How Research works

What it is

EdgeDepth Research is a search engine for recorded market microstructure. It answers a precise question: how often did an observable market condition actually occur? You describe the condition in plain words. The engine turns it into an exact, versioned query, scans the order-flow record on 660+ Binance USDT-M perpetual futures markets, and returns every verified occurrence: a count, the denominator it was counted over, per-occurrence evidence, a forward outcome distribution, and a replay link to the exact moment. The market set includes crypto and TradFi perps tracking Nvidia, QQQ, gold, silver and crude oil. They are USDT-settled derivatives, not direct primary-market feeds. The record grows daily.

This is occurrence research over market microstructure: trades, quotes, liquidations, open interest, funding, top-trader and global positioning, returns, volatility and completed candle formations, reduced to a closed library of 33 versioned features. It is not a screener guessing at the present and not a backtest simulating an account. It counts what happened in the record, shows its work, and stops.

What comes back

The occurrence rows are only one part of the result. Every scan also reports:

• the eligible symbol-bucket denominator and exclusions;
• observed feature values at each anchor;
• the distribution of forward returns and excursions across every match, computed before paging and never available as a filter;
• direct replay links to the matching market and minute;
• a reproducibility key that identifies the query and the exact dataset revision scanned.

For a where-only condition, a cohort comparison places the matched outcome distribution beside the baseline distribution from every other eligible predicate-false bucket. It reports both counts and mandatory caveats. It does not manufacture lift, win rates, statistical significance or a causal claim.

Why deterministic matters

Most market research can’t be checked. A number gets posted, the data moves on, and recomputing it is impossible even for its author. EdgeDepth Research is deterministic: the same query over the same record returns byte-identical results, on any account, any day, any number of times.

That works because every result carries a reproducibility key with five parts: the query grammar version, the metadata normalization version, the feature library version, the dataset revision, and the query’s canonical hash. In plain words: which language the question was asked in, which rules cleaned the labels, which frozen code computed every feature, which state of the record was scanned, and which exact question ran. Same key, same bytes. Change any part and the key changes with it.

The key is the receipt. A shared result URL re-runs the scan instead of retelling a summary. A published report pins its result set to its key forever: when the growing record adds occurrences, they append under a new revision, labeled, never merged silently into the pinned set. This is reproducible trading research in the literal sense: strangers can rerun it, diff the bytes, and catch us if they ever differ. (The same discipline runs the replay engine: see determinism.)

From research to replay

Every occurrence row deep-links into the terminal’s deterministic tick-by-tick replay at its anchor moment: a key-moment marker on the timeline, the setup window pre-rolled 30 minutes so the condition forms in front of you. Order book, tape and liquidations print exactly as they did in the permanent event archive. The count says how often; the replay shows what it looked like.

Published reports keep the same split: definition, counts, denominators and full match evidence are visible without an account. Only the replay behind each occurrence is gated (free account).

Browser, public API and MCP

The engine has one contract and three supported ways to use it:

• Browser: open Research, ask in plain words, inspect and confirm the exact chips, run scans and cohort comparisons, point at snapshots, publish reports, arm verified definitions as live alerts, and open matches in replay.
• Public API: use the REST API with an API key for registry-grounded scans, cohorts, snapshots, commonality and reports. Canonical response bytes and engine error codes pass through the HTTP surface.
• MCP for agents: connect to the read-only remote server at https://mcp.edgedepth.com/mcp, or run @edgedepth/research-mcp locally with npx. The tools let an AI agent read the feature registry, interpret prose into a proposed document, run scans and cohorts, inspect a moment, compute base rates and commonality, page results, and fetch published reports. MCP setup instructions cover Codex, Cursor, Claude and local stdio clients.

The MCP surface is deliberately read-only. It can research and retrieve, but it cannot create alerts or mutate your account. Alert arming stays in the browser after you inspect a verified result.

FAQ

Is this a signal service? No. Nothing here tells you to enter, exit, or expect anything. Research counts verified occurrences of conditions you define and shows the evidence. If a result looks actionable to you, that interpretation and its consequences are yours.

What does “byte-identical” actually mean? Re-running a query under the same reproducibility key returns the same bytes. Not similar results: the same result file, checkable with a diff or a hash. If the record grew since, the key’s dataset revision changes and says so.

Where does the record come from? EdgeDepth’s own lossless capture of Binance USDT-M perpetual futures (trades, quotes, liquidations, open interest, funding and positioning on 660+ crypto and TradFi markets), the same capture that powers replay. Features are computed from it by frozen, versioned code, and the record grows daily.

The rest of this topic

This page is the shape of the thing. The details each have their own page: what you can ask (the full feature library, windows and sequences), what it refuses to tell you (the grammar-enforced honesty rules), from a moment (point at a market and a minute instead of describing it), and how live alerts work (a verified definition, standing). For client setup, see Research from the browser, API and MCP.

---

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.

> Run a query on the record