Model Drift

New: Model Drift Eval

A new Drift tab in the prompt toolbar lets you bless any run, batch, or CSV sweep as a gold baseline, then re-check it against the current model at any time. Each row is scored on four dimensions — lexical similarity, semantic (embedding) similarity, assertion pass-rate delta, and LLM-judge equivalence — with configurable pass/fail gates and a trend view showing drift over time.

• Bless a baseline from any completed run
• Local (on-device ONNX) or API-based embeddings for the semantic score — off by default, opt in under Settings → Workspace → Embeddings
• Gate view with per-dimension thresholds (mean and worst-row)
• Trend view with inline drift history charts

Improved: LLM Verify assertions in batch results

Checklist-based llm_verify assertions now render properly in the batch results table — clear "LLM verify" label, pass threshold badge, percentage score, and the checklist text shown inline instead of being squeezed into the generic assertion format.

New: Help & Community link

Added a Discord link in the sidebar for support, updates, and roadmap discussion.