Different candle-chart providers (PocketOption, TradingView, our own Polygon-driven chart, …) paint different candles for the same underlying at the same moment — different wicks, slightly offset OHLC, different "completed" timestamps. That disagreement is the data we want to study. This lab captures multiple providers simultaneously, runs the LLM verdict pipeline on each one independently, and persists the inter-provider divergence as a time series. It makes zero predictions and fires no wagers.

1. ARM SESSION — locks one symbol + 2–4 provider source labels (e.g. pocketoption tradingview). A unique partial index makes it physically impossible to have two active divergence sessions on the same symbol.
2. CAPTURE EACH SOURCE — one Capture Screen button per source. Each binds a persistent, off-DOM <video> element and shows its own freshness pill updating in real time. If any source goes amber/red you re-share it — the lab will not run a tick with a stale source.
3. SYNCHRONIZED TICK — every N seconds, the page atomically snapshots every live source in the same browser animation frame and POSTs all N images in a single request. The server overwrites the client's tick_ts_ms with its own wall-clock to defeat client clock drift.
4. PER-SOURCE LLM VERDICT — server runs the engine + LLM on each image in parallel (thread pool). We capture each provider's direction, confidence, and pattern label independently.
5. DIVERGENCE INDEX + κ — we compute three independent metrics per tick (d_direction, d_confidence, d_pattern) and a session-cumulative Cohen's / Fleiss' κ over the direction labels. κ corrects for chance agreement, which is what makes the result publishable rather than a toy.

Sanity check: point two browser tabs at the SAME chart and capture both — κ should converge to ≈ 1.0 and d_direction to ≈ 0. If it doesn't, the pipeline is broken before any real provider comparison is meaningful.