Research CenterField guide

What an AI trading agent needs before it trades

A proposed trade is only the beginning. Before a decision, an agent needs current market context, a readable signal mix, and historical evidence with its limits attached.

Leroy ResearchSeptember 2, 20268 min read

105bn+

evaluated trades

Rolling archive scale shown on Leroy.

41

registered sensors

The signal state an agent can inspect.

5

match-strength tiers

The strongest available tier is surfaced.

2010

archive begins

Historical coverage starts in 2010.

An AI trading agent can produce a confident answer from a short prompt. Confidence is not context. The useful question is whether the proposal can be checked against a current market frame and the verified record of what followed similar setups.

That is the premise behind Leroy: an evidence layer for agentic trading. It gives the agent something inspectable to carry into its own reasoning without pretending that historical evidence is a prediction or that research software is a broker.

The prompt is not the evidence.

A model can summarize a chart, name a familiar pattern, or propose a trade without ever testing the setup. An evidence layer makes the proposal specific: what ticker, which side, what context, what match, and which outcomes are actually available.

A question becomes useful when every layer travels with the answer.

01

Current market context

The frame around the proposal: price, volume, benchmark relationship, session state, and other available conditions.

02

Strategy signals

The registered strategies evaluated against the setup, including agreement, disagreement, and neutral signals.

03

Historical matching

The strongest available match tier, with its basis and row count kept visible instead of silently pooled.

04

Observed Outcomes

Completed forward windows that show what followed similar setups. They are evidence about the past, not a forecast.

The path is deterministic enough to inspect. The result remains research, not an order.

Context turns a trade idea into a question we can test.

“Buy AMZN” is a direction. “Buy AMZN in this session, at this price, with this volume and benchmark relationship” is a research request. Leroy keeps that frame visible so an agent can decide whether the evidence applies to the question it actually asked.

What travels with the answer

Ticker and side

AMZN · buy

Optional setup

entry · stop · target

Outcome windows

15m · 30m · 60m

Boundary

research only

1

current frame

41

registered sensors

5

match tiers

3

outcome windows

These are response fields and archive dimensions, not a performance promise.

Agreement is a signal. Matching is the test.

Strategy agreement can make a setup interesting, but it does not make it proven. The next step is to search the prepared archive for the strongest available match and return the basis alongside the result. Exact and comparable observations stay separate; lower tiers are not quietly pooled to make a sample look larger.

5/5

Exact Setup

4/5

Cross-Ticker Setup

3/5

Signal Pattern

2/5

Contextual Analogue

1/5

Direction-Only Baseline

0/5

No Comparable Evidence

A no-match response is a valid research result. It is better to expose an evidence gap than to fill it with an unrelated baseline.

Outcomes should arrive with their limits attached.

A response can report what happened after similar setups at 15, 30, or 60 minutes. That is useful evidence for an agent to weigh. It is not a forecast, a guarantee, or a reason to remove human judgment. The boundary belongs in the response itself, not buried in a footer.

15m

Near-term outcome window.

30m

Intermediate outcome window.

60m

Longer observed window.

The answer should be readable by a person and inspectable by an agent.

A useful response is not a wall of prose. It is a compact record with stable fields: status, match strength, row count, outcomes, assumptions, and the research-only boundary. The same evidence can render in a terminal, travel through MCP, or open in the Research Center without changing its meaning.

{
  "request": {
    "symbol": "AMZN",
    "side": "buy",
    "horizon": [15, 30, 60]
  },
  "market_memory": {
    "match_strength": {
      "score": 4,
      "tier": "cross_ticker_exact_state"
    },
    "row_count": 184,
    "outcomes": { "15": {}, "30": {}, "60": {} }
  },
  "research_boundary": {
    "research_only": true,
    "live_order": false
  }
}

The shape is intentionally boring. Boring fields are easier to validate, compare, and pass between systems.

From market memory to a model.

Market Memory is the foundation: current context, strategy signals, and historical outcomes. We are developing a paper model on top of that foundation—testing how those signals combine, exposing its decision threshold, and keeping every result inspectable. The model is a research subject, not a black box we ask an agent to trust.

Open Research Center

Bring Leroy into the agent workspace you already use.

Connect through MCP, or install the CLI beside your existing workflow. Both paths return the same current context, historical matching, outcomes, and explicit research boundary.

Method note: Leroy’s archive and response contract change as the research system is updated. Figures above are rounded or structural examples unless explicitly described as response fields. Historical and simulated outputs are not actual customer returns, forecasts, or investment advice.

Read the live contract in the MCP instructions or start with the CLI guide.