# A11Y Production Scaling Config (Qwen3-Coder patcher) - Dec 30, 2025
#
# Purpose:
# - Same as config-a11y-prod-dec29.yaml, but swaps the build-fixer "patcher"
# model to Qwen3-Coder on Vertex MaaS.
#
# Notes:
# - This does NOT enable vision judging; pipeline.skip_judge remains true.
# - Deterministic gates remain observe-only (enforce=false) for throughput.
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: 0.8
- provider: vertex
model: minimaxai/minimax-m2-maas
publishable: true
variants: 1
max_tokens: 65000
temperature: 0.8
patcher:
# Vertex-only: do NOT use OpenRouter (no credits).
provider: vertex
# Verified working via the Vertex OpenAI-compatible endpoint:
# model="qwen/qwen3-coder-480b-a35b-instruct-maas"
model: qwen/qwen3-coder-480b-a35b-instruct-maas
publishable: true
max_tokens: 16000
temperature: 0.2
vision_judge:
provider: gemini
# skip_judge=true, so we explicitly disable the vision model to prevent any Gemini calls.
model: null
publishable: false
max_tokens: 2000
temperature: 0.1
pipeline:
skip_judge: true
deterministic_gates_enabled: true
deterministic_gates_enforce: false
axe_gate_enabled: true
axe_fail_impacts: ["critical"]
axe_timeout_ms: 60000
lighthouse_gate_enabled: true
lighthouse_preset: desktop
# Give Lighthouse extra headroom when multiple renders run concurrently.
lighthouse_timeout_ms: 240000
lighthouse_min_scores:
performance: 0.35
accessibility: 0.70
best_practices: 0.70
seo: 0.60
megamind_enabled: false
refinement_enabled: false
creative_director_mode: false
max_fix_rounds: 2
polish_loop_enabled: false
generate_edit_tasks: false
tasks_per_niche: 7
total_niches: 100
# Slightly higher headroom for provider responses under concurrency.
model_timeout_ms: 150000
build_timeout_ms: 240000
render_timeout_ms: 120000
budget:
# Parallelize tasks to make 500-task runs feasible.
# Keep render concurrency bounded; renderer finds open ports, but higher concurrency increases CPU + port-race risk.
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"
