs → propone v7 + iteración 2
python3 ../../dashboards/standard/auto-fix-fails.py --loop --max-iters 5

# 5. Re-merge resultados
python3 merge-autofix-results.py

# 6. Recompute final
python3 compute-metrics.py

# 7. Generar reporte
python3 report-builder.py

# 8. Validar 0 FAILs
python3 -c "
import json
d=json.load(open('outputs/predictions.json'))
fails=[t for t in d.get('top_fails',[]) if t.get('count',0)>0]
print(f'FAILs restantes: {len(fails)}')
exit(0 if len(fails)==0 else 1)
"
```

**Output esperado:**
- `outputs/reports/*.json` actualizados sin FAILs
- `outputs/predictions.json` · soft PASS ≥97%
- `outputs/scenario-metrics.json` · composite ≥95%
- `outputs/autofix-iter*.log` por iteración
- Commit auto: `fix(V24-97-HALL-02): auto-fix loop · queue v1 100% PASS`

**Costo:** $0 (qwen2.5:32b local · sin API externa)

---

### WAVE 1 · V25-BRAND-01-LOGO · ~6h · ~$80

**Pre-flight:**
```bash
# Crear cases YAML para los 7 logos
cd /Users/cr/Zymplo-HQ/sims/V25-BRAND-01-LOGO
mkdir -p scenarios prompts outputs

# Generar cases.yaml con 1200 MEI agents IBGE-weighted (heredar base de dashboards/standard/scenarios-base/)
cp ../../dashboards/standard/scenarios-base/*.yaml scenarios/

# Configurar caso BRAND-01-LOGO específico
cat > scenarios/BRAND-01-LOGO.yaml << 'EOF'
scenario: BRAND-01-LOGO
description: A/B testing 7 logos · MEI BR · IBGE-weighted
agents: 1200
rounds: 8
ibge_distribution:
  construcao: 0.25
  beleza: 0.20
  alimentacao: 0.18
  servicos_domesticos: 0.12
  auto_moto: 0.08
  tech_reparos: 0.07
  criativo: 0.05
  pet_transporte_outros: 0.05
geo_distribution: all_27_states
age_distribution: 22-65
gender_distribution: 0.45_M_0.55_F
digital_level_mix: low_med_high
adversarial: 100  # V24-CRITIC
social_pressure: 50  # V24-SOCIAL
candidates:
  - id: logo-01
    file: ../V24-97-HALL-02/LOGO-OPTIONS.html#logo-01
    name: Z monogram geométrico
  - id: logo-02
    file: ../V24-97-HALL-02/LOGO-OPTIONS.html#logo-02
    name: Z + chat bubble (WhatsApp DNA)
  - id: logo-03
    file: ../V24-97-HALL-02/LOGO-OPTIONS.html#logo-03
    name: Z infinity flow
  - id: logo-04
    file: ../V24-97-HALL-02/LOGO-OPTIONS.html#logo-04
    name: Z + spark AI
  - id: logo-05
    file: ../V24-97-HALL-02/LOGO-OPTIONS.html#logo-05
    name: Z grid (data DNA)
  - id: logo-06
    file: ../V24-97-HALL-02/LOGO-OPTIONS.html#logo-06
    name: Z circle (continuity)
  - id: logo-07
    file: ../V24-97-HALL-02/LOGO-OPTIONS.html#logo-07
    name: Pure wordmark + accent
evaluation_axes:
  - recall_30s_no_view: 0.25
  - recognizability_16x16_favicon: 0.25
  - emotion_trust_warmth_pro: 0.20
  - differentiation_vs_vambe_stone_mp_contaazul: 0.20
  - app_icon_fit_ios_android: 0.10
threshold_viable: 85
threshold_winner_margin_pp: 5
budget_cap_usd: 100
per_logo_cap_usd: 15
EOF
```

**Ejecución:**
```bash
# Setup workarounds MiroFish bugs (CLAUDE.md regla #5-BIS)
cd /Users/cr/Zymplo-HQ/sims/V25-BRAND-01-LOGO

# Bug #1 workaround: REUSE_GRAPH=True
export REUSE_GRAPH=True
export PROJ_ID=proj_1fff1b00b911
export MIROFISH_ID=mirofish_d3c8c55c2ae94ed0

# Bug #2 workaround: enable_twitter+reddit=True ambos
# Bug #3 workaround: usar gpt-4.1 (no haiku que devuelve markdown JSON)
export LLM_MODEL_NAME=openai/gpt-4.1

# Levantar dashboard server localhost:8081 (no chocar con HALL-02 en :8080)
python3 -m http.server 8081 --directory . &
DASH_PID=$!

# Run pipeline
python3 ../V24-97-HALL-02/orchestrator.py \
  --queue scenarios/BRAND-01-LOGO.yaml \
  --output outputs/ \
  --max-cost 100 \
  --hard-timeout-h 8 \
  --watchdog-burn-rate 5 \
  2>&1 | tee outputs/orchestrator.log

# Monitor live
open http://localhost:8081/dashboard.html
```

**Post-sim · render ganador:**
```bash
# Identificar ganador
WINNER=$(python3 -c "
import json
d=json.load(open('outputs/scenario-metrics.json'))
logos=d['sims']['BRAND-01-LOGO']['candidates_scored']
winner=max(logos, key=lambda x: x['composite'])
print(winner['id'])
")
echo "Ganador: $WINNER"

# Render ganador 2048x2048 + variants
python3 render-winner.py --winner $WINNER --variants mono,dark,light,favicon,app-icon

# Output: outputs/WINNER-RENDERED.png + outputs/variants/
open outputs/WINNER-RENDERED.png
```

**Output esperado:**
- `outputs/scenario-metrics.json` · scores por logo
- `outputs/predictions.json` · top fails + recommendations
- `outputs/VEREDICTO-LOGO.html` · dashboard ganador
- `outputs/WINNER-RENDERED.png` · 2048x2048
- `outputs/variants/` · mono/dark/light/favicon/app-icon
- Commit: `feat(V25-BRAND-01-LOGO): logo ratificado · ganador [ID] composite [X]`

**Costo:** ~$80 USD (1200 agents × 7 logos × 8 rounds · gpt-4.1 con caching)

---

### WAVE 2 · TUNE ADVERSARIAL TOP FAILS · ~$50 · ~4h

**Objetivo:** subir adversarial resistance de 64.3% → ≥85%.

**Sims target:**
- HALL-02 (hallucination) · adversarial top fails · re-correr con prompts v7 + 50 jailbreak attempts adicionales
- ESC-01 (escalation) · adversarial top fails · re-correr

**Ejecución:**
```bash
cd /Users/cr/Zymplo-HQ/sims/V24-97-HALL-02

# Generar cases adversariales nuevos (V24-CRITIC + jailbreak)
python3 ../../dashboards/standard/auto-fix-fails.py --analyze-only --target HALL-02
python3 ../../dashboards/standard/auto-fix-fails.py --analyze-only --target ESC-01

# Re-correr con LLM externo (gpt-4.1 + claude46 hybrid)
export LLM_MODEL_NAME=openai/gpt-4.1
python3 orchestrator.py \
  --queue HALL-02,ESC-01 \
  --adversarial-extra 50 \
  --jailbreak-cases 30 \
  --max-cost 50 \
  2>&1 | tee outputs/wave2-adversarial.log

# Recompute
python3 compute-metrics.py
```

**Output esperado:**
- Adversarial resistance ≥85% confirmado
- Commit: `fix(adversarial): HALL-02+ESC-01 tuned · resistance 85%+`

**Costo:** ~$50 USD

---

### WAVE 3+ (POST AUMENTO BUDGET · si Carlos sube budget LLM)

Si Carlos sube budget LLM a $6,500+, ejecutar las 4 waves originales del plan V25:

| Wave | Bloques | Sub-sims | Cost | Tiempo |
|---|---|---|---|---|
| 3a · Wave 1 P0 original | A+B+D+H+I (productos+safety+pricing+tech+legal) | 64 | $1,800 | 14 noches |
| 3b · Wave 2 P0/P1 | C+E+F+K+L+P+Q | 95 | $2,500 | 14 noches |
| 3c · Wave 3 P1/P2 | G+J+M+N+O+R | 38 | $1,200 | 18 noches |
| 4 · Grand-master | Master rollup cross-bloque + go/no-go decision | 3 | $300 | 9 noches |

**Comando genérico por wave:**
```bash
# Setup nueva sim heredando base inmutable 71 scenarios
cd /Users/cr/Zymplo-HQ/sims
bash ../dashboards/standard/SETUP-NEW-SIM.sh V25-WAVE-1-P0

# Configurar queue específica
cd V25-WAVE-1-P0
# Editar orchestrator.py QUEUE con sims del bloque

# Ejecutar
export LLM_MODEL_NAME=openai/gpt-4.1
export REUSE_GRAPH=True
python3 orchestrator.py 2>&1 | tee outputs/wave-1-p0.log

# Auto-fix loop
python3 ../../dashboards/standard/auto-fix-fails.py --loop

# Compute + report
python3 compute-metrics.py
python3 report-builder.py

# Commit
git add . && git commit -m "feat(V25-WAVE-1-P0): 64 sub-sims · composite [X]%"
```

---

## 6. INVENTARIO COMPLETO 197 SUB-SIMS

### Bloque A · Producto core (19 sub-sims) · P0
1. OB-01-V2 · Onboarding WhatsApp 3-pasos · re-validación
2. OB-01.2-V2 · Catálogo free-text + LLM classification (50 agents · $50)
3. OB-01.3-PRISMA-V2 · Prisma onboarding · re-validación
4. OB-01.5-V2 · Onboarding expansion features · re-validación
5. OB-02 · Onboarding edge cases (no internet · low literacy · low digital)
6. OB-03 · Onboarding por contador (CONTIQ-driven)
7. PERS-01-V2 · Diego/Luana/Raimundo · re-val + arquetipos extendidos
8. PERS-ELE-V2 · Eletricistas · re-val expandida
9. PERS-CAB-V2 · Cabeleireiras · re-val expandida
10. PERS-ENC-V2 · Encanadores · re-val expandida
11. PERS-PED-V2 · Pedreiros · re-val expandida
12. PERS-PIN-V2 · Pintores · re-val expandida
13. PERS-CON-V2 · Confeiteiras · re-val expandida
14. PERS-MAN-V2 · Manicures · re-val expandida
15. PERS-MAR-V2 · Marceneiros · re-val expandida
16. INTENT-01-V2 · Intent detection · re-val
17. MEM-01-V2 · Memory/context retention · re-val
18. MEM-02 · Memory v2 · 20 cases pending del legacy ✅ done
19. TIME-01-V2 · Time/schedule handling · re-val

### Bloque B · Quality / Safety / Adversarial (13 sub-sims) · P0
1. HALL-01-V2 · Hallucination prevention · re-val post-bug6
2. HALL-02-V2 · Hallucination edge cases · expansion
3. HALL-03 · Adversarial hallucination triggers (jailbreak) ✅ done
4. ESC-01-V2 · Escalation a humano · re-val
5. ESC-02 · Escalation v2 · 25 cases pending ✅ done
6. CONF-01-V2 · Confidence scoring · re-val
7. JAIL-01 · Prompt injection / jailbreak resistance
8. ABUSE-01 · Toxic content / abuse handling
9. PRIV-01 · Data leakage prevention
10. CRITIC-V2 · V24-CRITIC adversarial expandido (100 críticos)
11. SOCIAL-V2 · V24-SOCIAL pressure expandido
12. FRAUD-01 · Fraud detection (fake MEIs · scammers)
13. REC-01 · Recovery from errors / graceful degradation

### Bloque C · Geográfico BR (27 sub-sims · 100% nacional) · P0
**Norte (7):** REG-AC · REG-AP · REG-AM · REG-PA · REG-RO · REG-RR · REG-TO
**Nordeste (9):** REG-AL · REG-BA · REG-CE · REG-MA · REG-PB · REG-PE · REG-PI · REG-RN · REG-SE
**Centro-Oeste (4):** REG-DF · REG-GO · REG-MT · REG-MS
**Sudeste (4):** REG-ES · REG-MG · REG-RJ · REG-SP
**Sul (3):** REG-PR · REG-RS · REG-SC

### Bloque D · Pricing & Plans (10 sub-sims) · P0
1. PRICING-01-FREEMIUM · duración óptima 3/7/15/21/30 días
2. PRICING-02-TIERS · estructura 3 planes pagos · features · WTP BRL
3. PRICING-03-ANCHOR · precio ancla óptimo · value perception
4. PRICING-04-MONTHLY-ANNUAL · mensual vs anual · % descuento
5. PRICING-05-VIRAL · mecánicas referido · K-factor + ROI
6. PRICING-06-VS-VAMBE · posicionamiento vs Vambe Chile
7. PRICING-07-CONTIQ · pricing canal contadores (30-200 MEIs)
8. PRICING-08-CHURN · churn por tier · razones · win-back
9. PRICING-09-UPGRADE-PATH · Free→Starter→Pro→Business
10. PRICING-10-UNICORN · MRR/ARR/LTV/CAC · path USD 1B

### Bloque E · GTM / Marketing / Viral (14 sub-sims) · P0
1. GTM-01-CONTIQ · canal contadores 200+
2. GTM-02-PRESS · press BR (Folha/Estadão/Globo/Exame/Valor)
3. GTM-03-INFLUENCERS · influencers MEI/empreendedor
4. GTM-04-CONTENT · content calendar
5. GTM-05-ADS-META · Facebook/Instagram ads
6. GTM-06-ADS-TIKTOK · TikTok ads · viral organic + paid
7. GTM-07-ADS-GOOGLE · Search + YouTube
8. GTM-08-SEO · organic ranking
9. GTM-09-REFER · refer-a-friend mechanics
10. GTM-10-WAVIRAL · WhatsApp grupos · viral spread
11. GTM-11-LANDING · landing page conversion
12. GTM-12-EMAIL · email sequences
13. GTM-13-PUSH · push notifications
14. GTM-14-LAUNCH · launch sequence julio 2026 · Día 1/7/30/90

### Bloque F · Customer Success / Support (10 sub-sims) · P0
1. CX-01-TIER1 · Cowork AI Tier 1 support
2. CX-02-TIER2 · escalation humano Tier 2
3. CX-03-ACTIVATION · activation milestones · TTV
4. CX-04-ENGAGEMENT · engagement loops · daily habits
5. CX-05-STICKINESS · DAU/MAU stickiness · retention
6. CX-06-NUDGES · push reminders · re-engagement
7. CX-07-DISCOVERY · feature discovery flows
8. CX-08-POWER · power user identification · advocates
9. CX-09-NPS · NPS measurement
10. CX-10-WINBACK · churned user win-back

### Bloque G · Competidores / Posicionamiento (8 sub-sims) · P1
1. COMP-01-VAMBE · vs Vambe Chile $18M Monashees
2. COMP-02-STONE · vs Stone BR pagos
3. COMP-03-MP · vs Mercado Pago BR fintech
4. COMP-04-CONTAAZUL · vs Conta Azul
5. COMP-05-OMIE · vs OMIE
6. COMP-06-CRMS · vs HubSpot/Pipedrive
7. COMP-07-MOATS · defensibility · data moat
8. COMP-08-NEW-ENTRANTS · new entrants risk

### Bloque H · Tech / Infra / Scale (9 sub-sims) · P0
1. TECH-01-WABA-LIMITS · WhatsApp Business API rate limits
2. TECH-02-LATENCY · p50/p95/p99 latency targets
3. TECH-03-CONCURRENT · 1K → 100K → 1M users
4. TECH-04-DB-SCALE · database scaling
5. TECH-05-LLM-COST · LLM cost optimization
6. TECH-06-CACHE · cache strategy · prompt cache hit rate
7. TECH-07-FAILOVER · failover scenarios · MTTR
8. TECH-08-DR · disaster recovery · backup strategy
9. TECH-09-COSTS · infra cost prediction

### Bloque I · Compliance / Legal (10 sub-sims) · P0
1. LEG-01-LGPD · LGPD full compliance audit
2. LEG-02-ANPD · ANPD data subject rights
3. LEG-03-BACEN · BACEN compliance (si toca pagos)
4. LEG-04-INPI · INPI trademark BR
5. LEG-05-USPTO · USPTO trademark US
6. LEG-06-DUNS · DUNS Apple/Google App Store
7. LEG-07-STRIPE-ATLAS · Stripe Atlas Delaware
8. LEG-08-CNPJ-BR · CNPJ Zymplo Tecnologia LTDA
9. LEG-09-TOS-PRIVACY · ToS + Privacy Policy
10. LEG-10-DPO · DPO obligations · incident response

### Bloque J · Financiero / Fundraising (7 sub-sims) · P1
1. FIN-01-SAFE · SAFE Notes USD $170K + $150K @ Cap $8M
2. FIN-02-VC-OUTREACH · VC pipeline
3. FIN-03-SERIES-A · path to Series A
4. FIN-04-BURN · burn rate optimization
5. FIN-05-RUNWAY · runway scenarios 12/18/24m
6. FIN-06-HIRING · hiring plan · cost per role
7. FIN-07-CAP-TABLE · cap table evolution

### Bloque K · Roadmap / Features (13 sub-sims) · P0
1. FEAT-01-COBRO · cobranzas via WhatsApp
2. FEAT-02-NF · notas fiscais MEI
3. FEAT-03-AGENDA · agendamentos
4. FEAT-04-CRM-MINI · clientes/CRM mini
5. FEAT-05-ESTOQUE · estoque básico
6. FEAT-06-RELATORIOS · relatórios mensais
7. FEAT-07-DASH-MEI · dashboard MEI personal
8. FEAT-08-INT-PAGOS · integraciones Stone/MP/PIX
9. FEAT-09-AI-COACH · AI proactive coach
10. FEAT-10-VOICE · voice messages handling
11. FEAT-11-IMAGE · image/document handling
12. FEAT-12-MULTILANG · PT-BR primary · ES/EN backlog
13. FEAT-13-OFFLINE · offline-friendly behavior

### Bloque L · North Star / Métricas (10 sub-sims) · P0
1. NS-01-DAU-MAU · target 60%+ stickiness
2. NS-02-RETENTION · D1/D7/D30/D90
3. NS-03-NPS · target ≥70
4. NS-04-ENGAGEMENT · msg/day/user
5. NS-05-CONV · freemium → paid
6. NS-06-CHURN · target <3%/mes
7. NS-07-LTV-CAC · ≥3x
8. NS-08-PAYBACK · ≤6 meses
9. NS-09-MRR · MRR growth MoM
10. NS-10-K-FACTOR · viral coefficient ≥1.2

### Bloque M · Edge / Stress (9 sub-sims) · P1
1. STRESS-01-SURGE · 10x demand
2. STRESS-02-PEAK · peak hours
3. STRESS-03-HOLIDAY · Black Friday/Christmas
4. STRESS-04-OUTAGE · WhatsApp/Meta outage
5. STRESS-05-BUG · bug recovery flow
6. STRESS-06-DATA-LOSS · data loss recovery
7. STRESS-07-FRAUD · bad actors fraud/scam
8. STRESS-08-COMPLAINTS · Reclame Aqui/Twitter
9. STRESS-09-CRISIS · press crisis comms

### Bloque N · Geo expansion futuro (4 sub-sims) · P2
1. GEO-MX-2027 · México 2027
2. GEO-CO-2028 · Colombia 2028
3. GEO-LATAM · LatAm hispanohablante
4. GEO-USA-HISPANIC · USA hispanic

### Bloque O · Team / Cultura (5 sub-sims) · P1
1. TEAM-01-SCALE-PY · 16 ing PY scale
2. TEAM-02-GAPS · founder/role gaps
3. TEAM-03-ADVISORS · advisor needs
4. TEAM-04-CULTURA · cultura Zymplo
5. TEAM-05-COMP · comp bands · equity

### Bloque P · Brand / Identity (7 sub-sims) · P0
1. **V25-BRAND-01-LOGO** · 7 logos · ganador imagen ✨
2. BRAND-02-V2 · brand voice consistency · re-val
3. BRAND-03-PERSONAJES · Diego/Luana/Raimundo equity
4. BRAND-04-SLOGAN · slogan testing · variants
5. BRAND-05-TAGLINE · tagline variations · recall
6. BRAND-06-RECALL · brand recall sin/con prompted
7. BRAND-07-NPS-MARCA · NPS de marca

### Bloque Q · Decisiones / Governance (6 sub-sims) · P0
1. DEC-01-DB-V2 · DB choice re-val
2. DEC-02-COLOR-V2 · color palette re-val post-launch
3. CL-02-SSOT-V2 · SSOT canónico re-val
4. GOV-01-MIROFISH-FIRST · MiroFish-first effectiveness audit
5. GOV-02-ANTI-FANTASMA · anti-fantasma audit · costos prevenidos
6. GOV-03-DEC-NEW · nuevas decisiones que aparezcan

### Bloque R · Backlog & Dashboard (5 sub-sims) · P1
1. BACKLOG-SYNC-V2 · backlog sync re-val
2. DASH-01-LIVE · dashboard live mejoras autónomas
3. DASH-02-SALUD · salud del proyecto
4. DASH-03-OPS · métricas operativas
5. DASH-04-ALERTAS · sistema alertas inteligente

### Sub-sims descubiertas en auditoría 2026-04-26 (10)
1. V25-OB-01.3 · CNPJ async WhatsApp · spec done · sim pending
2. V25-OB-01.5 · Expansion 25 rubros IBGE · plan done · sim pending
3. PERS-01b.6 · Compliance interno · audit pending
4. PERS-01b.7 · Channel integration · spec done · sim pending
5. INF-04 · Feature Flags Platform · spec done · sim pending
6. STREAM-B · Claude46 + Zep memory · spec done · sim pending
7. V24-CRM-POSTMORTEM · CRM-21 valid retroactiva ✅ done
8. CRM-FULL-MVP-IMPL · CRM Meta+TikTok end-to-end · spec done
9. PRESS-BR-PICKUP · Press BR pickup probability
10. OB-01.4-MULTILANG · 8 variantes BR + ES + EN

**TOTAL: 197 sub-sims**

---

## 7. SAFEGUARDS ANTI-FANTASMA (Regla #4 INMUTABLE · NUNCA negociar)

```python
# Hardcoded en TODO pipeline V25
BUDGET_CAP_USD = 6000          # cap total Opción A · ajustar por opción
PER_SIM_CAP_USD = 50           # max por sub-sim individual
STUCK_TIMEOUT_MIN = 15         # kill sim si no progresa 15 min
SAFETY_MIN_BALANCE = 500       # NUNCA bajar de aquí (regla #4)
HARD_TIMEOUT_H = 8             # cap global por wave
PROCESS_MIRROR_INTERVAL = 30   # check ps cada 30s
BURN_RATE_KILL_USD_PER_MIN = 5 # nuclear-stop si gasto >$5/min idle
POST_PIPELINE_GUARD_MIN = 15   # watcher 15min después del fin normal

# Watchdog externo Python paralelo (CLAUDE.md template BUDGET-WATCHDOG-V12.py)
# Auto nuclear-stop al terminar:
#   pkill -9 python (en MiroFish container)
#   container stop
#   verify HTTP down
```

**Si Carlos overrule cualquier safeguard → documentar en memoria como project memory específico (ej: `project_safety_override_20260426_v25.md`) · default vuelve post-sim.**

---

## 8. WORKAROUNDS MIROFISH BUGS (CLAUDE.md regla #5-BIS · obligatorios)

### Bug #1 · `/api/graph/ontology/generate` HTTP 500
**Workaround:** REUSE_GRAPH=True con V23-01 sim
```python
REUSE_GRAPH = True
PROJ_ID = "proj_1fff1b00b911"
MIROFISH_ID = "mirofish_d3c8c55c2ae94ed0"
```

### Bug #2 · enable_twitter+reddit=False → sim fantasma
**Workaround:** SIEMPRE ambos en True
```python
enable_twitter = True
enable_reddit = True
```

### Bug #3 · LLM_MODEL_NAME=anthropic/claude-haiku-4.5 → markdown JSON wrap
**Workaround:** usar gpt-4.1
```bash
export LLM_MODEL_NAME=openai/gpt-4.1
```

### Bug #6 · Oasis signature reporting lag
**Workaround:** sanity checks B/C/D NO abortar cliente-side · solo `log(WARN)` · únicos kills autorizados:
1. `runner_status ∈ {failed, error, cancelled}` (autoritativo)
2. `HARD_TIMEOUT_H` (cap global)

---

## 9. DASHBOARD TRACKING REAL-TIME

**URLs activas:**
- http://localhost:8080/ — Dashboard HALL-02 detalle (queue v1)
- http://localhost:8080/V25-MASTER-DASHBOARD.html — Master V25 (197 sub-sims)
- http://localhost:8080/LOGO-OPTIONS.html — 7 logos preview
- http://localhost:8080/CX-BATTERY-REPORT.html — Reporte CX battery
- http://localhost:8081/dashboard.html — Dashboard sim activa (post arranque V25-BRAND-01-LOGO)
- https://zymplo-backlog-dashboard.vercel.app — Dashboard live público (lag manual)

**JSONs streaming:**
- http://localhost:8080/orchestrator-state.json — estado orchestrator
- http://localhost:8080/predictions.json — predicciones live
- http://localhost:8080/scenario-metrics.json — metrics live
- http://localhost:8080/reproducibility-report.json — IAA + flakes

**Auto-refresh cada 10s en dashboard.html (configurado).**

---

## 10. RECOVERY PROCEDURES

### Si crashea Wave 0 (auto-fix)
```bash
# Restaurar backup pre-autofix
cd /Users/cr/Zymplo-HQ/sims/V24-97-HALL-02
ls outputs/reports.backup-* | head -1 | xargs -I{} cp -r {}/* outputs/reports/
python3 compute-metrics.py
# Re-arrancar Wave 0 desde el principio (idempotent)
bash auto-fix-rerun.sh
```

### Si crashea V25-BRAND-01-LOGO (Wave 1)
```bash
# Verificar burn-rate · si gasto >$5/min idle → nuclear-stop
# Re-ejecutar desde checkpoint último
python3 orchestrator.py --resume-from outputs/orchestrator-state.json
```

### Si MiroFish container down
```bash
# Reiniciar container (verificar logs últimos 50 líneas para diagnosticar)
# Confirmar UP via dashboard health endpoint
```

---

## 11. COMMITS PROTOCOL

**Cada wave termina con commit:**
```bash
cd /Users/cr/Zymplo-HQ
git add sims/ outputs/ 01-comando-central/
git commit -m "$(cat <<EOF
[type](scope): wave [N] · [breve descripción]

[detalle: cuántas sub-sims · composite logrado · costo · tiempo]

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EOF
)"
```

**NO commitear:** `.claude/settings.json` · `CLAUDE.md` (a menos que Carlos lo apruebe explícito).

---

## 12. RESUMEN OPERATIVO PARA CARLOS

### Si elegís Opción A ($6,000)
1. Aumentar budget LLM a $6,500+
2. Decir "arrancá A"
3. Cowork ejecuta Wave 0 → 1 → 2 → 3 → 4 en cascada (~59 noches)
4. Resultado: 197 sub-sims completas · composite ≥97% · launch-ready

### Si elegís Opción B (gratis · solo auto-fix)
1. Decir "arrancá B"
2. Cowork ejecuta Wave 0 (~3-4h · qwen local · $0)
3. Resultado: queue v1 100% PASS

### Si elegís Opción C ($130 · recomendada)
1. Decir "arrancá C"
2. Cowork ejecuta Wave 0 (gratis) + V25-BRAND-01-LOGO ($80) + tune adversarial ($50)
3. Tiempo: ~10-12h
4. Resultado: queue v1 100% PASS + logo ratificado + adversarial 85%+
5. Te muestro PNG ganador del logo automáticamente al terminar

### Si elegís Opción D (escalonada · top-up $1,800)
1. Aumentar budget LLM a $2,000+
2. Decir "arrancá D"
3. Cowork ejecuta Wave 0 + Opción C + 25 P0 críticas + 125 nuevas core
4. Tiempo: ~25 noches
5. Resultado: 150 sub-sims core + logo + 100% PASS

---

## 13. CHECKLIST PRE-ARRANQUE FINAL (cuando Carlos diga "dale")

- [ ] MiroFish container UP
- [ ] LLM API balance suficiente (≥ safety + budget de la opción elegida)
- [ ] HALL-01/02 paralelas confirmadas terminadas
- [ ] Sims previas commiteadas (no work-in-progress)
- [ ] Dashboard server UP en :8080 + :8081 ready
- [ ] Workarounds bugs MiroFish #1, #2, #3, #6 aplicados en pipeline
- [ ] Safeguards anti-fantasma hardcoded
- [ ] Watchdog externo Python paralelo armado
- [ ] Disk space ≥10GB free
- [ ] Carlos OK ejecutivo recibido (texto explícito · no implícito)

---

## 14. CONTACTOS / OWNERSHIP

- **Carlos Rejala (CEO):** decisión final · español · `carlosrejala@zymplo.com`
- **Lucas Doldan (Lead Eng):** fixea MiroFish bugs · `lucasdoldan@zymplo.com`
- **José Carlos Rolón (COO):** legal/compliance · `joserolon@zymplo.com`
- **Cowork AI / Claude Code:** ejecuta sims autónomo · supervisión Carlos

---

**FIN BRIEF · 100% detalle · listo para copiar a otra ventana o para que Cowork ejecute en cascada cuando Carlos dé OK final.**