TerriScan: An Incident-Evaluated, Doctrine-Governed Multi-Agent LLM System for Recalculable Urban Indicator Production in the Global South

City-level indicators are difficult to ground across heterogeneous statistical systems in the Global South, where large language model (LLM) agents accelerate multilingual source discovery but risk unsupported values and fabricated execution reports. We present TerriScan, a doctrine-governed multi-agent system built while producing a 142-indicator matrix for ten emerging centralities in eight countries. A versioned charter separates production, model review, deterministic validation and non-delegable human decisions. We evaluate it as a four-month longitudinal design case study with one instrumented four-day period; its lot evidence is stratified, and six lots required substantive interception. During 15–19 July 2026, eight defect classes were registered, each with an identifiable corrective and no recorded intra-class recurrence; exposure denominators were published where countable; and seven earlier qualifying corrections predating the register are reported. The arbiter origin recurred across classes; the exhaustive hash-resolution control remained planned. No unsupported value detected by recorded controls remained in the engraved matrix. The system stopped when work required unrecorded human decisions, but our audit found its absence rule unenforced—181 of 210 absence-state cells named no consulted source. We claim no minimal or universal architecture, but show how incident records, deterministic controls and decision boundaries make urban data production auditable and capable of principled refusal.