The brain

The Brain is the single coordinator at the centre of Forven. Every strategy that moves through the pipeline and every piece of work an agent produces passes through it. It's hub-and-spoke: agents return their output to the Brain, the Brain decides what happens next, and agents never task each other directly.

The Brain page on a fresh install. In Manual mode the Brain runs no cycles, so almost everything is empty.

Forven is research software. Nothing here predicts results, and nothing here is financial advice.

What the Brain does

The Brain combines deterministic routing and policy checks with its own model calls. It takes strategies from the pipeline, dispatches work to the specialist agents, and drives strategies through the real stages:

quick_screen → gauntlet → paper → live_graduated

At each step the gates decide whether a strategy may advance, and the Brain can't override them. It never trades: there is no AI path to an order.

When it runs. Brain cycles follow the system mode. In Manual — the fresh-install default — the Brain runs no cycles at all. Semi and Auto start them.

The Brain page (/brain)

The page has three tabs.

Overview

Autonomy state, actions, blockers, and memory.

• Current Brain State, with when its memory was last updated and Refresh.
• Tiles: Memory (characters used of 2,000), Active runs, Approvals and Decisions — each links to the matching list.
• Needs Attention — signals worth a look.
• Memory Snapshot — its current working note, with Edit.
• Active Brain-Assigned Runs — pending, running, blocked or failed runs the Brain dispatched; Open Runs jumps to Agents → Runs.
• Recent Brain Activity and Repeated Failures (task types with three or more Brain-assigned failures).

Decisions

What the Brain saw, what it decided, the tasks it spawned, and how each decision turned out. Filter by cycle, action type, strategy or outcome (Success, Failure, Mixed). Click a row for the full decision: the situation, the action taken and its linked runs. Empty until the Brain's first cycle.

Working Notes

Short-term operational notes the Brain carries between cycles — its memory (below), with Edit and a Recent changes history.

Hub-and-spoke

There is deliberately no trading agent. The Execution Trader was retired; positions open and close only through the execution engine and your own controls, so the Brain has no one to delegate trading to. Every result returns to the Brain — one place where decisions are made and audited.

A strategy's journey

1. A strategy is registered — by an agent, by you, or by an AI client over MCP — and lands in quick_screen.
2. If it clears the quick screen, it moves on to the gauntlet: walk-forward, Monte Carlo, parameter jitter, cost stress and the rest.
3. By default the door into paper only stops strategies that are broken or clearly losing. The deflated Sharpe and the other selection-bias checks are recorded there and enforced later, at paper → live, along with its one-shot held-back test.
4. When it passes, it moves to paper automatically — the default promotion mode is auto — or waits in the queue if the paper pool is full. Switch Auto promotions off on Approvals and it waits for your approval instead.
5. In paper, the execution engine trades it on simulated fills from its own signals, behind the risk gates; the Risk Manager watches. No agent places an order. A paper strategy that's clearly losing is retired automatically.
6. Going live is never the Brain's call: it needs the paper → live gate and your typed GO LIVE with a per-asset ceiling.
7. When a strategy fails a gate on its merits it goes to the graveyard, and the Brain queues a post-mortem for the Quant Researcher. A strategy archived as untestable isn't a merit verdict and skips the post-mortem.
8. Each decision, and later its outcome, is recorded on the Decisions tab.

Gates and guardrails

The Brain enforces the gates; it doesn't own them. They're fail-closed — missing evidence means not yet, never pass. Full reference: promotion gates.

The hardest controls sit below the Brain. The kill switch and daily-loss halt act on their own, with no approval and no vote from the Brain. See risk controls.

Brain memory

The Brain keeps one persistent note — its memory — and puts it in its own prompt at the start of every cycle.

• It's capped at 2,000 characters and is never silently truncated: an edit that would exceed the cap is refused, so the Brain has to trim its own notes.
• Every change is recorded with before and after excerpts — the Recent changes list on Working Notes.
• When memory is unchanged between cycles, the prompt cache hits; a rewrite causes an expected miss. Diagnostics shows the hit rate.

Brain memory is the orchestrator's scratchpad. Shared, longer-term knowledge lives in quant skills, where the Brain can propose a rewrite for your approval.

Code changes and approvals

Runs that change code or config — code edits, code fixes, config changes, strategy archives — need your approval when the Brain or an agent starts them. Work you start directly doesn't. The Brain can't approve its own requests.

Its model

The Brain's model is set on Agents → Setup → Agent models, like any agent's, and it only ever calls a model you connected and selected. A keepalive restarts a Brain cycle that times out, so one stuck call can't strand the loop. See models & providers and cost controls.

Forven is a local research lab, not a profit engine. Past results don't predict future outcomes, and nothing here is financial advice.

Related

• Agents — the specialists the Brain dispatches to
• Runs — every run the Brain starts
• Approvals — what waits for you
• Autonomy modes — when the Brain runs at all