Decision Lifecycle

01

The most common model for decision-making infrastructure is the pipeline: generate, validate, deploy. Information flows in one direction. A candidate enters the system, passes through gates, and either reaches execution or is discarded. The pipeline model is intuitive, easy to build, and fundamentally incomplete.

The incompleteness is not in the forward path — pipelines are perfectly adequate at moving candidates from generation to deployment. The failure is in what happens after deployment. In a pipeline, outcomes are recorded but not routed. The system knows that a strategy succeeded or failed, but that knowledge sits in a database table rather than flowing back into the generation process that produced the strategy in the first place.

A lifecycle differs from a pipeline in one structural property: the final stage connects back to the first. When learning — the seventh stage — feeds its observations back into exploration — the first stage — the system becomes self-improving. Without this closure, you have a record-keeping system that happens to trade. With it, you have infrastructure that gets better at making decisions over time.

02

The lifecycle operates through seven stages, each with a distinct purpose, timescale, and set of outputs. The stages are not arbitrary divisions — they reflect the fundamental activities required to move from “this might work” to “this is working” to “here is what we learned.”

Exploration is where candidates are born. The foundry runs three sweeps per day, each producing approximately 20 candidates through a combination of deterministic templates and AI generation. Exploration is deliberately cheap — approximately $0.05 per candidate — because the system’s selective power comes from later stages, not from restricting generation. Coverage-aware budget allocation boosts under-served regimes, and a 20% exploration budget ensures that the system does not converge prematurely on known-good patterns.

Evidence is the first quantitative gate. Each candidate undergoes vectorbt backtesting with a 70/30 in-sample/out-of-sample split, composite scoring with fee-efficiency adjustment, and robustness labeling. A validation run takes 30 to 120 seconds. Approximately 31% of generated candidates pass this stage — the first major reduction in the lifecycle funnel.

Simulation is where time becomes the evidence mechanism. Approved candidates enter paper trading at a 1-minute tick cadence with full exit profile evaluation (ATR-based breakeven, trailing stops, partial take profit). The minimum observation period is seven days with at least five closed trades. Paper trading never auto-pauses — by design, the system needs to observe how strategies behave in all conditions, including unfavorable ones.

Qualification is the 14-criteria evaluator that runs every six hours. It assesses outcome metrics (win rate, ROI, drawdown), process metrics (observation duration, trade count, fee viability), and regime-specific proof. The gate is fail-closed: if evidence is uncertain, the strategy is blocked. Only 22.2% of paper-traded strategies achieve qualification.

Allocation is advisory, not prescriptive. A 6-component intelligence score ranks qualified strategies by deployment attractiveness, considering equity-curve Sharpe, regime fit, correlation diversity, and behavioral stability. Allocation does not gate deployment — it informs the decision about which qualified strategies deserve capital first.

Execution is the narrowest point in the funnel. Live trading operates on Hyperliquid through a 5-layer gate architecture with regime gate persistence, qualified scout synchronization, and reconciliation heartbeats. Positions are managed at 1-minute tick cadence with the same deterministic exit evaluation as paper trading. Only 5.1% of qualified strategies are promoted to live capital.

Learning is the stage that transforms the lifecycle from a pipeline into a cycle. Six deployed learning phases continuously capture observations: exit attribution, regime transition survivability, fee-bleeder detection, indicator effectiveness, divergence prediction, and strategy correlation. The weekly learning summary synthesizes these into structured guidance that is injected directly into the next generation prompt — closing the loop.

03

The lifecycle is formalized as a state machine with explicit promotion and demotion paths. Each transition requires specific evidence, and the system enforces that evidence programmatically rather than through human review.

The primary forward path is: draft → validated → approved → active (paper trading) → paused → promoted (live). Each arrow represents a gate with specific criteria. The draft-to-validated transition requires a passing backtest. The approved-to-active transition requires portfolio admission. The active-to-promoted transition requires 14-criteria qualification plus user intent.

Demotion paths are equally important. A qualified strategy can be degraded if paper trading performance deteriorates. A live strategy can be killed — closing exchange positions and revoking wallet access — if the kill switch is activated. A foundry strategy can be retired if ROI falls below -5% after 14 days. These demotion paths ensure that qualification is not permanent: the system continuously re-evaluates and revokes permission when evidence degrades.

One transition is notably absent: paper trading never pauses automatically for performance reasons. This is a deliberate design decision. The system needs to observe strategy behavior in unfavorable conditions — a strategy that is paused during drawdowns produces biased evidence. The only automatic pause in paper trading is the regime gate: strategies are paused when the market enters a regime where they have no proof of viability, and resumed when a favorable regime returns.

04

The feedback architecture is the most important structural property of the lifecycle — it is what distinguishes a cycle from a pipeline. Ten feedback loops are fully closed in the production system, meaning that observations from later stages directly and measurably alter behavior in earlier stages.

The most significant loop is the weekly learning summary. Every Monday at 4am UTC, the system aggregates backtest patterns, paper trading outcomes, exit attribution data, regime transition matrices, fee-bleeder rates per family and timeframe, and the top 10 qualified paper performers with their specific indicator patterns. This synthesis is injected directly into the AI generation prompt, so that the next generation cycle is informed by the accumulated evidence of every previous cycle.

Other loops operate at different cadences and granularities. The exit profile advisor adjusts exit mechanism rotation based on per-family exit attribution data — if a particular exit profile consistently underperforms for momentum strategies, the generator rotates away from it. Fit scores from paper trading feed into the search orchestrator, boosting families that have demonstrated strong performance patterns. Regime classification feeds directly into family filtering, excluding families that are inappropriate for current market conditions.

The annotation-driven loops are particularly illustrative. When the 6-hour evaluation batch detects HIGH severity degradation or overfitting, two things happen simultaneously: the strategy is paused for allocation purposes, and an inverse strategy is automatically generated — the same structure with flipped direction and operators. This is a concrete example of a failure being routed into a new exploration rather than simply recorded.

Not all loops are equally mature. Four additional loops are partially closed — data flows downstream but with limited measurable impact. Four more are open: data is collected and stored but no downstream system consumes it yet. The M2 regime transition model, the M6 anomaly detector, strategy DNA fingerprinting, and behavioral intelligence as a standalone signal are all in this category. The feedback architecture is a living system with gaps — but the gaps are known and tracked.

05

The lifecycle, viewed end-to-end, forms a funnel with severe attrition. Less than 1% of generated strategies survive to live deployment. This conversion rate is not a bug to be optimized — it is a feature of a system that takes the cost of unqualified execution seriously.

The funnel’s shape reveals the relative selectivity of each stage. Validation passes approximately 31% of candidates — the broadest gate, designed to filter out structurally unsound strategies quickly. Qualification passes 22.2% of paper-traded strategies — a narrower gate that requires sustained evidence. Live promotion selects only 5.1% of qualified strategies — the most selective gate, reflecting the gap between “this strategy has earned the right to execute” and “this strategy deserves real capital now.”

The time asymmetry across the funnel is equally revealing. Generating a candidate takes 1 to 5 minutes and costs approximately $0.05. Qualifying it requires a minimum of seven days of paper trading with at least five closed trades. Executing it involves positions held for minutes to months. The system is designed so that the cheapest operation (generation) produces the most candidates, while the most expensive operation (live execution) receives the fewest — each one backed by accumulated evidence from every preceding stage.