Stage 1: Intent
Stage 1Stage 2: Evidence
Stage 2Stage 3: Judgment
Stage 3Stage 4: Action
Stage 4Platform Docs / Architecture
Platform Architecture
The complete technical blueprint of the Entrestate Decision Infrastructure — a unified "Pipeline-to-Tunnel" operating model that transforms raw market signals into defensible, institutional-grade outcomes.
The One-System Model
The Entrestate architecture is not a collection of competing modules — it is a single, linear deterministic flow. The 10-Phase Pipeline serves as the factory where raw data is refined. The 5-Layer Evidence Stack acts as the warehouse for high-integrity inventory. The 4-Stage Decision Tunnel functions as the storefront where complexity is hidden to expose defensible outcomes.
The critical architectural link: Phase 7 (Investment Score) directly produces the ranking signal used in Stage 3 (Judgment), while Phase 8 (Intent Tagging) feeds Stage 1 (Intent Parsing). Every phase in the pipeline has exactly one downstream consumer in the tunnel, ensuring a deterministic path from raw sensor ingestion to the final Decision Object.
The 10-Phase Sequential Data Pipeline
The foundational Data Layer that transforms chaotic market signals into a structured inventory of scored UAE projects. Each phase refines data from "Raw Sensors" toward "Actionable Belief."
| Phase | Name | Output | Downstream Consumer |
|---|---|---|---|
| 1 | Source Ingestion | Raw HTML / JSON Objects | Internal Data Layer |
| 2 | Entity Extraction | Project Names, Initial Prices | Stage 2: Evidence (L5 Raw) |
| 3 | Developer Registry | 481 Canonical Developers | Stage 2: Evidence (L1 Verified) |
| 4 | Price Verification | Verified AED Prices | Stage 2: Evidence (L1 Canonical) |
| 5 | Yield Calculation | Rental Yield % (from L4 DLD) | Stage 2: Evidence (L1 Canonical) |
| 6 | Stress Testing | Grade A-F Resilience Scores | Stage 3: Judgment (Circuit Breaker) |
| 7 | Investment Score | Composite Score (0-100) | Stage 3: Judgment (Ranking Signal) |
| 8 | Intent Tagging | outcome_intent[] Array | Stage 1: Intent (Query Routing) |
| 9 | Quality Scoring | quality_score (0-100) | Stage 4: Action (Display Filter) |
| 10 | Evidence Compilation | evidence_sources JSONB | Stage 4: Action (Transparency Footnotes) |
Key Pipeline Mechanics
- Phases 1-5: Transform Raw Sensors into L1 Canonical Verified Truths. Developer Registry (Phase 3) normalizes 481 developers. Price Verification and Yield Calculation establish L1 baseline against L4 DLD historical data.
- Phases 6-7: Stress Testing assigns A-F grades based on market pressure resilience. The Investment Score synthesizes timing, resilience, and yield into the primary 0-100 ranking signal.
- Phases 8-10: Intent Tagging enables query routing. Evidence Compilation packages all sources into JSONB for footnoted transparency.
The 5-Layer Evidence Stack
Our "Sensor vs. Judge" philosophy: external platforms (Property Finder, DLD, RERA) are sensors that detect market movement. The Entrestate Adjudication Engine is the final judge. Data integrity increases as the layer number decreases.
L1
Canonical
(Highest reliability)Audited static truths: normalized developer names, verified AED prices, confirmed handover dates, geospatial coordinates. No ROI calculation occurs without finalized L1 values.
L2
Derived
(High reliability)Calculated truths from L1 data: Investment Scores (0-100), Stress Grades (A-F), and verified rental yield percentages. Powers the objective Market Score component.
L3
Dynamic
(Medium reliability)Living states responding to market pressure: live inventory levels, price momentum, BUY/HOLD timing signals. Adjusts Data Confidence based on market events.
L4
External
(Low-Medium reliability)Raw sensor data from DLD, RERA, Property Finder, and Bayut. Provides context but is never the final judge of truth — awaits internal adjudication to reach L1 status.
L5
Raw
(Lowest reliability)Unprocessed HTML/JSON artifacts, regex snippets, and raw PDF brochures before AI extraction. Entry point for Static Truth Recovery. Never used for direct decisioning.
The 4-Stage Decision Tunnel
The tunnel hides data complexity to expose only defensible outcomes. It moves the system from a "search bar" to a "decision engine," preventing Intent Collapse — the failure to distinguish between users with identical queries but divergent goals.
1
Intent Parsing
The HALTS Mechanism
Natural language is converted into a structured TableSpec JSON via the TableSpec Compiler. If a query is ambiguous, the system HALTS — it refuses to guess and becomes an active interrogator, forcing profile calibration before execution.
Prevents Intent Collapse: 'best property' resolves differently for Conservative vs Speculative profiles.
2
Evidence Collection
Exclusion Policy
The system retrieves records from the canonical graph while applying a strict Exclusion Policy. Distressed sales, internal transfers, and duplicates are filtered to protect L1 price integrity. Only projects passing L1 verification advance.
Data hygiene is the non-negotiable prerequisite for judgment.
3
Judgment
The 65/35 Engine
Properties are ranked via a 65/35 weighting engine: 65% Market Score (objective quality based on timing, yield, stress resilience) and 35% Match Score (subjective alignment with the user's risk profile and time horizon).
The Investment Score (Phase 7) provides the primary ranking signal for this stage.
4
Action
Decision Object Factory
Evaluated data is transformed into Decision Objects — professional artifacts (PDFs, memos, decks) ready for executive presentation. Every number is footnoted to its specific L-layer, scrape date, and algorithm version.
Three-page output: Portfolio Summary, Risk Matrix, Evidence Appendix.
Profile Calibration & Scoring Engine
The scoring engine shifts weights based on the calibrated investor profile. A project's "value" is never an objective constant — it is a variable function of the user's risk/return profile and time horizon. The translation mechanism is fully auditable: Profile → Weights → Query → Artifact → Footnoted PDF.
| Metric | Aggressive | Conservative | Impact |
|---|---|---|---|
| L1 Canonical Price | 0.15 | 0.35 | Entry-point safety/margin |
| L1 Canonical Yield | 0.10 | 0.30 | Cash-on-cash returns |
| L2 Investment Score | 0.35 | 0.15 | Algorithmic upside potential |
| L2 Stress Grade | 0.05 | 0.25 | Circuit breaker for safety |
| L3 Timing Signal | 0.25 | 0.05 | Entry-window heat |
| L3 Price Momentum | 0.10 | -0.10 | Rewards/penalizes market heat |
Conservative Focus
Targets "Safe Yield." Filters for completed projects with verified L1 yields. L2 Stress Grade acts as a mandatory circuit breaker (only A/B). Requires L4 DLD historical occupancy >85%.
Aggressive/Speculative Focus
Prioritizes capital growth with disproportionate weight on L2 Investment Score (0.35) and L3 Timing Signals (0.25). Identifies market leaders before peak pricing. 71% of UAE inventory is eligible.
The Decision Object Factory
Final outputs are rendered as three-page high-trust PDFs with full evidence footnotes:
Portfolio Summary
Top 5 projects ranked by profile_score with key metrics and scoring rationale.
Risk Matrix
Visual cross-reference of stress grades vs development timelines with circuit breaker flags.
Evidence Appendix
Full Evidence Drawer: every number footnoted to its L-layer, scrape date, and algorithm version.
Footnote Examples
- 1 Price: L1 canonical (PF verified, last scraped 2026-03-01)
- 2 Yield: L1 canonical (calculated from L4 DLD rental data)
- 3 Score: L2 derived (investor_score_v1 algorithm v1.0)
AI Documentation Support
Ask the AI assistant to explain any architectural component, compare scoring profiles, or generate investor-ready summaries from this documentation.
Open AI assistantInteractive Architecture Map
Mind Map
Read the system as connected operating lanes instead of a crowded horizontal board.
Deep-Dive Articles
Each node in the architecture map has a dedicated article with scope, execution details, and operational context.
Open all articles