Model Judgment Within Deterministic Harness

Description

A cross-cutting design commitment the paper identifies: across all subsystems, Claude Code trusts the model's judgment within a rich deterministic harness rather than constraining its choices. Quantified by the estimated 1.6% decision-logic ratio (98.4% operational harness). Hierarchical permissions preserve safety invariants across agent boundaries, and assembleToolPool() merges built-in and MCP tools into a unified interface, but the model retains full latitude over which tools to invoke. Trade-off: good local decisions can produce poor global outcomes when bounded context prevents global awareness.

Key claims

Relations

Sources

src-20260423-0cff68d3291b