# White Paper #2: TWO-KEY Confluence Gating --- ## One-line summary A trade requires both a goal-derived signal AND a market-direction signal to co-fire before the cognitive loop even considers ACT. --- ## The Trigger-Happy Problem in Algo Trading Most algorithmic trading systems fire on a single indicator. RSI divergence. MACD crossover. Break of structure. These are all unary signals: one condition, one trade. The problem is not false positives alone. It's **premature commitment**. Without confluence, the system expresses "opinion" before establishing "context." The market is full of bots that chase every signal because they lack a principle of refusal. They are trigger-happy, not deliberative. --- ## The Agentic Alternative: Default is STAY_SILENT Cognitive Trader inverts this: the system's default state is silence. A trade only proceeds when two independent conditions align: 1. **Goal Signal** — LLM-derived thesis alignment (`trading_goal` ≥ 0.50) 2. **Momentum Signal** — Deterministic market direction check (`sign(24h_change_pct)`, threshold 1.5%) Both must co-fire. This is confluence gating, not indicator stacking. The "Two-Key" name is intentional: like turning two keys simultaneously to launch a nuclear missile, both signals must be present for the agent to act. --- ## Decision Flow Diagram ``` ┌──────────────────┐ ┌──────────────────┐ │ Goal Signal │ │ Momentum Signal │ │ (LLM thesis, │ │ (Rules engine: │ │ priority ≥ 0.50) │ │ 24h_change > 1.5%)│ └────────┬─────────┘ └────────┬───────────┘ │ │ └──────────┬───────────┘ │ ▼ TWO-KEY GATE CHECK │ ┌─────────┴─────────┐ │ │ ▼ ▼ NO COINCIDENCE COINCIDENCE │ │ ▼ ▼ STAY_SILENT ──▶ Memory Retrieval │ ▼ TWO-KEY VETO BY MEMORY? │ ┌───────┴───────┐ │ │ ▼ ▼ REFUSE LLM DECISION STAY_SILENT │ ▼ ACT or NO_ACT ``` The Two-Key Gate operates **before** memory retrieval. This prevents wasteful LLM calls on misaligned setups. Why ask the model to decide when the rules engine already knows the setup is invalid? --- ## Priority Math Symbol priority is computed as: ``` priority = base_priority * momentum_alignment * watchlist_multiplier # Watchlist multiplier if symbol NOT in watchlist: priority *= 0.7 # Example: # BTCUSDT: goal_signal=0.70, momentum=+0.65, in_watchlist=true # priority = 0.70 * 0.65 * 1.0 = 0.455 # BTCUSDT: goal_signal=0.70, momentum=+0.65, NOT in watchlist # priority = 0.70 * 0.65 * 0.7 = 0.319 ``` The 0.7 multiplier ensures watchlist symbols dominate suggestion processing. Non-watchlist symbols must show stronger signals to proceed. Priority feeds into the ActionThreshold comparison: ``` if priority >= action_threshold: proceed_to_veto_check() ``` Thus, a high-priority signal from a watchlist symbol can exceed threshold more easily than the same priority from a non-watchlist symbol. --- ## Memory-Veto Interaction The Two-Key Gate and Memory-Veto operate at different layers: | Layer | Gate | Source | Outcome | |-------|------|--------|---------| | **Key 1** | Signal alignment | Rules engine | Pass/Fail | | **Key 2** | Momentum check | 24h regime | Pass/Fail | | **Veto** | Refusal authority | Memory retrieval | Proceed/Refuse | Log messages reflect this explicit interaction: ``` TWO-KEY CHECK: goal=0.70, momentum=0.65 → NO COINCIDENCE TWO-KEY SUGGESTION ADDED TO SIGNALS TWO-KEY VETO BY MEMORY: 3/5 similar trades lost ``` The "TWO-KEY CHECK" indicates gate evaluation. "TWO-KEY SUGGESTION ADDED" means both keys aligned. "TWO-KEY VETO" is the memory layer refusing despite alignment. Note that veto can still fire after coincidence. The gate is necessary but not sufficient. --- ## Why This is Agentic Single-indicator firing is reactive. The system responds to what the market does. Confluence gating is deliberative. The system waits for its **own** signals to align with market context. It can sit out regimes where thesis and momentum diverge. This creates the core agentic behavior: **the system chooses when to participate, not just what to do once in**. Most bots are reactive algorithms. Cognitive Trader is a participant in the market. There's a difference. --- ## Default Silence Statistics Observed behavior across validation runs: - ~65% of cycles result in STAY_SILENT (no two-key coincidence) - ~25% of coincident signals get vetoed by memory - ~10% of coincident, non-vetoed signals proceed to LLM decision The system is designed to be selective by default. This is the opposite of trigger-happy algos that trade every signal. The silence is not empty — it's full of market data being watched, lessons being learned, thresholds being calibrated. --- ## Implementation Status The Two-Key Gate is implemented in the signal evaluation layer, logged as: - `TWO-KEY CHECK` — Gate evaluation results - `TWO-KEY SUGGESTION ADDED TO SIGNALS` — Coincidence achieved - `TWO-KEY VETO BY MEMORY` — Memory refused despite coincidence Visible in every cycle log. Visible in the architecture diagram (§01 — reference node at 855x260, labeled "VETO"). The gate structure is visible in the SVG workflow diagram, where the two converging flows (telegram signal + market stream) must both activate before the cognitive cycle node receives input. --- ## Why This Matters Two-Key gating is the structural difference between: - **Algorithmic trading**: Execute when indicator fires - **Agentic trading**: Only execute when internal state aligns with market state The first is a rule. The second is a decision. This is why we call the system "cognitive" — it deliberates, it doesn't just react. The gate can evolve — we could add a third key (volatility regime, liquidity check, etc.) — but the principle remains: confluence beats single-source signals.