Ask the chart. Get told when it doesn't know.
Follow-up questions answered from the evidence a read actually measured — cited line by line, and refused outright when the measurement was never taken.
Three of the four ask the read to argue against itself. That is deliberate — the questions worth asking a grade are the ones that could lower it.
The answer is assembled from a fixed list
A follow-up never triggers a fresh read of the market. It is handed the analysis that already exists — and nothing else.
The read is already computed
Your question is scoped to it
Numbers are checked against the evidence
Or it refuses, and names what is missing
A refusal is a complete answer
Most chart assistants have no way to answer “I don’t know”, because they have no record of what they measured. This one does, so it can.
- status
- answered
- answer
- prose, 70 words at most
- evidence
- the ids it rests on
Short by design. An answer that needs three paragraphs is usually reasoning past what was measured.
- status
- refused
- missing
- the measurement it does not carry
- evidence
- empty — nothing supports an answer
There is no fallback that invents something. A thesis bullet that fails grounding is swapped for its deterministic twin; an arbitrary question has no twin, so the honest move is the only move.
Both shapes are the schema the answering layer is constrained to, not screenshots of a session. Filling them with invented prose would be the exact failure this page is about.
The same question, asked two ways
“Is there anything here that argues against the long?” — put to a general assistant holding a chart image, and to a read that knows what it measured.
General-purpose AI
UngroundedAnswers every question, because it cannot tell what it knows
"Volume is thinning into the move, which suggests weakening conviction."
Plausible, and unfalsifiable — nothing says whether volume was measured or inferred from the picture.
"Watch the 4H — it may be losing momentum."
A hedge, not a measurement. No higher timeframe was necessarily read at all.
"Overall the setup still looks constructive."
Adjectives standing in for a number, which is the failure mode that costs money.
ORIN
MeasuredAnswers from the evidence list, or declines
Names the factors scoring against the direction, and cites the evidence ids behind each.
Those factors were scored before the question existed, so the answer cannot be shaped by it.
Points at the higher-timeframe leg by its own read — or says none could be loaded.
The legs carry whether they were fetched or aggregated, so the answer can be honest about which.
"This read did not measure that." — when the question goes outside the evidence.
A refusal with the missing measurement named. The one answer a general model never gives.
Neither column is a transcript. The left is the failure mode general models exhibit on chart questions; the right is the behaviour the schema enforces.
What this does not do
Four limits worth knowing before you rely on it.
It will not tell you what to trade.
Directive language is prohibited outright — no buy, no sell, no enter now, no price predictions. It explains what the read supports and where it is thin. The decision is yours, which is the line between decision support and advice.
It cannot see anything you did not have measured.
News, order flow, your broker, the position you are already in — none of it is in the evidence list, so none of it can be in an answer. Ask about any of them and you get a refusal, correctly.
It will not argue you into the trade.
It is instructed not to contradict the thesis already on screen, and where a question presses on something the read got wrong, to say what the read actually supports. It is not a second opinion that can be shopped for a better one.
Questions are counted, because each one is a model call.
3 per read on a free account, 10 on Pro, then a further question spends one of that day's analyses. Nothing about the feature is withheld — the count is the cost of running it.
The grade itself is computed deterministically before any of this runs — see the confluence engine for how the nine factors are weighed. Asking questions never changes a score.
Questions traders actually ask
Every ORIN analysis produces evidence — measured levels, scored factors, a risk plan. Asking the chart means putting a follow-up question to that finished read: why the grade landed where it did, what argues against it, which evidence is carrying it. The model answering is given only what was measured, so it is interpreting a record rather than forming a fresh opinion about the market.
Structurally, no. The answering model is instructed that every number it writes must already appear in the evidence, factors or risk plan it was handed, and it must cite the evidence ids its answer rests on. An answer carrying a number the read never computed is rejected before it reaches the screen rather than shown with a caveat.
It refuses, and names the measurement it is missing. That is a designed outcome rather than an error state: the schema carries an explicit refusal status, so "this read did not measure overnight volume" is a complete answer. Guessing is the single thing the layer is built to prevent, so there is no fallback that invents one.
A general model answers every question, because it has no way to know which numbers it measured and which it inferred — so it produces a plausible level either way. ORIN answers from a fixed evidence list and refuses outside it. The narrower behaviour is the useful one: you can tell the difference between a thing that is known and a thing that sounds known.
No, and it is instructed not to. Directive language — buy, sell, enter now — is prohibited, as are price predictions. It explains what the read supports and where it is weak; the decision stays yours, which is also what keeps ORIN decision-support software rather than advice.
3 per analysis on a free account and 10 on Pro, after which a further question spends one of your ordinary daily analyses. Each question is a real model call, which is why it is counted at all — the allowance is the cost, not a gate on the feature.
Within a read, yes — the questions and answers stay on that analysis. Across reads, no. Each analysis is answered from its own evidence, so a question about one chart can never be answered using numbers measured on a different one.