WP-02 — Adaptive Risk Engine: Two-Tier Protection Model

AI agents that manage real or simulated capital must be constrained by something more durable than prompt instructions. The Adaptive Risk Engine is the constraint layer that sits between Treeova's agents and any executed action that affects a position. It enforces deterministic guardrails first, and only then allows context-sensitive adaptive behavior on top.

Standard tier (always on). A fixed set of deterministic guardrails that fire regardless of agent state: position-level stops, safety floors, the penny option guard, and phantom-fill protection. These rules are not subject to agent override. They define the worst-case behavior the platform will tolerate.

Adaptive tier (modulated). Layered on top of the standard tier, the adaptive tier modulates trailing behavior — when to arm a trail, how aggressively to tighten it, when to take partial exits — based on the position's conviction score, the current market regime, and per-position risk state. The adaptive tier can never weaken a standard guardrail; it can only strengthen the user's protection inside the standard envelope.

A foundational rule of the engine: the AI agent does not perform risk arithmetic. The agent's role is to choose which protection mode is appropriate given context — for example, escalating from a loose trail to a tight trail when conviction drops. The numerical consequence of that choice (the exact stop price, the exact trail distance, the exact share or contract count) is computed by deterministic platform code.

This separation has two benefits. First, it makes risk behavior fully auditable: every numerical risk decision is reproducible from inputs. Second, it removes a class of failure modes — an LLM hallucinating a price or quantity simply cannot translate into an unsafe action because the LLM is never the source of the number.

Trailing behavior is modeled as an explicit state machine with four states:

• Idle. No trail active. Standard-tier stops apply.
• Armed. A trail has been authorized but not yet engaged. The platform is watching for the deterministic arming condition.
• Trailing. The trail is engaged. The platform tracks the high-water mark for the position direction and updates the protective level deterministically.
• Triggered. The trail has fired. The platform initiates an exit order through the appropriate execution path.

State transitions are driven only by deterministic conditions on price, time, conviction, and risk-throttle setting. The agent can request a transition; the platform decides whether the conditions permit it.

Very-low-value option contracts produce pathological behavior under naive stop logic — a one-tick adverse move can crystallize a disproportionate percentage loss. The penny option guard enforces a minimum sensible exit price and routes such positions through a dedicated decision path that respects intrinsic value when market data is unreliable.

Phantom-fill protection prevents simulated executions in the paper environment from booking unrealistic gains or losses caused by stale or one-sided quotes. The principle generalizes: the engine refuses to act on a price it cannot trust.

Each agent runs against a persisted risk-throttle setting that governs how aggressively it can size positions, how fast it can compound, and how it reacts to drawdown. The throttle is the authoritative server-side value; client-side state cannot relax risk by mistake. Changes to the throttle are recorded in the audit log alongside all other administrative actions.

Every risk-relevant decision — arming a trail, triggering a stop, throttling an agent, applying a safety floor — is recorded as a structured event tied to the position, the agent, and the user. The audit trail makes it possible to reconstruct exactly why the engine acted the way it did at any point in a position's life.

• The engine cannot eliminate market risk. Gaps, halts, and broker outages can all produce outcomes worse than the modeled stop.
• Adaptive behavior is only as good as the conviction and regime inputs it consumes. In ambiguous regimes the adaptive tier is intentionally conservative.
• Specific numerical thresholds (floor values, trail distances, throttle scaling) are intentionally withheld and may evolve. The architecture and the principles documented here are stable; the constants are not.
• Past performance does not guarantee future results. The Adaptive Risk Engine is a discipline layer, not a guarantee.