Methodology

Market data

Price, volume and open interest come from Hyperliquid's public market data, read every minute. Nothing else feeds the analysis — no news, no social sentiment, no macro series, no other exchange.

That is a real limitation and it is deliberate. A system that reads only what the market did can be checked against what the market did. One that also reads headlines cannot.

Analysis

Four things are classified independently on every completed hourly bar: trend, momentum, volatility and participation. Each gets one label from a small fixed vocabulary — there is no free-form scoring and no overall verdict, because a single number would hide disagreement between the four.

Every threshold is measured against that asset's own recent history rather than a fixed figure shared across markets. Volatility, for instance, is a percentile of the same asset's own trailing range, so a naturally livelier market is not permanently described as turbulent.

Confirmation

A label changing once is not a change. It has to persist across several readings and clear a materiality threshold before it counts, and each dimension then waits out a cooldown before it can change again.

This is what stops the product narrating noise. Most hours produce nothing at all, which is the correct output for a quiet market.

AI interpretation

Only a confirmed change gets written about. A language model turns the classification into one or two sentences, working from a fixed vocabulary — it is transcribing a result, not forming a view.

It is given no room to analyse: it cannot compute, infer or judge, and every claim it makes must cite the specific data behind it.

Checking

Every generated line is checked mechanically before anyone sees it. The checks confirm that each number appears in the source data, that each claim's evidence actually supports it, that no forward-looking or advisory language has crept in, and that a claim about one dimension is not resting on evidence about a different one.

A line that fails is not shown. It is regenerated once with the reason, and if it fails again a plain deterministic summary is written instead. Nothing that failed the checks is ever published.

What this cannot tell you

• It does not predict. Every sentence is about what already happened.
• It is not advice, a signal, or a recommendation.
• It reads market data only, so it cannot account for news, earnings, regulation or anything else outside the price.
• It can be wrong or incomplete.