πŸ›οΈ Trading Brain by ARIA Evidence-linked Β· outcome-proven Β· human-gated capital
About Trading Brain

Built for professional trust, not black-box automation.

Trading Brain by ARIA is your Personal AI Trading COO: evidence-linked, outcome-proven, user-operable. It preserves sources, claims, shadow decisions, no-actions, outcomes, scoreboards, operations, support, and review posture so a professional user can inspect what the system knows, what it still cannot prove, and where human authority is required.

How to read this page

In the order that matters to you: what you get, how the workflow produces it, then the trust methodology and architecture underneath.

1
User outcome
See what deserves attention today, why something can wait, and where evidence is still missing β€” before any capital decision.
2
Product workflow
Sources β†’ claims β†’ shadow decisions and no-actions β†’ outcomes β†’ evidence cards β†’ your review. Each step has an auditable artifact.
3
Trust & architecture
The North Star V4 readiness model and the five-layer architecture explain why the proof holds β€” and why capital authority stays human-only.

The five-layer architecture

Each layer has a strict role. They do not collapse into each other. This separation is what makes "AI operates the company; AI cannot authorize capital" enforceable rather than aspirational.

L1
trading-agent
Runtime truth: account, positions, orders, risk, trades. The only source of truth for live state.
L2
Hermes
Orchestrator. Routes user requests, runs workflows, writes Obsidian notes. Never holds broker authority.
L3
TrustMem
Durable lessons only. Repeated failure patterns, agent reliability, recurring regimes. Never live state.
L4
Obsidian
Operating notebook. Research, playbooks, decisions, reviews, scorecards, CEO briefs. Never live broker truth.
L5
Synapse
Optional multi-agent DAG for complex analysis (signal review, portfolio review, weekly review). Not required for runtime queries.

What this means for you: Hermes coordinates the interaction, Trading Brain structures the evidence and review, TrustMem preserves why you watched and why you waited, and Synapse resurfaces unresolved review work β€” so nothing falls through the cracks. None of these layers can approve a trade or move live capital; that authority stays human-only.

The product trust model

Trading Brain is what a trading intelligence product looks like when professional trust surfaces are first-class. The system does not only generate candidates; it also exposes data freshness, credited sources, evidence gaps, outcome feedback, support issues, and known operational degradation.

The system can run research, monitor topics, collect evidence, generate candidates, write shadow decisions, replay portfolios, label outcomes, update scoreboards, generate evolution candidates, build promotion review packets, write CEO briefs, and produce code through Codex or Claude Code β€” all without your involvement. What it cannot do: submit a broker order, draft an execution, approve a trade, promote a strategy into production, mutate config, or expand risk budget.

That boundary is enforced in code, in tests, in CI gates, in post-deploy smoke checks, and in the Trust & Operations surface. It is not a tagline.

The operating principles behind the product

Trading Brain's design choices come from a small set of principles. Every product decision is checked against whether it improves proof, operability, and boundary discipline.

πŸ”¬

Outcome proof before confidence

Process quality is necessary; measured trading and replay results are the proof; feedback into the next cycle is the flywheel. A high-confidence answer with weak provenance or no measured outcome is treated as unproven.

πŸ“

Evidence cards before promotion

Promotion review must expose credited sources, corroborating evidence, evidence gaps, scoreboard support, and why the system cannot auto-promote.

πŸ”’

Capital authority is human-only

The system can research, replay, score, and propose. It cannot submit, approve, promote, or mutate runtime live state. LIVE-mode submissions require typed confirmation. Halt is one click and reversible.

🧾

Every decision has an evidence chain

From source data to claim, candidate, shadow decision, outcome, scoreboard, and promotion packet, each step is auditable. "AI thinks this is a buy" without an evidence chain is treated as noise, not a signal.

🚧

Layers do not collapse

Runtime truth, durable lessons, operating notebook, and orchestration are separate by design. Live state never lives in Obsidian. Lessons never get conflated with today's PnL. The architecture enforces the boundary.

πŸ”

Operations are product signals

Incidents, support tickets, notification traceability, freshness gaps, and safe repair evidence are part of maturity. A system that cannot explain degradation is not product-ready.

The actual evolution loop

This is the loop ARIA runs every day. Each step has artifacts. Each artifact is auditable. You decide where to interrupt, override, or approve.

Stage 1 Β· Discover

Topic signals β†’ research candidates

Scheduled research generates topic signals from market structure, earnings, filings, news, and source tracking. Topic signals feed a deterministic candidate generator. Output: backtest_candidates.generated.yaml.

Stage 2 Β· Replay

Shadow decisions β†’ outcomes

Each candidate gets a shadow decision committed to a ledger before any draft. Forward returns at fixed horizons (1d / 3d / 5d / 10d / 20d) label each outcome. Outputs: shadow_decisions.jsonl, signal_outcomes.jsonl.

Stage 3 Β· Score

Strategy scoreboard

Outcomes aggregate into strategy and product-maturity scoreboards: outcome proof, calibration, expectancy, evidence quality, learning-loop completeness, operability, support readiness, and authority discipline. Hard vetoes block any score. Output: strategy_scoreboard.json.

Stage 4 Β· Propose

Evolution candidates + promotion review packet

When a candidate clears walk-forward across multiple users, sample-size gates, and hard vetoes, a human-readable review packet is generated. Output: evolution_candidates.jsonl, promotion_reports/YYYY-MM-DD.md.

Stage 5 Β· Approve

Your gate

You read the packet, the evidence chain, and the failure modes. You approve, reject, or send back. LIVE submissions require typed confirmation. Promotion requires explicit human action β€” never automated.

Stage 6 Β· Learn

Postmortem β†’ durable lessons

Real fills get post-trade reviews. Durable lessons are written to TrustMem (only generalizable patterns, not today's PnL). Tomorrow's loop starts with sharper priors.

Who this is for

Trading Brain is built for a narrow user. The architecture and discipline make it a poor fit if you want a "trading bot" or a black-box signal stream.

🧠

Process-driven operators

You care about calibration, evidence, and walk-forward gates more than headline returns. You want the loop to be auditable end to end.

πŸ“ˆ

AI infrastructure investors

Your universe of attention is AI compute, optical, memory, semis, power, cooling. A focused universe helps; a 24-name screener does not.

πŸ›‘οΈ

Capital authority guardians

You believe AI should not authorize capital deployment. The boundary built into Trading Brain's architecture matches your operating principle.

βš–οΈ

Skeptical adopters

You have seen what "AI trading" usually looks like. You want proof, hard vetoes, evidence chains, and an explicit human gate before you trust anything.

Look at the loop. Approve nothing you don't understand.

Beta access starts in paper mode. Sign in with Telegram QR or username/password. Read the evidence chain on every signal. Halt is one click. Live promotion requires typed confirmation. Decide whether Trading Brain's discipline matches your own β€” then decide what to promote.

Request Beta Access Go to Login