Silvic V3 System Documentation

Overview

Silvic V3 is an autonomous AI trading desk that trades spot crypto on DEXs with real funds. Claude Sonnet analyzes markets and makes trading decisions. Hard-coded risk limits prevent catastrophic losses. A persistent Order Management System handles real-time stop-losses, trailing stops, and conditional entries between AI cycles.

Core Philosophy

Real-time by default. Event-driven, not batch. Stops execute in seconds.

AI autonomy within guardrails. The AI decides what to trade. The Risk Manager enforces limits.

On-chain is truth. Wallet balances reconciled every cycle. Redis is cache.

No leverage. Spot only. Real tokens in a real wallet.

Schedule & Timing

What happens and when. All times in PST. The system has three layers of timing: scheduled cycles, real-time OMS, and event-driven triggers.

Recurring Schedules

What Happens Every Cycle

Full Cycle (every 4h, ~$0.27-0.77)

Kill switch check, data collection (6 sources), screener (R1/R1b/R4/R5/R6), Analyst (Sonnet, 7 tools), PM (Sonnet or Opus if conviction >0.7), Risk Manager (13 checks), executor queue. Can open new positions, close existing, adjust stops, place conditionals.

Fast Cycle (every 2h, ~$0.12)

Kill switch check, data collection, screener only (no Analyst). PM reviews open positions using screener data. Can adjust stops, close positions, place conditionals. Cannot open new positions (no analyst views). Primary purpose: profit management between full cycles.

What Happens in Real-Time (OMS, every price tick)

Price Monitoring

WebSocket (Arbitrum WETH/PENDLE) + API polling every 15s (all other tokens, both chains). Runs 24/7.

Stop-Loss Execution

Every price tick is checked against stops. Fires within seconds. No AI cycle needed.

Trailing Stop Ratchet

Max price tracked on every tick. Trailing stop = max_price * (1 - trail%). Automatically tightens as price rises. Never loosens.

Auto-Breakeven

When position gains 5%+, fixed stop moves to entry price automatically. No AI cycle needed.

Trail Tightening

When position gains 15%+, trailing stop % tightens from 10% to 6%. Locks in more profit on big winners.

Take-Profit

If price reaches take_profit_price, position is closed immediately. Safety net for between-cycle moves.

Conditional Order Fills

Limit-like entries trigger when price crosses the target. Deviation check prevents bad fills.

Event-Driven Triggers (30-min cooldown)

Emergency Drawdown

Fires when any position drops 8%+ from entry OR daily portfolio drawdown exceeds 3%. Informational alert to Discord (no AI cycle invoked yet -- planned for next sprint).

Opportunity Upside

Fires when any position gains 10%+ unrealized. Same: Discord alert only.

Volatility Event

Fires when any token moves 5%+ in 1 hour. Same: Discord alert only.

Architecture

Six decoupled components communicate via Redis pub/sub. Each can fail independently without taking down the others.

AI Cycle Pipeline

Each cycle flows through 8 stages. Total time: 2-4 minutes, dominated by LLM API calls.

1. Kill Switch Check

If active, abort immediately. No API calls, no actions.

2. Data Collection

6 collectors run in parallel (funding, flows, sentiment, macro, on-chain, price history). Guarantees fresh data for the analyst.

3. Pre-flight Checks

Circuit breaker status, trigger cooldowns, portfolio value, gas balance.

4. Pre-Screener

Deterministic filter. Matches tokens against funding rate, momentum, and other signals using cached Redis data. Zero LLM cost. Focuses the Analyst on promising candidates.

5. Analyst (Claude Sonnet)

7 tools: prices, funding, sentiment, macro calendar, on-chain metrics, recent large swaps, exchange flows. Scans all tokens including Base chain. Produces MarketViews with asset, direction, conviction, and key factors.

6. Portfolio Manager (Sonnet or Opus)

Receives analyst views, portfolio state, bucket exposure, playbook rules, regime score, and benchmark context. Auto-upgrades to Opus when analyst conviction >= 0.7. Decides: open positions (with mandatory trailing stop + take-profit), close positions, adjust stops, place conditional entries, or no action.

7. Risk Manager

13 deterministic checks. No AI. First failure rejects the action. Checks: whitelist, params, stop-loss required, position size, exposure, correlation, pool liquidity, price impact, idempotency, daily loss, drawdown, lifetime drawdown.

8. Executor Queue

Approved actions are pushed to Redis. The executor consumes and executes on-chain via Uniswap Trade API.

Screener Rules

The pre-screener runs before the AI analyst. It uses cached data from Redis to flag tokens that match playbook rules. Zero LLM cost. The analyst prioritizes flagged tokens but also scans all others.

Profit Management

Three layers of profit protection: OMS real-time automation, AI cycle evaluation, and trailing stop mechanics. The goal is to let winners run while never giving back significant gains.

At Entry (PM sets these)

Position Lifecycle Example

Thesis Hold Periods

Mean reversion theses need time. The PM will not exit before the minimum hold unless the stop is hit or the thesis is fundamentally broken.

Order Management System

The OMS runs continuously, independent of AI cycles. It protects the downside and captures profits in real-time.

Price Monitoring

WebSocket connection to Uniswap V3 pool events (Arbitrum WETH/PENDLE). API polling every 15s for all other tokens on both chains. Prices cached in Redis with 5-min TTL.

Stop-Loss Execution

Every price tick is checked against open positions. If price hits the stop level, the position is closed immediately via the executor. Secondary price verification prevents false triggers.

Trailing Stops

MFE (max price since entry) tracked on every tick. Trailing stop = max_price * (1 - trail%). Ratchets up only, never down. Locks in profits while allowing upside.

Auto-Breakeven

When a position gains 5%+ from entry, the fixed stop automatically moves to the entry price. You can't lose money on that trade anymore. Persists to Redis immediately.

Dynamic Trail Tightening

When a position gains 15%+, the trailing stop percentage tightens (e.g., from 10% to 6%). Locks in more profit on big winners. One-way ratchet, persists immediately.

Take-Profit Execution

If price reaches the take_profit_price, position is closed in real-time. The PM sets an initial TP at entry and can raise it on each cycle if momentum continues.

Conditional Orders

Limit-like entries with TTL. When price hits the target, the order executes. Expired orders are cleaned up automatically. Deduplication prevents double entries for the same token.

Trigger Engine

Monitors for emergency drawdown (8% position loss or 3% daily portfolio loss), volatility spikes (5% 1h move), and opportunity signals (10% unrealized gain). Currently informational (Discord alerts with 30-min cooldown). Event-driven AI cycles planned for next sprint.

Data Sources

The Analyst has 9 tools to query market data. All data is pre-cached in Redis by the collector.

Regime Score

A 5-factor composite score that adjusts position sizing and conviction thresholds:

Risk Controls

All risk limits are hard-coded constants. The AI cannot override them. 13 sequential checks, first failure rejects the action.

Execution Flow

Swap Execution

AI proposes trade with token, size, stop, and thesis. Risk Manager validates. Approved actions go to Redis queue. Executor reads queue, gets Uniswap API quote, validates price impact <3%, validates contract target against allowlist, signs Permit2 permit, broadcasts transaction, waits for receipt.

Contract Allowlist

Every transaction target (both approval and swap) is validated against a hardcoded set of known Uniswap contracts (Universal Router, SwapRouter02, Permit2). Transactions to unknown addresses are blocked. Prevents wallet drain if the API is compromised.

Multi-chain

Arbitrum (7 tokens) and Base (6 tokens). Chain resolved per-trade from token config. Separate RPC connections and nonce managers per chain.

Token Whitelist

Arbitrum: WETH, PENDLE, UNI, ARB, LINK, CRV, AAVE. Base: WETH, AERO, VIRTUAL, MORPHO, ZORA, LINK. WBTC is reference-only (benchmarks, regime) but never traded.

Model Assignment

Normal cycle: ~$0.17 (Analyst) + ~$0.12 (PM Sonnet) = ~$0.29. High-conviction cycle: ~$0.17 + ~$0.53 (PM Opus) = ~$0.70. Fast screener: ~$0.12 (PM only). Estimated daily: ~$3-5 depending on Opus upgrades.