When it says 70%,
that happens about
70% of the time.

Point your phone at any candlestick chart. Crypto or stocks, any app, any website. You get a probability, an 80% price range, and a confidence score that has been checked against what actually happened.

Coming to theApp Store

Not submitted yet. No email capture, no waitlist — there is nothing to sign up for.

See how confidence is measured

BTC/USDT · 1h · 61 candles read · last close $66,210Hover to see what it forecast42.2% likely up within 1 candle

Fig. 01 — the number everyone else buries

About half the time, we are wrong about direction.

Measured directional accuracy is 51–55%. Market drift is right about 47% of the time. That gap is real, and it is small, and we are not going to dress it up. What this product actually sells is the second number: when it says 70%, that happens about 70% of the time — and that claim we can show you.

Directional accuracy, measured: 51–55%; Against 47% for drift, on data no model or prompt was tuned on. Every figure carries a cluster-bootstrap 95% interval.
Held-out windows: 729; Non-overlapping, across 26 symbols, on a split reserved until the day it was run. Once, as designed.
80% range, observed: 84.9%; At ten candles. The band is supposed to contain the outcome 80% of the time, and it does.

Fig. 02 — how far it reaches

Push it further out and it gets less sure.

One real scan, dragged through the three horizons it forecast. The next candle is the easy one. Ten candles out, the probability has slid 4.5 pointstoward the floor and the range is 3.1× wider. Nothing here is smoothed to make that look better.

Drag, or use the arrow keys

BTC/USDT · 1h · 61 candles read

Likely up

42.2% within 1 candle. With three outcomes the floor is 33.3%, so this call sits 8.9 points above guessing.

80% range: −0.28% … +0.29%; 0.57 points wide
Median move: +0.03%; of similar setups

78/100STRONGsame at every horizon

Top 10% of what this engine produces, about one forecast in ten. These landed right 40.2% against a 34.2% base. A real edge, and still nowhere near a certainty.

Resolved+0.26%inside the 80% rangedirection right

One archived scan · BTC/USDT 1h · 2026-07-22 · resolved against what the market did next. Not a simulation.

Fig. 03 — what you actually do

Point it at a chart

A photo or a screenshot, from any exchange, broker or trading app. Crypto or stocks — if it has candles, it can be read.

Watch what it read

Claude vision digitizes the candle bodies and wicks and calibrates the price axis. The scan below read 61 candles. You see the extraction before you see the forecast, because a bad read should be obvious.

Get a probability, a range, and a score

Three horizons, an 80% price range for each, and one confidence score whose deciles are checked against backtests. Ten to twenty-five seconds.

The last six candles this scan read, as they came off the image.
time	open	high	low	close
07-21 22:00	66385.00	66388.35	66204.00	66304.02
07-21 23:00	66304.02	66576.00	66228.55	66556.16
07-22 00:00	66556.15	66739.89	66525.54	66620.97
07-22 01:00	66620.97	66661.79	66282.65	66433.30
07-22 02:00	66433.29	66532.26	66298.00	66314.12
07-22 03:00	66314.12	66371.27	66176.00	66210.00

Fig. 04 — stated against observed

This is the chart every competitor hides.

Each dot is a bucket of forecasts. Its position left to right is what the model claimed; up and down is what actually happened. A perfectly honest model sits on the diagonal.

The shaded corridor is where an honest model's dots should land 90% of the time given how many scans are in each bucket. Thin buckets get a taller corridor, because we know less there. Scatter inside it is sampling noise. Scatter outside it is a miss, and two of these are.

How this is measured, and what could still be wrong with it

stated probability → ↑ observed frequencyn = 729 windows · 5-candle horizonECE 2.9%7 of 8 bins inside the corridor

Fig. 05 — the states it can return

The weak reads get the same room as the strong ones.

Confidence is muted when it is low. The information never is — same layout, same size, same detail. Hiding a weak read would be the one thing that makes every strong read worthless.

AVAX/USDT1h

86/100STRONG

Forecasts in this band landed right 40.2% of the time, against a 34.2% base rate.

80% range, 10 candles: 7.10 pts
Resolved: 1/1 right

XRP/USDT1h

48/100TYPICAL

Forecasts in this band landed right 34.5% of the time, against a 34.2% base rate.

80% range, 10 candles: 2.01 pts
Resolved: 2/3 right

ETC/USDT1h

15/100WEAK

Forecasts in this band landed right 30.9% of the time, against a 34.2% base rate.

80% range, 10 candles: 3.31 pts
Resolved: 2/3 right

Stocks and ETFs

No directional call on this market.

We tested three ways of predicting which way a stock moves and none beat a coin flip, so we do not show one. How far it might move is a different question, and that we can answer: the 80% range on daily equities contained the outcome 80.7% of the time at ten days, with a mean width of 13.36%.

42 equities · daily · range only · ATR-scaled interval (the arm that shipped)

Fig. 06 — the five inputs

A calibrated 70% and a 51% hit rate are not in tension.

One is a claim about how often the model is right overall. The other is a claim about whether its confidence means anything. A model can be barely better than chance and still tell you honestly which of its calls to trust — and that is exactly what the tiers do.

Signal strength
How far the models lean from a coin flip.
this scan: 0.164
Historical agreement
Whether the model and the historical analogs point the same way.
this scan: 0.600
Analog similarity
How closely past setups resemble this one.
this scan: 0.662
Chart readability
How cleanly the candles came off your image.
this scan: 1.000
Regime familiarity
How often the models have seen markets behaving like this.
this scan: 0.718

The full method, including what would falsify it

Fig. 07 — what it will not do

01
It never tells you to buy, sell, size or time anything.
It describes what a chart looks like and what similar charts did next. That is the whole job.
02
When the price axis is unreadable, it quotes no prices at all.
A designed refusal, not a blank field. You get the shape of the forecast and an explicit note that the numbers are missing.
03
A failed or unreadable scan never uses a credit.
Refunded on any non-200. Retry as many times as you need.
04
No countdown timers, no discount that disappears.
The price is the price. There is no urgency to manufacture, so none is manufactured.

Fig. 08 — the record, misses included

Eight scans, taken as they came.

8 scans sampled at even intervals through the 728 others in the final-gate run, sorted by confidence. Not curated; misses included.

INJ/USDT1h

2026-06-03 · 41 candles

18WEAK

Direction: 1 of 3 right
Inside 80% range: 3 of 3
Moved: −7.04%