# Production Config (Creative Risk + Stacked TITAN prompt + Qwen patcher) - Dec 30, 2025
#
# Purpose:
# - Preserve the "creative risk" workflow (promptgen injects Creative risk dial)
# - Avoid generic/default aesthetics by using the stacked TITAN_UI system prompt
# - Keep deterministic gates recorded (axe + Lighthouse) while skipping subjective vision judging
# - Use Vertex MaaS only (NO OpenRouter)
#
# Usage (example 50 tasks):
# .venv/bin/python -m titan_factory.cli run --max-tasks 50 --run-id prod-creativerisk-qwen-dec30 -c config/config-prod-creativerisk-qwen-dec30.yaml
#
# Notes:
# - Color collapse fix: promptgen injects deterministic `brand.mood` + `brand.accent` lines now.
# - This config stacks the built-in UIGEN prompt with prompts/titan_ui_system_long.txt for higher quality.
models:
planner:
provider: vertex
model: deepseek-ai/deepseek-v3.2-maas
publishable: true
max_tokens: 2000
temperature: 0.6
ui_generators:
- provider: vertex
model: moonshotai/kimi-k2-thinking-maas
publishable: true
variants: 1
max_tokens: 65000
temperature: 0.95
- provider: vertex
model: minimaxai/minimax-m2-maas
publishable: true
variants: 1
max_tokens: 65000
temperature: 0.95
patcher:
provider: vertex
model: qwen/qwen3-coder-480b-a35b-instruct-maas
publishable: true
max_tokens: 20000
temperature: 0.2
vision_judge:
provider: gemini
# Use ADC/Vertex so we spend GCP credits, not API-key billing.
# Prefer stable Gemini 3 Flash; judge.py falls back to preview if needed.
model: gemini-3.0-flash
publishable: false
max_tokens: 65000
temperature: 0.7
force_adc: true
pipeline:
# Throughput: no subjective vision judging; deterministic gates still run during render.
skip_judge: true
generate_edit_tasks: false
# More variety than purely niche/local-business prompts.
task_prompt_pack: mixed
# Stacked mega prompt for UIGEN (built-in rules + TITAN_UI long system prompt).
uigen_prompt_variants:
- id: stacked_all
input_mode: both
parts:
- source: builtin
- source: file
path: prompts/titan_ui_system_long.txt
- source: inline
text: |
GLOBAL OVERRIDES (APPLY EVEN IF OTHER PROMPTS CONFLICT):
- Output must be STRICT: <think>...</think> followed by ONE valid JSON object.
- No emojis in any text. If an icon is needed, use simple inline SVG (keep icons minimal).
- Follow the provided brand accent exactly; do not default to violet/purple.
- Follow the Creative risk dial; avoid boring/generic layouts; keep UX clear and build-safe.
# Deterministic gates (record always; enforce only when tuned)
deterministic_gates_enabled: true
deterministic_gates_enforce: false
# Discard visually broken renders (blank pages, error overlays, 404s) even if they pass axe/LH.
broken_vision_gate_enabled: true
broken_vision_gate_min_confidence: 0.85
# Section-level creativity refinement: improve only weak sections, preserve what works.
creativity_refinement_enabled: true
creativity_min_section_score: 0.7
creativity_max_refinement_passes: 2
axe_gate_enabled: true
axe_fail_impacts: ["critical"]
axe_timeout_ms: 60000
lighthouse_gate_enabled: true
lighthouse_preset: desktop
lighthouse_timeout_ms: 240000
lighthouse_min_scores:
performance: 0.35
accessibility: 0.70
best_practices: 0.70
seo: 0.60
# Keep expensive loops off for throughput; creative risk is handled by prompts.
megamind_enabled: false
refinement_enabled: false
creative_director_mode: false
polish_loop_enabled: false
max_fix_rounds: 2
# Timeouts (milliseconds)
model_timeout_ms: 150000
build_timeout_ms: 240000
render_timeout_ms: 120000
budget:
task_concurrency: 5
concurrency_vertex: 10
concurrency_gemini: 1
concurrency_build: 5
concurrency_render: 2
requests_per_min_vertex: 60
max_total_tasks: null
stop_after_usd: null
export:
holdout_niches: 12
validation_split: 0.08
holdout_niche_ids: []
gcs:
bucket: null
prefix: titan-factory-outputs
upload_interval_tasks: 50
vertex:
endpoint_template: "https://{region}-aiplatform.googleapis.com/v1/projects/{project}/locations/{region}/endpoints/openapi/chat/completions"
