Tier 3 chat routes to Qwen Turbo — simple completions for under a tenth of a cent per call.
§ Architecture · 15 agents · 6 hard caps
Every cycle the desk runs a deterministic chain — discovery, analysis, bull-vs-bear review, decision, execution. A verdict only ships when consensus and risk both clear. Open positions cycle through a five-stage revaluation orbit every five minutes.
§ 01 · Pipeline
Each stage hands intelligence to the next. The whole chain runs server-side; the operator sees the trace, not the moving parts.
Market scout + biotech FDA screener via Google Search.
Four parallel specialists: technicals, fundamentals, news, sentiment.
Adversarial bull/bear with a mutual rebuttal round.
Consensus judge + trade architect (position sizing).
Risk analyst + PM final authority + Alpaca routing.
§ 02 · Risk gates
Every trade passes through six independent risk gates before execution. No agent can override them — they are enforced at the infrastructure level.
No single position exceeds 15% of total portfolio equity. Orders auto-reduced or blocked.
Two-plus positions or >25% weight in one sector blocks all new entries in that sector.
No new positions within five days of earnings. Prevents overnight gap risk from binary events.
Consensus confidence must exceed institutional minimum. Learnable threshold — trade debates adjust it.
At least three of four analysts must be directionally aligned. Weak alignment triggers lower confidence.
Cash reserve scales by macro threat level: LOW 10, MODERATE 15, HIGH 20, CRITICAL 30.
§ 03 · Self-learning
Every closed position triggers a structured post-close review. The system learns from both wins and losses, automatically adjusting scanner weights, sector confidence, and conviction thresholds.
Defends entry timing, thesis quality, sizing. Identifies what worked with specific data points.
Attacks every decision — better entry price? More convergence needed? Should we have waited? Missed signals?
Weighs both sides, assigns verdict, proposes concrete parameter adjustments for future trades.
Verdict taxonomy
Debate adjusts these parameters
§ 04 · Always-on infrastructure
These run independently on their own schedules — feeding intelligence, enforcing risk, learning from every trade.
AI researches geopolitics, Fed policy, sector rotation, supply chains. Generates threat-level digest that adjusts cash floor (10-30%) and injects sector-weight scoring modifiers into the scanner.
Feeds into
Position Sentinel scores each holding 0.0-1.0. Reeval PM issues HOLD/TRIM/EXIT with execution timing (immediate, 1h, 4h, end-of-day). Auto-submits stop orders to Alpaca.
Feeds into
Three-round review (Strategist vs Critic) analyses each closed trade. Judge assigns WIN/LOSS verdict and proposes adjustments to sector confidence, scanner weights, conviction thresholds.
Feeds into
The desk never stops re-asking
The same agents that put the trade on are the ones that decide whether it stays on.
§ 05 · Intelligence routing
When macro intelligence detects sector headwinds the system doesn't wait — it cross-references your portfolio, adjusts scanner scoring weights, and scales the cash floor by threat level.
AI + web search across geopolitics, central banks, sector rotation, supply chains.
Threat level → cash-floor adjustment + scanner weight modifiers per sector.
Example
#Airlines DOWN → UAL sentinel score drops → PM: TIGHTEN_STOP (1h)
Example
#Defense UP → scanner boosts macroAlignment for LMT, RTX, GD
"TIGHTEN_STOP UAL (airline headwinds) + boost scanner macroAlignment for #Defense sector"