Framework

Decision Infrastructure Framework

The canonical architecture for systems that transform exploration into qualified execution.

v0.2·working·Updated July 2026

Overview

Most decision systems are built as monoliths — generation and execution live in the same layer, validated by the same logic, governed by the same feedback loops. This works until the cost of a bad decision exceeds the cost of building infrastructure to prevent it.

Decision Infrastructure separates the pipeline into distinct architectural layers, each with its own responsibility, evidence requirements, and failure modes. The framework makes visible the structural arrangement that determines whether intelligence earns trust, resources, and action.

This framework is spatial, not temporal. The layers exist simultaneously and continuously. A decision does not "pass through" the framework once — it lives within it, accumulating evidence, responding to environmental change, and earning or losing trust over time.

Design Principles

These are governing constraints, not guidelines. Every architectural decision in the framework must be traceable to one of these principles.

01

Separation of Generation and Qualification

The system that generates possibilities must be architecturally separate from the system that evaluates them.

When generation and evaluation share infrastructure, the generator optimizes for passing evaluation rather than producing genuine diversity. Architectural separation prevents this collapse.

02

Evidence Before Allocation

No resource flows to a decision until evidence justifies the allocation.

Confidence without evidence is opinion. The framework gates every resource transition on accumulated, structured evidence — not conviction, narrative, or recency.

03

Continuous Qualification

Trust is not a one-time grant. Every qualified decision must continuously re-earn its position through ongoing evidence.

Environmental conditions change. A decision that was optimal yesterday may be harmful today. Qualification must be a living process, not a historical credential.

04

Failure Isolation

The failure of any individual decision must not compromise the infrastructure that evaluates all decisions.

Individual decisions fail. That is expected. The system fails when a single failure cascades into infrastructure damage — corrupting evaluation criteria, qualification thresholds, or resource allocation logic.

System Map

Five architectural layers, each with distinct responsibilities. Intelligence flows downward through increasingly rigorous gates. Evidence flows upward, informing every layer above.

1
Generation Layer3 components
Template EngineAI GeneratorFoundry
2
Validation Layer2 components
Backtest EngineRegime-Aware Scoring
3
Qualification Layer2 components
Paper Trading SimulationQualification Engine
4
Execution Layer2 components
Live Execution EngineRisk Guard
5
Learning Layer2 components
Outcome AttributionIntelligence Aggregation
Intelligence flows down|Evidence flows up

Layers & Components

Each layer has distinct responsibilities, inputs, outputs, and ownership. No component spans multiple layers.

1

Generation Layer

Produces candidate decisions with maximum structural diversity. Optimized for exploration, not quality.

Template Engine

Deterministic candidate generation with diversity constraints

In: Market context, Regime classification, Learning summaryOut: Candidate specsOwns: Generation templates and diversity quotas

AI Generator

LLM-based candidate generation for structural patterns beyond templates

In: Market context, Regime classification, Indicator rankingsOut: Candidate specsOwns: AI generation prompts and validation pipeline

Foundry

Automated research pipeline — mutations, cross-asset transfer, inverse mining

In: Existing strategies, Learning summary, Regime catalogOut: Mutated candidatesOwns: Foundry workspace, admission/retirement logic
2

Validation Layer

Tests candidates against historical evidence. Produces structured confidence assessments, not pass/fail verdicts.

Backtest Engine

In-sample and out-of-sample statistical validation

In: Candidate spec, Historical price dataOut: Validation results with confidence intervalsOwns: Backtest execution and scoring formulas

Regime-Aware Scoring

Evaluates performance within specific market conditions

In: Validation results, Regime catalogOut: Regime-conditional performance metricsOwns: Regime classification and fit scoring
3

Qualification Layer

Gates resource allocation on accumulated evidence. The critical boundary between exploration and deployment.

Paper Trading Simulation

Forward-testing in live market conditions without capital risk

In: Approved strategies, Live market dataOut: Forward performance recordOwns: Paper accounts, position tracking, isolated capital

Qualification Engine

Multi-dimensional qualification assessment

In: Backtest results, Paper performance, Regime proofOut: Qualification status with evidence trailOwns: Qualification criteria, cohort analysis, divergence detection
4

Execution Layer

Manages resource deployment for qualified decisions. Enforces safety invariants.

Live Execution Engine

Translates qualified decisions into market actions

In: Qualified strategy, Live market data, Risk parametersOut: Executed trades, Position stateOwns: Order routing, position management, exit evaluation

Risk Guard

Circuit breakers and exposure limits

In: Portfolio state, Market conditionsOut: Kill signals, Pause directivesOwns: Drawdown limits, exposure caps, correlation guards
5

Learning Layer

Transforms outcomes into institutional knowledge. Feeds evidence back to every layer above.

Outcome Attribution

Decomposes results into parameter-level evidence

In: Closed positions, Entry conditions, Regime contextOut: Attribution records, Parameter guidanceOwns: Exit audit trail, decision fingerprints

Intelligence Aggregation

Synthesizes evidence across strategies, regimes, and time

In: Attribution records, Qualification data, Regime transitionsOut: Learning summaries, Generation guidanceOwns: Evaluation batch, fit scores, fee-bleeder detection

States & Transitions

A decision moves through these states as it accumulates evidence. Each transition is gated — there are no free promotions.

Candidate
Draft
Validated
Active (Paper)
Qualified
Active (Live)

Candidate

Generated but not yet validated. May be structurally invalid.

Entry: Produced by any generation source

Exits: Passes pre-validation → Draft · Fails pre-validation → Rejected

Draft

Structurally valid, awaiting historical validation.

Entry: Passes spec validation and diversity checks

Exits: Passes backtest → Validated · Fails backtest → Failed

Validated

Historical evidence supports the strategy. Ready for forward testing.

Entry: Backtest results meet minimum thresholds

Exits: Approved for paper trading → Active (Paper)

Active (Paper)

Running in live market conditions without capital risk.

Entry: Admitted to portfolio with paper capital allocation

Exits: Qualifies → Qualified · Retires → Retired · Paused → Paused

Qualified

Evidence supports live deployment. Eligible for capital allocation.

Entry: Meets qualification criteria with regime proof

Exits: Promoted → Active (Live) · Disqualified → Paused

Active (Live)

Deployed with real capital. Subject to continuous qualification.

Entry: User promotes with capital allocation

Exits: Killed → Killed · Regime transition → Paused

FromToTriggerGuard
CandidateDraftPre-validation passValid spec structure
DraftValidatedBacktest completionComposite score ≥ threshold, IS/OOS coherent
ValidatedActive (Paper)Portfolio admissionDiversity constraints met, bot capacity available
Active (Paper)QualifiedQualification engineCohort divergence, regime proof, minimum duration
QualifiedActive (Live)User promotionExchange connection, capital available, fee status clear
Active (Live)PausedRegime transition or killToxic transition detected or user request

Boundary Rules

Boundaries define what crosses between layers and what does not. Every boundary has an explicit failure mode.

Generation → Validation

Crosses

StrategySpec (canonical format with indicators, risk parameters, exit profile)

Does Not Cross

Generation confidence, template source, AI model identity

Failure Mode

Invalid spec structure → rejected at pre-validation. Duplicate fingerprint → deduped.

Validation → Qualification

Crosses

Validation results with confidence intervals, regime-conditional scores

Does Not Cross

Raw backtest trades, intermediate calculations, engine internals

Failure Mode

Failed validation → terminal state. No retry without spec modification.

Qualification → Execution

Crosses

Qualified strategy with evidence trail, regime proof, capital allocation

Does Not Cross

Qualification scores, internal rankings, paper account details

Failure Mode

Disqualification during live → position closed, capital returned. No silent degradation.

Execution → Learning

Crosses

Closed position outcomes with entry conditions, exit mechanics, regime context

Does Not Cross

Open position state, unrealized PnL, real-time market data

Failure Mode

Missing attribution → exit logged without enrichment. Learning degrades gracefully.

Operational Semantics

How this framework operates in practice — cadence, consumers, and staleness rules.

DimensionValueRationale
Generation cadenceOn-demand (user) + 3x/day (foundry)User generation is interactive. Foundry sweeps run autonomously to maintain pipeline diversity.
Validation latency30-90 seconds per strategyBacktests run against 90-day rolling windows with IS/OOS split.
Qualification assessmentEvery 6 hours (batch)Balances evidence freshness against compute cost. Batch includes qualification, evaluation, and intelligence scoring.
Paper tick cadence1 minuteEvaluates all paper positions against live market data. Event-driven with polling fallback.
Live tick cadence1 minuteMatches paper cadence. Exchange orders placed on signal.
Learning aggregationWeeklyLearning summaries aggregate across all outcomes. Feeds back into generation prompts and template parameters.
Regime classificationHourly (snapshot) + weekly (catalog rebuild)Hourly snapshots capture current conditions. Weekly rebuild derives regime_periods from snapshot history.

Knowledge Lineage

Related Research

See this framework implemented in Orqis

Research — Warren Labs | Orqis