Changelog

Open testing — every scenario, every category, every pass/fail visible at /proof. Including individual prompt-response pairs.

Recommend-first never overrides catalog truth. AI now says 'I don't have an exact match — here are the closest' instead of fabricating.

Up to 3 attempts on transient upstream errors with jittered backoff. Result: AI errors dropped from 22 to ~0 per full test run.

When the AI can't respond after retries, the widget shows a soft chat bubble ('I'm having trouble — try again in a moment?') instead of a technical error toast.

Per-store adaptive quiz. The AI picks each question based on the catalog and prior answers. Quiz ends with real product cards through the full retrieval pipeline.

25 pages: ROI tracking, conversation browser, sync history, memory inspector with edit/pin/delete, in-dashboard test chat, sparkline charts wired to real per-day data.

conversation_outcomes + winning_conversations + store_memory + network_intelligence + product_intelligence. Per-store learning that compounds with every chat.

Replaced fake client-side typewriter with real server-side streaming. Perceived latency cut roughly in half.

Stable per-store context now reused across requests at a fraction of normal cost. Savings passed through into the founding-member pricing.

Multilingual (58 languages), hallucination resistance, multi-turn conversations, edge cases, adversarial safety, and more. Run on every release.