Upload has one dropzone and two modes whose contracts are opposites. The measured read transcribes only the chart’s labels — instrument and timeframe — then grades that market’s real candles, so no output number came from a pixel. The AI read is a model’s own reading of the picture, every field inferred.
The measured read
Only two things are taken off your image: which instrument it is and which timeframe. Everything after that is the ordinary pipeline running on real vendor candles for that market.
The consequence is worth stating precisely, because it is the entire value of this mode: no number in the output was derived from a pixel. The picture was used to answer "which chart is this", and nothing else.
The transcribed price axis is cross-checked against the fetched market. When the two are not plausibly the same chart, a mismatch warning is attached rather than the read being silently served — the usual cause is a screenshot of a different instrument from the one its label claims.
An illegible symbol produces a question, never a guess. A confident wrong ticker is worse than an admission, because the resulting read looks entirely normal.
The AI read
The beta mode asks a model to read the chart itself and return a direction, structure, a thesis, an entry, an invalidation and targets — all of it inferred from the image.
Every field carries a marker recording that it came from the image, and a banner that cannot be dismissed carries the provenance sentence. None of the measured vocabulary is drawn: no grade badge, no confidence figure, and no take, pass or watch controls.
Withholding that vocabulary is the point. Those elements are the visual signature of a measured read, and lending them to an inferred one would make the two indistinguishable at a glance.
The bridge between them
When the ticker resolves, the beta read carries a button that runs the real measured pipeline on the same instrument. Where the two disagree, the measured one is the one with data behind it.
Cost and availability
Both modes are on the free tier. An upload spends one ordinary daily analysis — including when the image turns out to be unreadable, because the model ran either way. The allowance display refreshes after every upload so the ledger is never behind the spend.
Frequently asked
- Which upload mode should I trust with numbers?
- The measured read. Its numbers are computed from real vendor candles, and nothing in its output came from a pixel. The AI read’s numbers were asserted by a model looking at your image, and they never enter your record.
- Why did my upload attach a chart-mismatch warning?
- The price axis transcribed from your image was cross-checked against the fetched market and the two were not plausibly the same chart. The usual cause is a screenshot whose label does not match the instrument actually pictured.
Updated Sep 1, 2026