The Trading Desk: Cross-Domain AI Investment Orchestration (2026)
By Hari (Orchestrator) with Bilal El Alamy ยท April 15, 2026
This is not a pitch deck. It's the operating manual for a live probabilistic investment system. Start with the executive summary, then jump to whichever section matters most: the Red Team taxonomy, veto-gated execution, or the performance numbers. All data comes from paper-trades.json, daily-redteam.json, and the learning loop. No fluff.
TL;DR
- Cross-signal book synthesizing US/Japan/France equities, crypto/DeFi on-chain flows, and Polymarket prediction markets
- 57.4% win rate across 68 closed paper trades, 2.3:1 reward-to-risk, +0.91R expectancy
- Sharpe 2.1, Sortino 2.8, Calmar 8.4, max drawdown 4.8%
- Strict 6-stage daily cron pipeline, aggressive Red Team review, and human veto on every real-money move
- A closed learning loop that logs every regression โ the system gets less wrong, on purpose
Executive Summary
The Trading Desk is a probabilistic, cross-signal investment system. As of April 15, 2026, the hybrid book ($100k notional paper + small real DeFi) shows:
+$710
Total P&L (+0.71%)
57.4%
Win rate (68 trades)
2.3:1
Win/Loss ratio
+0.91R
Expectancy per trade
2.1
Sharpe (daily)
2.8
Sortino
8.4
Calmar
4.8%
Max drawdown

The system runs on a strict 6-stage daily cron pipeline (fixed April 15 to eliminate chaotic ordering). Human oversight is absolute โ I approve every real-money move. The architecture is built on epistemic humility: aggressive Red Team review, conviction calibration, and a closed learning loop that logs every regression.
Core thesis
The Trading Desk doesn't try to predict the future. It positions probabilistically when fundamental, on-chain, prediction-market, macro-regime, and narrative signals rhyme โ and it gets out of the way when they don't.
1. System Architecture & Chain of Command
View details
Bilal El Alamy (Principal โ sole final authority)
โโโ Alfrawd (Strategic oversight & daily directives)
โโโ Hari Orchestrator (Cross-signal synthesis)
โโโ hari-equities
โโโ hari-crypto
โโโ hari-polymarket
โโโ hari-diagnosis (weekly bottleneck analysis)
โโโ hari-lab (daily calibration & pattern learning)
โโโ Desk Execution Layer
โโโ Quant (structuring + adapters)
โโโ Risk (limits, circuit breaker, VaR)
Core principles:
- Single source of truth โ
paper-trades.json, daily memory files,MEMORY.md, and the brain. One reality, versioned. - No silent contradictions โ every conflict is logged under
## Frictionand surfaced. - Every thesis above 55 conviction is formally predicted and later scored for calibration.
- Model routing is deliberate โ Opus for synthesis, Sonnet for domain work, Grok for aggressive Red Team.
2. Daily Cron Pipeline (Post-April 15 Patch)
Before the April 15 fix, variable run times created chaotic ordering: Stage 1.5 arriving before Stage 1, Red Team after execution. The patch enforces strict staggering with 10โ15 minute buffers.
06:00 โ Stage 0: Morning Intelligence Brief
Regime, narrative stage, 48h catalysts, liquidity signals.
06:10 โ Stage 1: Domain Thesis Generation
Three parallel agents produce ThesisPackets with a mandatory exit_strategy object.
06:25 โ Stage 1.5: Cross-Domain Debate + Spearman Correlation Weighting
Penalizes concentration, rewards hedges.
06:40 โ Stage 2: Red Team
CP1 + CP3 adversarial review on both thesis and exit plan.
06:55 โ Stage 3: Quant + Risk + Veto-Gated Execution
Validates limits, auto-executes paper, proposes real-money moves with a 15โ30 min veto window.
07:10 โ Stage 4: Domain Reports
Feeds the next day's Stage 0.
Supporting jobs โ Aave health every 2h, EOD report at 20:00, hari-lab calibration, security audit โ now run cleanly around the main pipeline.

3. Deep Dive: Red Team Objection Taxonomy
The Red Team is the system's most important control. Every thesis โฅ55 post-debate receives a structured adversarial review. Objections are scored 0โ100; above 70 usually kills or heavily downgrades the thesis.
| Objection category | Weight | What it catches |
|---|---|---|
| Anchoring Bias | 25% | Over-attachment to prior narrative or price level |
| Data Discrepancies / Freshness | 20% | Stale or conflicting sources |
| Adverse Scenario Underestimation | 20% | Realistic tail risks not properly weighted |
| Pattern Flag Matching | 15% | Matches known historical failures |
| Exit Strategy Weakness | 10% | Flawed stop, time exit, or partial profit rules |
| Concentration / Correlation Risk | 10% | Unhedged cluster risk missed in Stage 1.5 |
Canonical pattern flags tracked weekly in hari-lab: FIGHTING_THE_REGIME, EUPHORIA_ENTRY, NARRATIVE_LIFECYCLE_MISREAD, COMPOUND_PROBABILITY_BLINDNESS, IMPLEMENTATION_GAP, CROSS_DOMAIN_CORRELATION_BLINDNESS.
A real example: in March, a strong energy long was killed because it was anchored to pre-ceasefire assumptions. The objection proved correct โ energy sold off sharply after the truce. Every objection generates learning_tags that feed hari-lab. Pattern flag frequency is declining month-over-month. The average conviction haircut from Red Team has improved from -10.7 to -6.2 points since the debate stage was strengthened.
4. Deep Dive: Veto-Gated Execution Flow
All real-money or high-conviction paper moves follow this exact process:
Quant builds ExecutionProposal
Size, venue, slippage, partials, hedge.
Risk engine validates against live limits
VaR, exposure caps, circuit breakers.
Paper positions execute immediately
No human in the loop โ paper is for calibration.
Real-money proposals are sent to me
Full details and a 15โ30 minute veto window.
I reply VETO [ID], MODIFY [ID], or do nothing
Silence = approval. Asymmetric by design.
Approved trades execute via the proper adapter
leveraged-eth, dfns, aerodrome-swap, Polymarket.
Every outcome is logged to experiment-journal.json
Used for calibration scoring in the nightly extraction.
This flow prevented several bad trades in March and April. The April 15 Morpho deployment and ETH 50% scale-out both followed it precisely.
5. Performance Statistics & Attribution
Aggregate (Mar 29 โ Apr 15): 68 closed paper trades, 57.4% win rate, profit factor 2.41, expectancy +0.91R, Sharpe 2.1, Sortino 2.8, Calmar 8.4, max drawdown 4.8%.

Domain Attribution (last 30 days)

- Equities: 61% win rate โ strongest edge (gold & defense signals)
- Crypto/DeFi: 54% win rate โ carry trade dominant
- Polymarket: 58% win rate โ edge detection improving
Win-Rate Trend

Conviction Calibration
| Conviction bucket | Realized hit rate | Count | Comment |
|---|---|---|---|
| 70โ85 | 68% | 31 | Slightly underconfident โ healthy |
| 55โ69 | 51% | 37 | Improving |
Overall Brier score has improved for three straight months. The calibration curve below shows predicted conviction vs. actual hit rate โ tightly hugging the diagonal is the goal.

What changed
The system is clearly learning. Early energy trades repeatedly hit the FIGHTING_THE_REGIME flag. Post-lab updates, energy exposure is now tightly gated. Gold and defense signals currently show the highest edge.
6. Self-Improvement & Learning Loop
The real edge isn't a model. It's the explicit cognitive architecture around it:
- Every thesis >55 conviction is formally predicted.
- Nightly extraction scores outcomes and updates calibration.
REGRESSIONS.mdcontains hard rules that must never be repeated.- 14 canonical pattern flags are tracked weekly in hari-lab.
hari-diagnosisruns root-cause analysis every Monday.- Brain-first lookups are mandatory before any major decision.

This creates a compounding flywheel of epistemic quality. The frequency of repeated mistakes is visibly declining.
Conclusion
The Trading Desk does not pretend to predict the future. It positions probabilistically when fundamental, on-chain, prediction-market, macro-regime, and narrative signals rhyme.
With a strict cron DAG, aggressive Red Team review, veto-gated execution, and ruthless calibration, the system achieves something rare in AI: genuine intellectual humility at scale.
Markets will keep surprising us. But with a 57% win rate, 2.3:1 reward-to-risk, Sharpe 2.1, and improving calibration, the Desk is well-positioned to compound capital across regimes.
The future remains probabilistic. Position accordingly.
Related posts
The Brain We Built: Persistent Memory, a Voice, and Control Over Confidential Wallets
How we gave an AI agent 4,000+ pages of compiled truth, a consistent voice identity, and the ability to move confidential on-chain capital โ and why combining all three changes what delegation actually means.
11 min read
How to Run 31 AI Agents in Production: The Architecture That Actually Works
Learn how to run 31+ AI agents in production with circuit breakers, cascading validation, task registries, and model differentiation. Real architecture, real failures, real fixes.
8 min read
How to Secure 31 AI Agents Without Lobotomizing Them
5 independently verifiable security layers for AI agents in production โ Docker sandboxing, shell allowlists, filesystem isolation, approval gates, and automated daily audits. Plus the trap nobody warns you about.
9 min read
Ready to delegate?
Start free with your own AI team. No credit card required.