방법론

Two layers of knowledge. The platform keeps (1) authoritative structured data — leases, financial lines, utility bills, vendors, equipment — and (2) the source documents behind them. Analytics and AI answers consult the structured layer first; documents provide verification and context. Uploaded rent rolls and statements are historical sources with dates, not live data.

Provenance engine. Every important fact row carries: value, effective date, document date, reporting period, source document, source kind, a precedence rank, confidence, and — when a person confirms it — who verified it. Values are never overwritten; a better or newer source supersedes the old row and the trail stays visible on the property pages.

Temporal precedence. When sources conflict: signed documents (leases, amendments, executed agreements) outrank rent rolls and summaries; official public records outrank manual entry unless an administrator overrides; newer sources outrank older ones of the same class; human verification is authoritative. Values are also classed as current, historical, projected, contractual, user-entered, or AI-estimated — and the class is displayed.

Computed figures. Occupancy = leased SF ÷ total suite SF, dated today. NOI = effective gross income (revenues minus vacancy and bad debt) minus operating expenses, from booked monthly actuals; trailing-12 windows state their months. Budget variance compares YTD actuals against the adopted budget. Utility anomalies compare each recent billing period against the same month a year earlier (seasonal baseline); rises above the threshold are flagged with both periods shown. Valuation scenarios are direct capitalization (NOI ÷ cap rate) on stated assumptions — analytical aids, never appraisals.

Permissions before AI. Role capabilities and per-property grants are enforced when building AI context: records a user cannot open never reach the model. Organizations are fully isolated from one another. AI conversations are logged per property for audit.

AI transparency. Answers label figures with periods, cite source documents by number, distinguish record data from calculations from estimates, and say what is missing rather than inventing values. Model routing is provider-agnostic: different tasks (chat, extraction, classification, vision) can run on different models, currently Anthropic Claude with slots for other providers.

News. Real articles from named mainstream industry publishers via their public feeds: headline, the publisher's own summary, timestamp, always linking out. Nothing is rewritten; an empty state is shown rather than placeholder content.