Honest methodology. Every number we publish has a source. Every claim has a sample size. Every limitation is stated. Our 11.3% mean / 8.5% median figure (n=714 across 5 RERA portals) measures how closely the engine matches the cost developers self-report to RERA — a compliance figure that typically reflects a partial cost baseline, below independent quantity-surveyor estimates — so it is a coverage metric, not a true-cost accuracy claim; independent QS validation is detailed below. We will show you exactly how we got there, including the audit cycles that made our error rate go UP before it went down.
The engine pulls from 60 live data adapters organised into 6 tiers by reliability and freshness. No single tier is privileged — the pipeline resolves conflicts by weight.
Tier 1 — Statutory registers
CPWD DSR 2025 construction rate schedules · Multi-portal RERA cost ground-truth (4,760 raw scrapes filtered to 714 5-portal HEADLINE per T461-S63: HRERA n=283, MahaRERA n=204, KRERA n=199, UPRERA NCR n=19 Tier-5, Delhi RERA n=6) · TNRERA Chennai n=301 disclosed separately as Tier-5 · RERA registration tables (HRERA, K-RERA, TN-RERA, UPRERA, MahaRERA, Delhi RERA) · NBC 2016 occupancy and coverage rules · RBI dbie.rbi.org.in WPI Group 25+26 for FY18-FY24 inflation correction
Tier 2 — SEBI-filed disclosures
Signature Global DRHP + RHP FY23 — project-level GDV in table cells (n=5 revenue-verified) · Listed developer annual reports for cross-reference (Prestige, Brigade, Sobha, Tata, Godrej, Mahindra)
Tier 3 — Government remote sensing
ISRO Sentinel-2 imagery (cloud-cover, greenery, urban heat island proxy) · NHB Residex price indices (quarterly, 50 cities) · IMD weather normals for climate risk overlays
Satellite flood / heat / soil risk analysis (Climate Shield add-on) · ASI buffer zones · Ridge zone and Yamuna floodplain shapefiles · Road-width estimates from OSM
Tier 6 — Regulatory knowledge base
Jurisdiction rules engine — FAR / FSI limits, setback tables, ground coverage caps, basement norms for 40+ Delhi NCR sub-jurisdictions · Chennai CMDA bylaws · Haryana AHP policy conversion factors
Section 2
The Pipeline — Brief to Verdict
Each feasibility request passes through five deterministic stages before the verdict is locked. The same plot, the same inputs, produces the same verdict every time — tier-invariant by design.
01
Brief intake
Plot address, dimensions, road width, zone classification, land cost, selling rate, spec level, and funding structure. Validated against NBC 2016 minimum plot sizes and jurisdiction-specific rules before proceeding.
02
Khilji core engine
Computes permissible FAR / FSI from the jurisdiction rules engine (40+ NCR sub-jurisdictions, Chennai CMDA). Derives maximum permissible built-up area, sellable area, and unit count. Applies setback and ground-coverage constraints. Produces the GDV (gross development value) estimate.
03
NBC validator
Cross-checks all outputs against National Building Code 2016 occupancy norms, fire-egress minimums, and accessibility requirements. Flags non-compliant inputs before Monte Carlo runs begin.
04
Monte Carlo — 25 000 runs
Samples over uncertainty distributions for construction cost per sqm, selling rate, absorption timeline, and interest rate. Produces a probability-weighted IRR distribution, P10 / P50 / P90 outcomes, and a Verdict (PROCEED / MARGINAL / AVOID) with confidence band. 25 000 runs chosen to stabilise the P10 tail — increasing to 50 000 adds less than 0.3% precision at 4x compute cost.
05
Verdict — locked and tier-invariant
The Verdict is locked: the same inputs always produce the same P50 IRR and the same PROCEED / MARGINAL / AVOID call. Snapshot tests in CI enforce this — any determinism regression blocks the deployment pipeline. The tier (Quick / Full / Expert) changes the depth of explanation and add-ons, never the underlying verdict.
Section 3
Calibration — n=714 Back-test, 5-portal HEADLINE
The engine is back-tested against 714 Indian residential apartment projects across 5 RERA portals (HRERA + MahaRERA + KRERA + UPRERA + Delhi RERA). 4,760 raw scrapes filtered through quality gates. T461-S63 evidence-first calibration, two-run reproducible 2026-05-21.
Revenue MAPE
15.9%
n=5, verified-input (SEBI RHP)
RERA-declared agreement, mean
11.3%
n=714 · agreement w/ self-declared RERA cost (a partial compliance baseline, below independent QS estimates), NOT true-cost accuracy
RERA-declared agreement, median
8.5%
n=714 · agreement w/ self-declared RERA cost, NOT true-cost accuracy
Projects in corpus
714
5 RERA portals, 6 cities
Revenue MAPE is measured on SEBI-filed projects where project-level GDV appears in a table cell — not press releases, not acreage estimates. Cost agreement with RERA self-declared cost (a compliance figure that typically reflects a partial cost baseline, below independent QS estimates — a coverage metric, not true-cost accuracy) is published as a 5-tier cohort (T461-S63, evidence-first calibration 2026-05-21, two-run reproducible): 5-portal HEADLINE 11.3% mean / 8.5% median (n=714 — HRERA + MahaRERA + KRERA + UPRERA + Delhi RERA); NCR-only sub-cohort 15.1% / 10.2% (n=313 — Master's NCR refocus, gate <11% median hit by 0.8pp); UPRERA NCR Tier-5 IDC-noise sub-cohort 19.5% / 19.3% (n=19 — Form-5 Row 3C embeds Infrastructure & Development Charges the engine does not model, disclosed not hidden); TNRERA Chennai Tier-5 disclosed separately 23.8% / 19.9% (n=301); 6-portal full transparency including TNRERA 16.2% / 11.8% (n=1,039). Every input adjustment cites a Tier-1 primary URL + page + quoted text in the underlying JSON — 350+ HRERA + 222+ KRERA + 30+ MahaRERA + 23 UPRERA + 6 Delhi cited brochure spec overrides. WPI inflation correction (RBI dbie.rbi.org.in Group 25+26) applied to pre-2021 registrations. Faridabad's two plotted-colony projects are excluded — scope mismatch, not accuracy failure.
The engine's agreement with RERA self-declared cost (11.3% mean / 8.5% median on n=714 — a self-reported floor, not true-cost accuracy) is measured against projects in the back-tested geographies listed below. For cities outside this set the engine still produces an estimate, but the figure is indicative — accuracy is approximately ±35% until back-test data is collected and the model is re-calibrated. The report will show an explicit “uncalibrated location” banner when this applies so customers are never misled.
KRERA corpus n=199. BBMP + BDA jurisdiction rules. Calibration run T461-S63, two-run reproducible.
Mumbai Metropolitan RegionFull
MahaRERA corpus n=204. MCGM + MMRDA + SRA zones included. Calibrated at project level; sub-municipal FSI variations are approximate.
Chennai (Tamil Nadu cities)Partial
TNRERA corpus n=301, disclosed separately as Tier-5. CMDA bylaws engine implemented. MAPE 23.8% mean / 19.9% median — higher than NCR/Bengaluru due to TNRERA cost-reporting structure differences. Figures for Chennai carry an explicit partial-calibration notice.
All other citiesIndicative only
No RERA back-test data yet. The engine applies national CPWD DSR rates with a location factor but cannot claim calibrated accuracy. Reports for uncalibrated cities show a prominent ±35% uncertainty banner. Data collection via RTI pipeline is ongoing.
Specification level is required for a precise figure
Construction cost varies by approximately 2.2× between Basic and Luxury spec on the same plot. When a customer does not specify a level, the engine assumes Mid-spec as a default and returns a cost range (low–high) instead of a single point estimate. The report headline and any totals will reflect this range with an explicit banner. The four tiers are: Basic (economy/EWS-LIG finishes — CPWD basic schedule), Mid (standard RCC + ceramic tile + standard fittings — most common residential), Premium (premium marble/granite + branded fittings + false ceilings), and Luxury (imported stone + high-end MEP + smart-home integration). Rate references: CPWD DSR 2025 Schedule A (Basic/Mid), Schedule B (Premium), Market Addendum (Luxury).
Section 4
What Is Locked and Tested
✓
Tier-invariance test
Quick, Full, Expert, and Climate tiers must produce identical P50 IRR and identical PROCEED / MARGINAL / AVOID verdicts for any given set of inputs. A parametric test suite enforces this on every CI run. Tiers change depth of output, never the verdict.
✓
Snapshot tests — determinism
Reference inputs run on every PR. The P10 / P50 / P90 IRR outputs, sellable area, and GDV must match within 0.01%. Any regression blocks the deployment pipeline.
✓
BoQ allowance — Rs 9 500 / m²
The Bill of Quantities line-item allowance for structural cost is locked at Rs 9,500 / m² until the CPWD DSR 2024 schedule is published. Unlocks automatically on data refresh.
✓
Verdict canonical — 25 000 Monte Carlo runs
Run count is fixed. Increasing it to 50 000 improves P10 tail precision by less than 0.3% — not enough to change a verdict. Locking it ensures consistent performance across all tiers.
Section 5
What Is Not Yet Implemented
Honesty about scope is how you earn trust. All outputs are schematic-grade — they give architects and developers a defensible starting position, not a stamped drawing or a legal opinion.
—
NRI tax and TDS cascades
Non-resident income tax, TDS on property, FEMA remittance rules — all out of current engine scope. Refer to a tax professional.
—
MEP detailed equipment sizing
Module M8 gives system-level MEP (HVAC load estimation, plumbing riser sizing). It does not produce equipment-level BOQ or stamped drawings.
—
Structural design (RCC calculations)
The engine outputs structural sizing ratios and indicative slab and column proportions. These are not stamped structural drawings — a licensed structural engineer is required before construction.
—
Cost MAPE outside Haryana
HRERA is the only Indian RERA portal that exposes project-level cost in static, machine-readable HTML at scale. MahaRERA requires bearer-token auth; K-RERA, TN-RERA, TS-RERA, UP RERA require RTI or manual downloads. Until those land, cost accuracy claims are not geographically generalisable.
—
Plotted-colony and commercial cost models
The current engine assumes GFA times construction rate — a residential apartment model. Commercial / mixed-use / plotted-colony projects are formally tracked as a separate Tier-2 cohort (n=8, MAPE 65.3%) — structural category mismatch because HRERA total cost for these projects embeds land + infrastructure the engine does not model. Plotted-colony cost model (HRERA REP-I structure differs) and a dedicated commercial cost-driver model are both in the H3 backlog.
—
H3 modules — scaffolded, not implemented
Advanced MEP equipment schedules, structural frame optimisation, plotted-colony cost model, and cross-jurisdictional cost MAPE are all in the H3 development backlog. Scaffold exists; implementation is not yet shipped.
Section 6
Updates and Versioning
Engine version — each deployed engine build carries a commit hash. Reports include the engine version at generation time so any accuracy changes are attributable.
Cohort version — the back-test cohort is versioned separately (currently v8, 2026-05-21 snapshot, n=714 5-portal HEADLINE; T461-S63 evidence-first calibration, two-run reproducible). Methodology claims reference the cohort version, not just the engine version.
Data freshness — Tier 1 data (HRERA, RERA registers) is scraped quarterly. Tier 3 (NHB Residex, IMD) is updated on each quarterly release. Tier 4 commodity rates (MCX and NCDEX) are updated monthly.
What triggers a methodology revision — a change in the cohort composition, a new tier of data source, or a material change in the pipeline. Phrasing and layout changes to this page do not constitute a methodology revision.
Section 7
Questions About the Methodology
If you have questions about how a number was derived, a data source you want to verify, or a scope boundary you want clarified — write to us. We answer methodology questions directly, not through a helpdesk template.