# TITAN-4-DESIGN Dataset Factory Configuration
models:
# Stage 1: Planner (generates UI_SPEC)
planner:
provider: vertex
model: deepseek-ai/deepseek-v3.2-maas
publishable: true
max_tokens: 65000
temperature: 0.7
# Stage 2: UI Generators (generate code candidates)
# Note: Thinking models need extra tokens for <think> blocks + code
# INCREASED to 32k for FULL landing pages (hero + features + pricing + FAQ + footer)
# Kimi K2 supports 32k output, MiniMax M2 supports similar
ui_generators:
# Multi-temperature spread for diversity.
#
# NOTE: Candidate count per task = sum(variants for each entry).
# Current spread:
# - Kimi: 4 temps × 1 variant = 4 candidates
# - MiniMax: 4 temps × 1 variant = 4 candidates
# Total = 8 candidates per task (keeps throughput sane while still spanning temps).
# === Kimi K2: conservative → balanced → creative → bold ===
- provider: vertex
model: moonshotai/kimi-k2-thinking-maas
publishable: true
variants: 1
max_tokens: 65000
temperature: 0.70
- provider: vertex
model: moonshotai/kimi-k2-thinking-maas
publishable: true
variants: 1
max_tokens: 65000
temperature: 0.85
- provider: vertex
model: moonshotai/kimi-k2-thinking-maas
publishable: true
variants: 1
max_tokens: 65000
temperature: 1.00
- provider: vertex
model: moonshotai/kimi-k2-thinking-maas
publishable: true
variants: 1
max_tokens: 65000
temperature: 1.15
# === MiniMax M2: include the same 4 temp points ===
- provider: vertex
model: minimaxai/minimax-m2-maas
publishable: true
variants: 1
max_tokens: 65000
temperature: 0.70
- provider: vertex
model: minimaxai/minimax-m2-maas
publishable: true
variants: 1
max_tokens: 65000
temperature: 0.85
- provider: vertex
model: minimaxai/minimax-m2-maas
publishable: true
variants: 1
max_tokens: 65000
temperature: 1.00
- provider: vertex
model: minimaxai/minimax-m2-maas
publishable: true
variants: 1
max_tokens: 65000
temperature: 1.15
# NOTE: GLM-4.7 and MiniMax M2.1 are NOT on Vertex AI MaaS yet
# Only available: Kimi K2, MiniMax M2, DeepSeek V3.2
# Stage 3: Patcher/Fixer (fixes build errors)
patcher:
# Vertex-only: do NOT use OpenRouter (no credits).
provider: vertex
model: deepseek-ai/deepseek-v3.2-maas
publishable: true
max_tokens: 65000
temperature: 0.3
# Vision Judge (scores screenshots - NOT used for training)
# Uses Gemini native API for multimodal vision capabilities
vision_judge:
provider: gemini
# Prefer the stable Gemini 3 Flash name; the code will fall back to preview if not available
# in this Vertex project.
model: gemini-3.0-flash
publishable: false
max_tokens: 65000
temperature: 0.7
force_adc: true
# Refinement Loop Models (quality improvement based on vision feedback)
# These are DIFFERENT from patcher - patcher fixes build errors,
# refiners improve visual/UX quality based on judge feedback.
# Refine Reasoner: Plans targeted fixes from judge feedback
refine_reasoner:
provider: vertex
model: deepseek-ai/deepseek-v3.2-maas
publishable: true
max_tokens: 65000
temperature: 0.5
# Refine Coder: Applies visual/UX fixes (needs strong coding model)
refine_coder:
provider: vertex
model: moonshotai/kimi-k2-thinking-maas
publishable: true
max_tokens: 65000 # Match ui_generators for full rewrites
temperature: 0.7
# Megamind models (avoid homogeneity: do not fall back to planner for every sub-reasoner)
megamind_bold:
provider: vertex
model: moonshotai/kimi-k2-thinking-maas
publishable: true
max_tokens: 65000
temperature: 0.9
megamind_minimal:
provider: vertex
model: minimaxai/minimax-m2-maas
publishable: true
max_tokens: 65000
temperature: 0.8
megamind_safe:
provider: vertex
model: deepseek-ai/deepseek-v3.2-maas
publishable: true
max_tokens: 65000
temperature: 0.7
megamind_synthesizer:
provider: vertex
model: deepseek-ai/deepseek-v3.2-maas
publishable: true
max_tokens: 65000
temperature: 0.7
pipeline:
# Scoring threshold (0-10)
vision_score_threshold: 8.0
# Allow generator temps > 1.0 (hard-clamped in uigen to this cap).
generator_temp_cap: 1.3
# === WINNER SELECTION (Fix E) ===
# Blend judge score (0-10) with section creativity (0-1) so that distinctive
# candidates can win when judge scores are close.
selection:
# Creativity is the north star: pick the most distinctive candidate once shippable is satisfied.
# "creativity_first" ignores weights for ordering, but keep weights for logging/compat.
mode: creativity_first
judge_weight: 0.4
creativity_weight: 0.6
# === CREATIVITY GATE (north star) ===
# Keep in observe mode by default (enforce=false) so throughput isn't destroyed while we tune prompts.
creativity_gate:
enabled: true
min_avg: 0.7
enforce: false
# === MEGAMIND 3-PASS REASONING ===
# When enabled, uses 3 parallel sub-reasoners (bold/minimal/safe) to generate plans,
# then synthesizes them into a unified best-of-3 plan before UI generation.
# This increases plan diversity and quality at the cost of 4x planner API calls.
megamind_enabled: true
megamind_v2_enabled: true
# === STYLE ROUTING + EVOL (prevent monoculture) ===
style_routing_enabled: true
style_gates_enabled: true
style_gates_enforce: false
evol_enabled: true
evol_passes: 2
# === CREATIVE DIRECTOR MODE ===
# When enabled, replaces numeric scoring with qualitative feedback.
# This mode is more generous with creative risk-taking and focuses
# on production readiness rather than aesthetic preferences.
# Refinement is guided by specific feedback (missing_for_production)
# rather than score thresholds.
creative_director_mode: false
# === REFINEMENT LOOP THRESHOLDS ===
# Only used when creative_director_mode is false
refinement_enabled: true
# Don't over-polish: high thresholds tend to normalize creativity.
refine_pass2_threshold: 7.0 # Refine if score < 7.0
refine_pass3_threshold: 8.0 # Refine again if score < 8.0
max_refine_passes: 2
# === REFINEMENT SKIP POLICY (Fix F) ===
# Refinement tends to normalize distinctive layouts. If deterministic gates pass
# AND section creativity is high, skip refinement to preserve creative intent.
refinement:
skip_for_high_creativity: true
creativity_skip_threshold: 0.7
# === BROKEN VISION GATE (blank/error pages) ===
broken_vision_gate_enabled: true
broken_vision_gate_min_confidence: 0.85
# Max fix attempts per candidate
max_fix_rounds: 2
# Whether to attempt polish loop for low-scoring candidates
polish_loop_enabled: true
# Tasks per niche (7 = 6 page types + 1 edit)
tasks_per_niche: 7
# Total niches to generate
total_niches: 100
# Timeouts (milliseconds)
model_timeout_ms: 120000
build_timeout_ms: 240000
render_timeout_ms: 90000
# === DETERMINISTIC QUALITY GATES (AXE + LIGHTHOUSE) ===
# Runs measurable validation before subjective vision judging.
deterministic_gates_enabled: true
# Start in "observe" mode; flip true once thresholds are tuned.
deterministic_gates_enforce: false
axe_gate_enabled: true
axe_fail_impacts: ["critical"]
axe_timeout_ms: 60000
lighthouse_gate_enabled: true
lighthouse_preset: desktop
lighthouse_timeout_ms: 180000
lighthouse_min_scores:
performance: 0.35
accessibility: 0.70
best_practices: 0.70
seo: 0.60
budget:
# Concurrent requests per provider
concurrency_vertex: 5
# Rate limits (requests per minute)
requests_per_min_vertex: 60
# Optional limits
max_total_tasks: null # null = run all tasks
stop_after_usd: null # null = no budget limit
export:
# Number of niches to hold out for generalization testing
holdout_niches: 12
# Validation split (within non-holdout data)
validation_split: 0.08
# Specific niche IDs to hold out (optional, overrides random selection)
holdout_niche_ids: []
# Optional GCS backup
gcs:
bucket: null # Set to bucket name to enable
prefix: titan-factory-outputs
upload_interval_tasks: 50
# Vertex AI configuration
vertex:
endpoint_template: "https://{region}-aiplatform.googleapis.com/v1/projects/{project}/locations/{region}/endpoints/openapi/chat/completions"
