From a moment

Point instead of describing

Sometimes you have the moment, not the words. You watched a market do something at 14:32 and the question is not “how do I phrase this” but “what was true right there, and how often has THAT happened?”

From a moment answers it in that order. Point at a market and a minute and the panel reads the record as-of that bucket: every feature in the versioned library, the window aggregates leading into it, and which standing rulebook definitions fired on that exact minute. Values that read as unusual at a glance come pre-checked by fixed, documented rules; the rest sit unchecked below them.

Every checked row becomes a proposed chip. Observed values become thresholds by fixed significant-digit rounding, enum states lift verbatim, and every derived value enters the tray marked INFERRED: nothing runs until you confirm or edit each one. What leaves the panel is an ordinary versioned query, not a black box: you can read it, tighten it, share it, and re-run it forever. This is not similarity search. There are no embeddings and no nearest neighbors, and the output is a question you own.

Where you can point from

• The research page itself: open the panel, type a market and a UTC minute, read the store.
• Any archived event page: one click loads that event’s market and start minute into the panel.
• The replay viewer: scrub to the moment that interests you and jump to research at the playhead.

Each entry lands preloaded and shareable: the moment rides in the URL, so a colleague opening your link sees the same bucket read the same way.

What several moments share

The panel also compares moments: add 2 to 8 of them (one market at different times, or different markets entirely) and it renders a per-feature commonality table. The rules are deterministic:

• Numbers agree on the WEAKEST bound every moment satisfies: if three moments show liquidation intensity of 0.52, 0.61 and 0.83, the candidate chip is “at least 0.52”, never an average and never a similarity score.
• Enum states agree only when every moment shows the same non-default state. Three of four is shown as PARTIAL and never becomes a chip: a partial condition would exclude a moment you pointed at.
• Disagreements render as prominently as agreements. When your moments share nothing the library can express, the table says exactly that.

One click turns the agreeing rows into a standard occurrence scan over a window that spans every pointed moment. That makes the comparison self-validating: if the shared setup is real, every moment you pointed at appears in its own result set.

The honest part

You selected these moments BECAUSE they moved. Any intersection computed from them is hypothesis generation conditioned on the outcome, and the tool says so on the surface, not in a footnote. The scan that follows is the corrective: it returns every occurrence of the shared setup, and the outcome columns show the distribution over all of them, not over the handful you picked. Expect the pattern to deflate. That is the product working, not failing: with 33 features, any few moments share something, and a commonality that is common everywhere is uninformative.

FAQ

Is this similarity search? No. Nothing is embedded, ranked, or scored for closeness. Fixed rules turn observed values into explicit thresholds, and the result is a readable query under the same grammar, versioning and reproducibility key as any typed question.

Why does the commonality scan return so much more than my moments? Because that is the point. The scan reports every time the shared setup occurred, so you can see how often it led anywhere at all. If the setup only looks special on your selected moments, the full distribution will say so.

---

Occurrences describe only the recorded markets, dates and eligible buckets reported by the scan. They are observed examples, not a guarantee. Nothing here is trading advice.

> Point at a moment