A daily crypto, finance, and AI vocabulary puzzle about dollar rails, onchain risk controls, agent execution, and inference efficiency.
Stablecoin Rails
These terms cover backing, issuance, redemption, and how a token stays close to its target value.
- Reserve Assets: Reserve assets are the cash-like or liquid holdings that support a stablecoin's redeemability and target value.
- Redemption Window: A redemption window is the period during which a holder can exchange a token for its underlying value or backing.
- Mint Authorization: Mint authorization is the permission or control gate that determines who may create new stablecoin supply.
- Peg Deviation: Peg deviation is the gap between a token's market price and the value it is designed to track.
Oracle Checks
These terms describe price inputs, safety controls, and forced-close logic used to reduce bad decisions.
- Oracle Price: An oracle price is an external market quote supplied to an onchain protocol or risk engine.
- Price Feed: A price feed continuously delivers market data that a contract, app, or model can consume.
- Circuit Breaker: A circuit breaker pauses or limits activity when prices or system behavior move outside expected bounds.
- Liquidation Threshold: A liquidation threshold is the point where a position may be closed or reduced because collateral is too low.
Agent Runtime
These terms cover how an agent reasons, calls tools, and keeps short-lived context while completing tasks.
PlanningTool CallMemoryScratchpad - Planning: Planning is the process an agent uses to break a task into ordered actions before executing them.
- Tool Call: A tool call is a structured request for an external function, API, or capability during an agent run.
- Memory: Memory is stored context that helps an agent reuse relevant facts across a longer interaction or workflow.
- Scratchpad: A scratchpad is short-lived working context an agent uses while reasoning through a task.
GPU Throughput
These terms describe common ways to make inference faster, denser, and more efficient on shared hardware.
- Batching: Batching combines multiple requests so shared hardware can process them more efficiently.
- KV Cache: A KV cache stores prior attention keys and values so generation can continue faster.
- Quantization: Quantization reduces numerical precision so models use less memory and often run faster.
- Tensor Parallelism: Tensor parallelism divides model computation across multiple devices to increase usable throughput.