THE PIPELINE
Most AI chart tools are a prompt wrapped around a vision model. Verdict is a measurement pipeline with an analyst at the end of it. Here is every stage, in order.
01 · UPLOAD
Full resolution, untouched
Your chart reaches the engine exactly as you exported it. No compression, no downscaling — pixel extraction needs every pixel, so the pipeline preserves them.
02 · EXTRACTION
Candles become data
Computer vision locates the price axis, reads its labels, and measures every candle — open, high, low, close — from the pixels. User drawings, watermarks, and UI furniture are stripped before anything is measured.
03 · STRUCTURE MATH
Deterministic, checkable
Fair value gaps come from the three-candle definition with fill status checked against every later candle. Equal highs and lows require real swing points within tolerance. Sweeps and BOS/CHoCH come from actual swing breaks. Math doesn't hallucinate.
04 · JUDGMENT
The AI grades from facts
The model receives the computed fact sheet and does what it's good at: bias, premium/discount context, the grade, the confidence score, and a cold written read. In verified mode it cannot claim a structure the math didn't find.
05 · THE CARD
One frame, everything
Grade stamp, bias and confidence, observations you can check against the chart, computed overlays placed at exact coordinates, and watch-for triggers with real price levels that feed the re-verdict loop.
THREE MODES, HONESTLY LABELED
Confidence should be earned, not implied. The badge on each verdict names exactly which pipeline produced it.
You uploaded candle data directly (CSV/JSON). Structure is computed on exact values — the highest-fidelity mode.
You uploaded a screenshot and extraction succeeded. Structure is computed on candles measured from pixels, typically within a tenth of a percent.
Extraction declined this image — too exotic, too compressed, too low-contrast. The AI analyzes visually and the card says so. Treat it as an estimate.
Every miss makes the engine harder to break.
When extraction refuses a chart, that chart becomes an engineering case: the detector gets patched, the chart joins the permanent regression suite, and the engine can never silently fail on that style again. If you hit a chart Verdict won't verify, sending it in is the single most useful thing you can do.
SEND A CHART →