# Full Titan pipeline (200-task scale run, higher concurrency) — Dec 31, 2025
#
# Purpose:
# - Run the COMPLETE pipeline end-to-end at larger scale:
# Megamind (3-pass) → UI gen → build+patch → render → deterministic gates →
# broken-vision gate → vision judge scoring → refinement loop → Fix E/F selection →
# exports + portal.
#
# Notes:
# - No OpenRouter spend. All LLMs on Vertex MaaS. Gemini uses ADC/Vertex even if API keys exist.
# - Concurrency tuned for throughput while staying stable locally.
#
# Run:
# .venv/bin/python -m titan_factory.cli run \
# --max-tasks 200 --run-id full-refine-200 \
# -c config/config-full-refine-200-dec31.yaml
models:
planner:
provider: vertex
model: deepseek-ai/deepseek-v3.2-maas
publishable: true
max_tokens: 2000
temperature: 0.6
# 2 generators × 2 variants = 4 candidates per task
ui_generators:
- provider: vertex
model: moonshotai/kimi-k2-thinking-maas
publishable: true
variants: 2
max_tokens: 65000
temperature: 0.85
- provider: vertex
model: minimaxai/minimax-m2-maas
publishable: true
variants: 2
max_tokens: 65000
temperature: 0.85
# Build-fix patcher (Vertex MaaS)
patcher:
provider: vertex
model: qwen/qwen3-coder-480b-a35b-instruct-maas
publishable: true
max_tokens: 9000
temperature: 0.2
# Vision judge (multimodal). Force ADC/Vertex so we don't accidentally spend via API key billing.
vision_judge:
provider: gemini
model: gemini-3-flash
publishable: false
max_tokens: 1500
temperature: 0.0
force_adc: true
# Megamind per-reasoner models (used only when pipeline.megamind_enabled=true)
megamind_bold:
provider: vertex
model: moonshotai/kimi-k2-thinking-maas
publishable: true
max_tokens: 2500
temperature: 0.9
megamind_minimal:
provider: vertex
model: deepseek-ai/deepseek-v3.2-maas
publishable: true
max_tokens: 2500
temperature: 0.3
megamind_safe:
provider: vertex
model: deepseek-ai/deepseek-v3.2-maas
publishable: true
max_tokens: 2500
temperature: 0.5
megamind_synthesizer:
provider: vertex
model: deepseek-ai/deepseek-v3.2-maas
publishable: true
max_tokens: 3000
temperature: 0.4
pipeline:
# Full scoring + selection mode (required to validate Fix E).
skip_judge: false
vision_score_threshold: 0.0
# Fix E: weighted winner selection (judge score + section creativity)
selection:
judge_weight: 0.6
creativity_weight: 0.4
# Fix F: skip refinement for high-creativity candidates that already pass deterministic gates.
refinement:
skip_for_high_creativity: true
creativity_skip_threshold: 0.7
# Task prompts (preserves creative risk workflow)
task_prompt_pack: mixed
generate_edit_tasks: false
# Increase prompt universe diversity (not max-tasks; this controls generation of prompts/tasks.jsonl)
tasks_per_niche: 10
total_niches: 100
# 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 (axe + Lighthouse)
deterministic_gates_enabled: true
deterministic_gates_enforce: false # observe mode (keep throughput)
axe_gate_enabled: true
axe_fail_impacts: ["critical"]
axe_timeout_ms: 60000
lighthouse_gate_enabled: true
lighthouse_preset: desktop
lighthouse_timeout_ms: 240000
# Required shippable floor: 0 critical axe, LH a11y >= 0.80. (Perf/BP tracked in SQL.)
lighthouse_min_scores:
accessibility: 0.80
performance: 0.00
best_practices: 0.00
seo: 0.00
# Blank/broken page killer (runs in BOTH skip_judge and non-skip_judge after code patch)
broken_vision_gate_enabled: true
broken_vision_gate_min_confidence: 0.85
# Megamind + refinement loop
megamind_enabled: true
refinement_enabled: true
refine_pass2_threshold: 7.0
refine_pass3_threshold: 8.0
max_refine_passes: 2
creative_director_mode: false
# Keep extra expensive loops off at 200 scale.
polish_loop_enabled: false
creativity_refinement_enabled: false
max_fix_rounds: 2
# Timeouts (milliseconds)
model_timeout_ms: 150000
build_timeout_ms: 240000
render_timeout_ms: 120000
budget:
# User request: 10 tasks in flight.
task_concurrency: 10
# Concurrency caps by provider/stage
# NOTE: Vertex can return 429 RESOURCE_EXHAUSTED if concurrency is too high.
# Keep this conservative; task_concurrency still allows overlap while requests queue safely.
concurrency_vertex: 10
concurrency_gemini: 2
concurrency_build: 3
concurrency_render: 2
# Throughput throttle
requests_per_min_vertex: 60
max_total_tasks: null
stop_after_usd: null
export:
holdout_niches: 0
validation_split: 0.0
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"
