Architecture · 12-phase pipeline

How EMI Works — An autonomous analyst desk, end to end.

EMI is not a feed. EMI is a 12-phase intelligence pipeline that runs every 20 minutes, 24 hours a day. She ingests, resolves, correlates, scores, and only then — if three independent integrity gates agree — dispatches a trade. If the structure isn't there, she puts it on the watchlist and waits.

Start the 7-day trial → Read a live casefile →
  • 12 phases per cycle
  • ~20 min cycle cadence
  • 150 symbols / batch
  • 3 integrity gates before dispatch

Definition · in one paragraph

EMI, in one paragraph

EMI is an autonomous market-intelligence engine that reads 12,000 news articles per day, parses every fresh SEC filing, watches 22 activist hedge funds, monitors options flow and on-chain data, and routes every signal through a graph neural network that decides which ideas have enough convergent evidence to become a trade. The output is a paper-traded book — every entry, stop, target and outcome posted publicly so you can grade her work.

The 12 phases

One cycle, twelve phases.

  1. PHASE 1

    Article collection

    News, macro events, SEC filings and prediction-market signals are pulled from 350+ sources into the evidence spine. Nothing is dropped — full audit trail from byte-one.

  2. PHASE 2

    Entity resolution

    Tickers, executives, funds and beneficiaries are extracted and disambiguated. A 13D citing "the Issuer" gets matched to the right NASDAQ symbol.

  3. PHASE 3

    Evidence enrichment + memory recall

    Each candidate symbol pulls its dossier — past articles, filings, autopsies, theme priors. Fresh news wakes the full memory rather than starting cold.

  4. PHASE 4

    Event accumulation

    Symbols beyond the per-cycle batch (150) flow to the deferred queue and continue as a second pass in the same cycle. Nothing is lost; only sequenced.

  5. PHASE 5A

    Fast parallel correlation filter

    A lightweight pre-filter removes obvious noise — wrong asset class, wrong direction, broken price scale.

  6. PHASE 5B

    GraphSAGE GNN refinement

    The graph neural network — 200,000+ nodes, 50,000+ edges — scores convergence: which evidence corroborates which thesis. Output: an intelligence package with calibrated probability.

  7. PHASE 6

    Strategy + TA + trade economics

    Per-asset rules build the trade plan. Entry, stop, target, position size, reward-to-risk floor. Sub-floor geometry is rejected here, not later.

  8. PHASE 6.5

    Narrative pipeline

    The thesis is rendered into the trader briefing — the few sentences a desk operator can read in 30 seconds and act on.

  9. PHASE 7

    Alert dispatch + integrity gates

    Three independent integrity gates check the trade contract. Missing target? Missing TA snapshot? Stop wider than the asset cap? Dispatch is blocked. Tier C watchlist is the safety net.

  10. PHASE 8

    Outcome tracking + learning

    Every alert is tracked to resolution. Wins, losses, stops, time-to-target — feed into the autopsy engine and recalibrate signal-source weights.

  11. PHASE 9

    Watchlist promotion + conviction memory

    Tier C ideas are diagnosed for missing proof. Targeted enrichment runs. Confirmed setups get re-evaluated. Conviction memory dossiers carry the learning forward.

  12. PHASE 10–12

    Shadow observers, support agents, casefile API

    Trade-eligibility shadow, dual-horizon postgame, structure-readiness, executive briefs — and the casefile visualization layer. Every decision is audit-traceable.

The truth contract

What blocks a bad alert before it ever reaches you.

Tier A and Tier B trade alerts must clear three independent integrity layers. All three must pass. Any one says no, the alert is demoted to Tier C watch — research only.

  • Layer 1 — Alert Integrity Policy. Validates the trade contract: live current price, target price, technical snapshot. No stale data. No scale mismatch.
  • Layer 2 — Distributor null-contract guard. Defensive second line that re-checks the package after policy resolution.
  • Layer 3 — TRADE_CONTRACT_INCOMPLETE invariant. Final backstop. If anything is missing, dispatch is blocked at the wire.

This is the difference between a feed and a desk. A feed posts everything. A desk only fires when the trade is real.

What you get

Every alert ships with the receipts.

  • Entry price
    live
    Sampled at dispatch, not stale.
  • Stop level
    capped
    Per asset class. Forex 2.5%, futures 8%, crypto 12% hard cap.
  • Target ladder
    3
    T1, T2, T3 with staged management.
  • Reward / risk
    ≥ 1.5
    Asset-specific floor. Options ≥ 2.0. No exceptions.

Ready

Read a real alert from end to end.

Every dispatched idea has a public casefile — evidence, decision trace, broker fills, the autopsy if it lost.

Open a casefile → Start free trial →