AI Engine

The AI engine is an autonomous decision system that manages search strategy, resource allocation, and campaign orchestration. It replaces manual tuning with a unified OODA decision loop that continuously adapts to fleet state, discovery patterns, and cost constraints.

OODA Decision Loop

Every 30 seconds the engine executes a full Observe → Orient → Decide → Act cycle:

┌──────────┐     ┌──────────┐     ┌──────────┐     ┌──────────┐
│ Observe  │────▶│  Orient  │────▶│  Decide  │────▶│   Act    │
│          │     │          │     │          │     │          │
│ Snapshot │     │  Score   │     │  Select  │     │ Execute  │
│ fleet,   │     │  forms,  │     │  best    │     │ start/   │
│ costs,   │     │  weight  │     │  action  │     │ stop/    │
│ records  │     │  drift   │     │  plan    │     │ reconfig │
└──────────┘     └──────────┘     └──────────┘     └──────────┘
      │                                                   │
      └───────────────── Learn ◀──────────────────────────┘

Observe: WorldSnapshot

A single consistent view of the entire system assembled via parallel database queries in ~50ms:

Active workers, their capabilities, and current assignments
Running searches with progress, rate, and stall detection
Recent discoveries and their forms
Budget velocity and remaining compute budget
World record standings per form (scraped from t5k.org)
Cost model coefficients from calibration data

Orient: Scoring Model

Each candidate search form is scored using a 7-component weighted model:

Component	Weight	Description
`record_gap`	Dynamic	Distance to current world record — closer means higher payoff
`yield_rate`	Dynamic	Historical primes-per-core-hour for this form
`cost_efficiency`	Dynamic	Expected cost per discovery using the power-law cost model
`opportunity_density`	Dynamic	Untested candidate density in the target range
`fleet_fit`	Dynamic	How well the form matches available worker hardware
`momentum`	Dynamic	Recent discovery trend — reward hot streaks
`competition`	Dynamic	External search activity on competing platforms

Weights are learned via online gradient descent, comparing predicted outcomes against actual discovery data. The learned weights are persisted in the ai_engine_state database table.

Decide & Act

The decision phase selects from a set of possible actions:

Start search — Launch a new search on the highest-scored form
Stop search — Terminate a stalled or low-yield search
Reconfigure — Adjust sieve depth, worker count, or range parameters
Scale — Request more workers or release idle ones
Hold — No action needed (system is performing well)

Learn: Outcome Tracking

Every decision is recorded in the ai_engine_decisions table with reasoning text, confidence score, and eventual outcome. This audit trail enables:

Weight updates via gradient descent on prediction error
Post-hoc analysis of strategy effectiveness
Debugging poor decisions with full context replay

Cost Model

The cost model predicts compute time for a work block using a power-law regression fitted to historical data:

text

cost(digits) = a * digits^b

Where:
  a, b  = OLS-fitted coefficients on log-log work block data
  digits = candidate digit count

Fallback defaults (when insufficient data):
  factorial:    a=1e-6,  b=2.5
  kbn:          a=1e-7,  b=2.0
  palindromic:  a=1e-5,  b=2.2
  ...per form

Coefficients are recalibrated automatically as new work block completions arrive. The model is stored in the calibrations database table.

Drift Detection

The engine compares consecutive WorldSnapshots to detect significant changes that require immediate attention:

Worker change — Workers joining or leaving the fleet
Discovery — New prime found, potentially shifting strategy
Stall — Search making no progress for extended period
Budget alert — Spend rate exceeding budget velocity target

Safety Checks

Before any action is executed, safety gates are evaluated:

Budget gate — Cannot start new searches if remaining budget is below threshold
Concurrency limit — Maximum simultaneous searches per form
Stall penalty — Penalize forms that have recently stalled
Cooldown — Minimum interval between actions to prevent thrashing

Dashboard Integration

The AI engine state is visible in the dashboard at app.darkreach.ai/strategy:

Current scoring weights and form rankings
Decision history with reasoning and outcomes
Cost model curves per form
Drift event timeline

Configuration

The AI engine runs automatically when the coordinator starts with a database connection. Key configuration is via environment variables and the ai_engine_state table:

bash

# Tick interval (default: 30s)
AI_ENGINE_TICK_INTERVAL=30

# Budget limit (USD per day)
AI_ENGINE_DAILY_BUDGET=50.0

# Maximum concurrent searches
AI_ENGINE_MAX_CONCURRENT=8