Extract Indicator Values from Thinkorswim Chart Screenshots

Project We run a daily options workflow from a fixed Thinkorswim layout of IWM. Each capture is a single high-resolution screenshot with three time-frame panels. Studies include stacked EMAs, an Ichimoku-style cloud, Keltner channels, a squeeze indicator, money flow, and MACD. Example image (Attached) The current pipeline sends the image to a vision model with a structured prompt. Header prices are often correct. EMA levels, cloud position, and other study values are frequently hallucinated so the output JSON still looks complete. That is not acceptable for trading use. Constraints The chart template cannot be changed. Historical captures (thousands of files) use this same layout and must remain usable. Thinkorswim cannot be modified to print study values. Cropping and other image processing on the existing files is allowed. Deliverable A production reader that accepts one screenshot and returns validated JSON for all three panels. Unreadable fields must be null, with a confidence score. The system must not impute missing values. Required fields per panel Printed last price. Last value of each EMA, identified by color. Price versus cloud (above, below, inside). Cloud bias if visible. Keltner context if visible. Squeeze state. Money-flow direction and any printed value. MACD histogram direction. Optional directional bias (call or put) only after the extracted fields. Acceptance You will receive a labeled sample set. Success is a material reduction in extraction error versus the current prompt-only baseline, measured on a holdout of at least thirty screenshots, with an error log and method write-up. We are not looking for prompt tuning alone. Applicants should describe how they will map plotted series to the price axis and how they will prevent fabricated numbers.