Updates May 25

PGrid

• running evals for classification and specs search
  • deepseek v4 pro is worse than gpt-5.4-mini at classification
  • codex gpt-5.4-mini seems pretty good for search grounded specs search prompts, but still not up to gemini flash grounded quality
  • added specs search prompt evals and dual-client consensus mode for review replay evals
• added make command to bootstrap libraries and pnpm on remote vps. this needed a bit of iterating as gpt-5.5 kept getting it wrong
• work in progress on officeworks crawler + laptop support
  • did updates to officeworks crawler and crawler skill
  • enabled officeworks for laptops
• added support for single item classification
  • supports cpu/gpu/ssd/ram. this is to allow specific items to fast track through classification
  • added a cli command for this
• updated laptop variant specs search to include db examples for same model
• updated eval and logging tooling
  • added eval logging updates and made logging more consistent across categories
  • passed reasoning effort through for local codex