AI chart analysis

ChartLens reads a chart image and returns four things: the trend structure it can see, two to four support and resistance levels with the visual evidence for each, any clearly-formed classical pattern, and one plan with an entry zone, a stop, targets and the exact condition that would prove it wrong.

Step one: is this even a chart?

Before anything else the model decides whether the image is a legible price chart. A photo of a portfolio balance, an order book, a news article or a meme gets refused with a reason — and refusals never cost a credit. That check exists because the failure mode it prevents is the worst one: confident analysis of something that was never a chart.

Step two: only what is visible

The symbol and timeframe are reported only if they are legibly printed on the image. Indicators are referenced only if they are drawn on it. And if the price axis cannot be read, no numbers are quoted at all — you get levels described in relation to visible structure and an explicit approximate flag.

That last rule costs us something. An analysis with no numbers cannot be scored, so it is excluded from the public record entirely. We would rather publish a smaller honest record than quote prices off an axis nobody could read.

Step three: structure, levels, patterns

Trend comes first — the sequence of highs and lows, ranges, breaks of structure — and "sideways" is a valid answer given most of the time most markets are not trending.

Then two to four levels, each with the evidence for it: how many touches, which wicks, which consolidation shelf. A level with no evidence is not a level, and the requirement to cite evidence is what stops the model drawing lines anywhere it likes.

Patterns are named only from a fixed dictionary of 32 classical formations, with a confidence level and a status of forming, confirmed or failed. Returning no pattern is an encouraged answer.

Step four: one plan, and its invalidation

You get exactly one plan, not a menu. A bias of long, short, or wait; an entry zone; a stop; one or two targets; and — required, never optional — the specific condition that would prove the idea wrong.

The invalidation is the most useful line on the card. It is the difference between a trade you can manage and a position you are hoping about.

What the server checks before you see it

Model output is not trusted on its own. Before an analysis reaches your screen the server nulls every number if the axis was unreadable, recomputes reward-to-risk from the entry midpoint rather than believing the model’s arithmetic, drops any pattern name not in the dictionary, and downgrades a plan to wait if its levels are not internally coherent — a long whose stop sits above its entry is not shown to you with a note, it is not shown to you.