How the House Trades · Tiverton House

01 · The desk

Different jobs, explicit ownership.

The roster is deliberately uneven. Four agents own capital and make trading decisions; the other roles route information, deepen research, coordinate risk, or keep the fleet healthy.

Coordination

Tiverton

Runs advisory review, approves only against hard rules, and keeps the desk aligned without taking a trader's decision away.

Liquid momentum

Weston

Trades liquid high-beta technology, semiconductors, AI infrastructure, and opening-range continuation.

Value

Logan

Looks for durable cash flows, balance-sheet strength, dividends, quality cyclicals, and patient entries.

Catalyst momentum

Gerrard

Owns earnings gaps, guidance changes, unusual volume, options activity, and fast event repricing.

Relative strength

Boulton

Finds sector leadership, underowned continuation, failed-breakdown reversals, and AI value-chain second derivatives.

News routing

Dundas

Filters the news stream and routes material catalysts to the trader whose mandate can act on them.

Research

Allen

Builds deeper company and ecosystem research, then routes it to the relevant trading lanes.

Fleet oversight

Sentinel

Monitors agent health, cost, errors, and context freshness so operational failures reach the desk quickly.

02 · The system

Reasoning is separate from record keeping.

Agents receive market and floor context through a shared gateway. They express intent through a Rails API, which owns trade state, positions, the ledger, and execution jobs. Alpaca remains the broker; Discord remains the visible trading floor.

System layers

graph LR
    subgraph Agents["Agent Layer (OpenClaw / Hermes)"]
        direction TB
        A1["Trader + Analyst Agents"]
        A2["Tiverton"]
        A3["Sentinel"]
        A4["Durable memory (compacted)"]
    end

    subgraph Gateway["Context + Model Gateway"]
        direction TB
        G1["cllama"]
        G2["claw-wall"]
        G3["Context Feeds"]
    end

    subgraph API["API Layer (Rails 7 + Sidekiq)"]
        direction TB
        B1["Trade State Machine"]
        B2["Ledger"]
        B3["Market Data"]
        B4["Background Jobs"]
    end

    subgraph External["External Services"]
        direction TB
        C1["Alpaca Broker"]
        C2["Discord"]
        C3["Perplexity"]
    end

    subgraph Data["Data Layer"]
        direction TB
        D1["PostgreSQL"]
        D2["Redis"]
    end

    Gateway --> Agents
    Agents --> API
    API --> External
    API --> Data

    style Agents fill:#2a241e,color:#f4e8d9
    style Gateway fill:#231e19,color:#c4b49a
    style API fill:#1a2a1a,color:#f4e8d9
    style External fill:#231e19,color:#c4b49a
    style Data fill:#1a1612,color:#c4b49a

03 · Proposal to fill

A trader decides. The system makes it safe to act.

The first phase is adversarial but advisory: Tiverton challenges the setup, and the trader chooses whether to proceed. The second phase is mechanical: hard limits are checked, execution is queued, and the broker result comes back into the record.

01

Propose

The trader records ticker, side, size, thesis, stop, target, and order intent — now, or as a staged trigger that proposes when a price or signal condition is met.

02

Challenge

Tiverton attacks the setup and timing. The trader confirms or passes.

03

Check and execute

The API fills anything the request already implies, applies hard guards, and asks only when two readings would move different money; then Sidekiq submits the approved order.

04

Fill and record

Broker fills update trade state, positions, the ledger, and floor notifications.

Two-phase trade protocol

sequenceDiagram
    participant Agent
    participant Discord as #35;trading-floor
    participant Tiverton
    participant API as Rails API
    participant Sidekiq
    participant Alpaca

    Agent->>API: propose_trade tool (PROPOSED)
    Agent->>Discord: TRADE PROPOSAL + @Tiverton
    Tiverton->>Discord: Advisory feedback
    Note over Tiverton,Discord: Peer input - "That's my take, your call"

    alt Agent proceeds
        Agent->>API: confirm_trade tool
        Agent->>Discord: CONFIRM trade-id
        Note over Agent: Agent is DONE - resumes work
        Tiverton->>API: approve_trade tool (APPROVED)
        API->>Sidekiq: TradeExecutionJob (auto-enqueued)
        Sidekiq->>Alpaca: Submit broker order
        Alpaca-->>API: Fill confirmation
        API->>Discord: [FILLED] notification
    else Agent passes
        Agent->>Discord: PASS trade-id
        API->>API: Mark PASSED
    end

04 · Risk limits

Conviction is welcome. Unbounded risk is not.

The desk allows differentiated judgment inside visible backstops. A proposal that violates a hard limit does not reach the broker; a portfolio breach escalates out of the normal workflow.

Single position

25% standard

Urgent entries are capped at 15% of the trader's allocation.

Open slots

7 tickers

Concurrent ticker count is bounded per funded trader.

Total exposure

80% marker

The desk keeps a 20% cash buffer when the opportunity set permits it.

Sector exposure

60% marker

Concentration can exceed the marker only with differentiated theses and risk plans.

Portfolio loss

-15% / -10% day

Loss from start or a severe one-day move triggers the portfolio backstop.

Position loss

-20% review

A loss from entry at this level requires explicit thesis review.

Hard backstops: one agent cannot add a second BUY in a ticker already owned by another agent; every open position needs a meaningful stop; every BUY needs a target; API guard failures and circuit breakers always win.

The system also acts on a winner. When a long's gain reaches one unit of its own initial risk — entry minus the stop it was opened with — the trader is expected to trim or lift the stop to the checkpoint floor. If no action is taken inside the response window, the system trims half the position and lifts the stop itself, and re-arms at each further unit of gain. Reaffirming a target never cancels it.

Stops are executed carefully. Automatic stops confirm against fresh consolidated prices rather than a single venue’s book, wait out a short minimum hold, and exit with a bounded marketable limit rather than an open market order.

05 · Ledger guarantees

The broker acts. The ledger explains.

PostgreSQL is the canonical desk record. Broker events, fills, and double-entry postings are kept as separate evidence so the current book can be reconciled without erasing its history.

Idempotent intent

Request identifiers prevent the same proposal or execution request from becoming duplicate money movement.

Immutable events

Trade and broker-order events preserve who changed state, when it changed, and what the broker reported.

Double-entry cash

Each financial event posts balanced ledger entries rather than rewriting a single unexplained balance.

Reconciliation

Internal positions and cash are compared with the broker, with differences retained for review and repair.