AI Research Feed · Alpha Discovery · Judge Engine

Discover alpha with agents.Promote only what survives evidence.

Corrai turns agent collaboration, governed market data, and Alpha Canvas workflows into a live research feed. Every hypothesis, trial, dataset, and backtest is recorded, challenged, and judged before it can move toward production.

Request Early Access Read the Judge Methodology

No one-prompt strategies. No hidden trial history. No promotion without evidence.

AI Research Feed

Registered trial · RUN-8291-A3F2

LIVE

Agent Observation Log

Agent	Observation	State
DATA	Watching feed staleness and canonical coverage	open
ALPHA	Drafting BTC funding-rate reversal hypothesis	queued
RISK	Checking leakage, trial count, and cost assumptions	active

Hypothesis Abstract

Funding compression after volatility shock may mean-revert over 3 bars.

Source: market feed + canonical OHLCV + funding metrics. Trial history remains attached.

Judge docket

Blocked — DSR and walk-forward evidence insufficient

PIT data lineage
Cost-aware next-bar execution
DSR charged for repeated search

Illustrative research docket. Corrai does not provide investment advice or promise profitable strategies.

Alpha discovery system

The feed proposes. The canvas runs. The Judge decides.

Corrai is built around a controlled research loop: agents create candidates, workflows turn them into registered experiments, and evidence gates decide what survives.

Agent collaboration

Research, data, factor, and risk agents work from shared evidence. They can propose and investigate, but they cannot self-promote a strategy.

Research feed
Hypotheses
Tool use
Human gate

Alpha Canvas

Turn ideas into a reproducible DAG: data, factors, signals, labels, backtests, and validation nodes with run IDs and lineage attached.

DAG workflow
Run ID
Lineage
Replay

AI market feed

Keep market, on-chain, and alternative data governed before research starts: staleness, fingerprints, canonicalization, and feed health remain visible.

Scheduled feeds
Canonical data
Fingerprints
Staleness

Multi-agent orchestration

Harness intelligence without handing it the promotion key.

The agent system is designed as a research organization: parallel exploration, shared memory, tool permissions, verification runners, and explicit human approval for consequential actions.

Hypothesis Agent

propose

Turns market observations, prior dead ends, and data anomalies into precise hypotheses that can be registered and tested.

Data Agent

ground

Checks feed health, world/canonical coverage, staleness, lineage, and whether evidence is strong enough to support research.

Workflow Agent

execute

Builds Alpha Canvas DAG proposals and sends verification runs through controlled tools instead of free-form notebooks.

Judge / Risk Agent

challenge

Explains failed gates and risk issues, but trusted verdicts still come from structural Judge evidence, not LLM self-assessment.

From research feed to evidence docket

AI research feed

Agents surface candidate directions from data, memory, and current market structure.

Registered trial

Hypothesis, data version, parameters, and search context are frozen before evaluation.

Judge DAG

Validation nodes challenge the candidate with leakage, DSR/PBO, walk-forward, costs, and review gates.

Evidence docket

The result is not a hype chart. It is a record of what survived, what failed, and why.

Default-skeptical validation

What the Judge checks

Every run faces the same gates. Exceptions are recorded and signed off — never silent.

Purged CV + embargo

Time-series-aware validation. Random K-fold is refused.

DSR + PBO

Sharpe deflated by the number of registered trials. Multiple testing is priced in.

Cost-aware execution

Fees, slippage, next-bar fills — declared on every run.

Robustness

Walk-forward, regime survival, cross-market stability.

Trial lineage

Every experiment is registered. The search history is part of the evidence.

Review & signoff

No researcher self-approval for production.

Evidence operating system

Search history is evidence, not clutter.

A failed run is not thrown away. It becomes part of the research memory that steers the next search pass and taxes repeated discovery attempts.

registered experiments

Trial ledger

Every candidate keeps its hypothesis, parameters, data version, cost model, and search lineage attached.

dead ends + leads

Research memory

Agents can learn what to try next, but memory never weakens the Judge gates or hides failed evidence.

approval boundary

Human gate

Agent work enters the console for review. Production promotion requires evidence and human signoff.

No cherry-picked winners

Autopsies are part of the product.

Corrai is willing to show twelve hypotheses and zero survivors because finding false positives early is the point. That is how a research system earns trust.

Read: Twelve Hypotheses, Zero Survivors

No hiding the number of attempts behind a polished backtest.
No treating raw Sharpe as a verdict.
No letting agent enthusiasm override failed gates.
No promotion without reproducible evidence.

Built for researchers who want a stronger process

Not for signal buyers who want a finished answer.

Individual quants

Ship cleaner research. Keep every result reproducible from its Run ID.

Small quant teams

Standardize R&D and review evidence the same way, every time.

Institutional research

Auditable promotion gates, permissions, and risk signoff.

Boundaries: not designed for tick-level HFT matching replication; assumptions are declared per run.

Build one workflow. Review the evidence. Decide with confidence.

Request early access for reproducible, auditable quantitative research.

Developer · Team · Enterprise — no public pricing yet. Compare plans

Request Early Access