A chart-evidence checklist before interpretation.

Identify the document before reading the pattern.

A chart is a document with an origin, creation time, transformation history, and intended audience. Those properties matter before any line, candle, or annotation is interpreted. Record the instrument name and identifier, venue, currency, interval, timezone, visible date range, data vendor, and capture time. If one of these fields is absent, label it unknown rather than filling it from memory.

The title printed in an image is useful evidence, but it is not independent verification. A file can be renamed, a browser tab can be stale, and an annotation can cover the original symbol. The best check is to compare the image with a source that has an appropriate right to publish the relevant market data. If the source cannot be accessed, the interpretation should carry that limitation all the way through the review.

Record transformations as part of identity

Price histories can change when a vendor applies split adjustments, dividend adjustments, currency conversion, session filters, or custom aggregations. Logarithmic and linear axes can make the same move look materially different. Heikin-Ashi or other derived candles are not interchangeable with standard OHLC bars. A useful note states which transformation is visible and avoids comparing levels across incompatible settings.

The first question is not “What does the chart predict?” It is “What document am I actually looking at?”

Keep visible observations apart from derived claims.

An observation should point to something another reviewer can locate: a labeled axis value, a sequence of bars, a volume expansion, a marked moving average, or a repeated reaction near a visible region. A derived claim interprets those marks—for example, calling a sequence a trend, a failed breakout, or a volatility contraction. Both can be useful, but they answer different questions.

Use three columns when taking notes: visible, derived, and unknown. “The last three visible closes are above the blue line” belongs in visible, assuming the closes and line can be inspected. “Momentum has recovered” is derived and needs a defined measure. “The blue line is a 50-session average” remains unknown unless the legend or chart configuration establishes it.

This separation becomes more important when a model reads the image. Models often turn an ambiguous mark into a fluent label. Require every important statement to identify its evidence type, and downgrade or remove a claim when the cited region cannot be found. Eloquence is not a substitute for traceability.

Audit the crop, scale, and time horizon.

A crop directs attention, but it also removes competing explanations. A narrow window may hide an earlier gap, a much larger range, the start of a trend, or a nearby higher-timeframe level. A price-only capture can conceal volume. An intraday view can omit overnight moves. Before interpreting the center of the image, list what each edge prevents you from checking.

Look for axis compression and inconsistent spacing. A truncated price axis can magnify a modest move. An irregular time axis can conceal weekends, market closures, or missing observations. When a drawing extends beyond the captured area, do not infer its unseen anchor. If a level has no readable number, describe its relative location instead of inventing precision.

Compare at least one wider and one narrower view

When permitted data is available, inspect a wider interval to understand regime and a narrower interval to understand how the visible move formed. The goal is not to search until one view confirms a preferred story. It is to discover whether the original story survives reasonable changes of scale. Record the views that weakened the thesis as carefully as the views that supported it.

Corroborate the facts that carry the conclusion.

Not every visible fact needs the same level of verification. Focus first on claims that would change the interpretation: the latest bar, an earnings date, the identity of a gap, an indicator setting, a corporate action, or whether the data is delayed. A clean review names the confirming source, its timestamp, and the scope of what it confirms.

Independence matters. Two websites can display the same upstream feed, and two articles can repeat one original error. Count sources by their underlying evidence, not by browser tabs. Prefer a primary filing for a filing fact, an exchange or appropriately licensed provider for a market-data fact, and the chart configuration for an indicator parameter.

If a decisive fact cannot be corroborated, state the consequence explicitly: “This scenario remains provisional because the capture time is unverified” is more useful than a generic disclaimer at the bottom of the page.

Build alternatives around observable conditions.

A responsible scenario is a conditional map, not a promise. Start with at least two plausible paths and an unresolved path. For each one, write the condition that would make it more relevant, the observation that would weaken it, the time horizon under discussion, and the source limitation that could alter the reading.

Avoid using a confidence percentage unless it comes from a validated, applicable forecasting process with documented calibration. A precise number attached to an informal chart reading can create confidence without evidence. Plain ordinal language—such as “better supported by the currently visible evidence”—is often more honest, provided the supporting evidence is listed.

Do not hide the no-decision outcome

Sometimes the most useful result is that the image cannot distinguish among the alternatives. That may call for a fresher source, a wider interval, confirmation after a scheduled disclosure, or simply no action. The checklist is working when it makes an unsupported decision harder, not when it always produces a dramatic conclusion.

Review multimodal output line by line.

A vision-language model can misread tiny labels, confuse colors, infer an indicator that is not present, or follow instruction-like text embedded in an image. It can also produce a coherent explanation whose individual claims cannot be tied back to the pixels. Review the result as an untrusted draft, even when the prose sounds cautious.

For each material sentence, ask four questions: Can the cited region be located? Is the statement visible or inferred? Does it introduce external knowledge? What observation would show it is wrong? Reject fabricated precision, especially unreadable prices, dates, and indicator parameters. If the original image is too small to audit, obtain a better source instead of asking the model to guess.

Never place credentials, account balances, personal financial records, material nonpublic information, or another person’s confidential data into a chart-analysis workflow. A useful research process minimizes sensitive input before any model or external service is involved.

Use a compact worksheet for the next reviewer.

1. Document identity: instrument, venue, interval, timezone, range, source, capture time, and transformations.
2. Visible observations: numbered statements tied to inspectable regions.
3. Derived interpretations: each linked to one or more observations.
4. Unknowns: missing labels, hidden context, stale-data risk, and unresolved source rights.
5. Corroboration: source, timestamp, and exact claim confirmed.
6. Alternative paths: condition, disconfirming evidence, horizon, and consequence of uncertainty.
7. Review decision: accept as a draft, revise, obtain more evidence, or stop.

Keep the worksheet beside the image rather than replacing it. Future reviewers should be able to see the source document, the reasoning path, and the unresolved questions without reconstructing the author’s memory. That small amount of friction is valuable: it turns an impression into an artifact that can be challenged.