Every update, tracked in the open.

The big one — SAM went from a chatbot that talked about predictions to one that actually computes them. Real 10,000-trial Monte Carlo simulations now run in the Worker itself: team sports (NFL, NBA, MLB, NHL, Soccer) pull live season scoring stats from ESPN automatically, and UFC pulls real striking/takedown stats from ufcstats.com — both feed a genuine randomized simulation loop, not an LLM guess (UFC falls back to an estimated power rating only if live fighter stats can't be found). Every pick logs to Airtable automatically with league, matchup, confidence, and result, and SAM reads its own graded track record before each new prediction to self-calibrate and avoid repeating past misses. Predictions are now expanded, naming 2-4 key players per side with the specific stat that matters, not just team-level summaries. Added live ESPN tools for game schedule/venue and injury reports, factored directly into predictions. SAM's reasoning is now visible — types out live, then collapses into a tappable 'Thought for Ns' pill. After every response, tappable follow-up suggestions appear (injuries, schedule, another matchup — deliberately never anything gambling-related, like parlays). Replaced the 'How much do you love America?' quiz gate with a single responsible-use disclaimer and an Enter button. Unified the red/white/blue patriotic theme (flag-stripe bar, eagle branding) across every page, added full 'SAM X.X' version naming everywhere, and general UI polish (lighter blue message bubbles, cleaned-up thinking indicator). Fixed a live bug where multi-step tool calls (e.g. schedule + injury lookups together) could fail with a Gemini 400 error ('missing thought_signature') — the Worker now correctly echoes back Gemini's full response turn on each tool round-trip instead of a stripped-down version, and properly handles multiple tool calls fired in parallel. Added a third backup Gemini API key (GEMINI_API_KEY_BACKUP2) to the fallback chain for extra reliability. Fixed the reasoning display going blank on most real predictions — thinking is now collected across the entire tool-calling exchange (schedule/injury/simulation lookups), not just the final turn, since Gemini often explains itself on the turn where it decides to call a tool and stays quiet on the follow-up turn. When all API keys are exhausted or rate-limited, the site now shows a calm 'Cooling Down' message instead of raw error JSON.

Initial launch — patriot gate screen, dual-key Gemini fallback, stat-based predictions only (no gambling lines, spreads, or odds, ever).