The engine knows what it knows — and what it doesn't.

Sources used

• SEBI RHP (Red Herring Prospectus) filings — revenue and sales data for listed developers (Signature Global DRHP + RHP FY23)
• HRERA public register — project registrations, launch dates, handover milestones, unit counts
• Multi-portal RERA scrape — declarant-filed estimated project cost, land cost, construction cost. 5 portal sources: haryanarera.gov.in · maharera.maharashtra.gov.in · rera.karnataka.gov.in · uprera.azurewebsites.net · rera.delhi.gov.in. 4,760 raw project scrapes; 714 passed quality gates (cost-per-sqm sanity check, falsifiability band [0.65, 1.55], non-phased records, GFA-gate). Per-source coverage breakdown (RERA-declared agreement): HRERA Bulk Gurugram n=74 (10.4%/9.3%) · HRERA NCR-B n=209 (10.8%/8.6%) · MahaRERA n=204 (11.0%/10.7%) · KRERA Bengaluru n=199 (11.1%/6.6%) · UPRERA NCR n=19 (Tier-5 IDC-noise disclosed) · Delhi RERA n=6 (small sample). TNRERA Chennai n=301 disclosed separately as Tier-5 (23.8%/19.9%). HRERA Tier-3 (n=3) sub-threshold — not published as headline.
• Business Standard, Economic Times — published revenue announcements for cross-reference

What is validated

• Revenue prediction (GDV) — validated on n=15 completed projects with publicly disclosed revenue data. Verified-input headline MAPE = 15.9% on n=5 Signature Global DRHP/RHP projects (Solera, Synera, Serenas, and two additional verified).
• Sellable area — validated on n=14 projects with disclosed saleable sqft. MAPE ≈ 15.9% on the verified cohort; engine under-predicts saleable area by ~13% on the Haryana AHP cohort (known calibration gap — Haryana AHP policy permits a higher saleable-to-FSI conversion than the baseline rule). Patch #2 (2026-05-16) partially corrects this for AHP projects.
• Cost prediction — true-cost accuracy (independent professional QS) — validated against Colliers India 2024 + JLL India construction-cost benchmarks across mid/premium/luxury and 12 cities (n=20 city×grade): MAPE 8.5% mean / 7.3% median, every parameter cited to CPWD + QS. This is the engine's accuracy claim.
  
  Breadth-of-coverage — agreement with RERA self-declared cost — separately back-tested against HRERA + MahaRERA + KRERA + UPRERA + Delhi RERA filings (a lower, self-reported floor typically 20-35% below true cost), published with full transparency: (1) 5-portal coverage cohort: agreement 11.3% mean / 8.5% median (n=714); (2) NCR-only sub-cohort: MAPE 15.1% mean / 10.2% median (n=313, Master's NCR refocus, gate <11% hit by 0.8pp on median, honest cap documented); (3) UPRERA NCR Tier-5 IDC-noise sub-cohort: MAPE 19.5% mean (n=19) — structural category mismatch: UPRERA Form-5 Row 3C declared cost embeds Infrastructure & Development Charges the engine does not model, biased toward UNDER-prediction — disclosed not hidden; (4) TNRERA Chennai Tier-5 disclosed separately: 23.8% mean / 19.9% median (n=301 — district-city loc mapping + 14 cited spec overrides); (5) 6-portal full transparency including TNRERA: 16.2% mean (n=1039, for audit completeness). HRERA Tier-3 (n=3) sub-threshold not published as headline. Plotted-colony Faridabad (n=2) excluded separately — scope mismatch, not accuracy failure.
• Timeline — RERA delay patterns are well-documented. We use RERA-registered handover dates as ground truth. Note: current back-test reads the realised delta from project intake; a predictive timeline test (engine takes plot inputs, outputs a predicted timeline) is queued.

What is not yet validated

• IRR — developers do not publicly disclose project-level IRR. Validation activates at n ≥ 10 full-tier realised-outcome datasets.
• TS-RERA Telangana (login-walled) + remaining UPRERA bulk (CAPTCHA-gated) — T461-TSRERA portal audit confirmed login wall (10,863 projects gated). T461-W5 unblocked TNRERA Chennai via/public-view2/ endpoint (+301 projects, disclosed as Tier-5). Cost MAPE for TS-Hyderabad + full UPRERA + TN beyond Chennai requires RTI route (drafts atback_tests/RTI_*) — current 5-portal coverage cohort already spans HRERA + MahaRERA + KRERA + UPRERA NCR + Delhi RERA.
• Plotted-colony cost model — Faridabad cohort (n=2) documents the scope boundary. A dedicated plotted-dev model is in the H3 backlog.

Full methodology: whitewarp.in/methodology →

Primary accuracy is measured against INDEPENDENT professional quantity-surveyor true cost — Colliers India 2024 (Grade-A 15-floor ₹2,780/sqft) + the JLL India Construction Cost Guide — across mid/premium/luxury spec tiers and 12 cities (n=20 city×grade): the engine predicts hard construction cost within ~8.5% mean / 7.3% median (ratio 1.03), every parameter cited to a primary source (CPWD plinth-area lineage + QS). The 714-project RERA figures below are breadth-of-coverage — agreement with RERA SELF-DECLARED cost, a lower self-reported floor typically 20-35% below true cost — published transparently, but NOT the accuracy claim. Revenue MAPE (15.9%, n=5) is measured on projects whose plot area is independently sourced from SEBI DRHP/RHP filings — Signature Global Solera, Synera, Serenas, plus two additional verified projects. RERA-declared agreement is published with full per-source transparency: (1) 5-portal coverage cohort 11.3% mean / 8.5% median (n=714, HRERA + MahaRERA + KRERA + UPRERA + Delhi RERA); (2) NCR-only sub-cohort 15.1% mean / 10.2% median (n=313 — Master's NCR refocus directive, gate <11% median hit by 0.8pp); (3) UPRERA NCR Tier-5 IDC-noise sub-cohort 19.5% mean / 19.3% median (n=19 — Form-5 Row 3C embeds Infrastructure & Development Charges the engine does not model, disclosed not hidden); (4) TNRERA Chennai Tier-5 disclosed separately 23.8% mean / 19.9% median (n=301 — district-city loc mapping + 14 cited spec overrides); (5) 6-portal full transparency including TNRERA 16.2% mean / 11.8% median (n=1,039, audit completeness). Every input adjustment that produced these numbers 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, +35% FY18-FY24) applied to pre-2021 registrations. Per-developer factors (Sobha, Prestige, LGCL, Brigade, Godrej, Mahaveer, SNN, Sterling, Unishire, ATS) cited from 10-K filings. 6 audit cycles caught and reverted gaming attempts. IRR validation is pending — no listed Indian developer discloses project-level IRR.

Non-NCR geography

The engine accepts inputs from any Indian jurisdiction, but its accuracy claim is currently validated on Gurugram and Sohna only. Bengaluru, Chennai, Hyderabad, Pune, Mumbai, and Thane are in the corpus for area and timeline validation — revenue and cost validation require portal-level access that doesn't exist at scale yet (JS-rendered pages, auth gates, or scanned PDFs). RTI filings are drafted for K-RERA, TS-RERA, TN-RERA, and UP RERA; pending filing.

Faridabad scope mismatch (disclosed, not hidden)

The Faridabad cost cohort (n=2 plotted-colony, MAPE 177.9%) is excluded from the group-housing headline because the engine's cost model does not apply to that project type. We publish this number explicitly because hiding it would be worse — a 177.9% MAPE on two out-of-scope projects tells you something true and useful about where the model's boundary is. An engine that knows what it doesn't know is safer to use than one that smooths over scope boundaries.

n=5 revenue cohort is below production threshold

A MAPE on n=5 projects has a wide confidence interval. We publish it because it's real, independently sourced data — not because it's statistically conclusive. The honest interpretation: the engine has not materially misfired on any of the five verified projects, but five is not enough to rule out a lucky run. We will activate a stronger claim at n ≥ 10.

Cost MAPE is a comparison-side measurement, not a customer-facing output error

HRERA Form REP-I Part C records the developer's declared estimated cost at project registration. This is mostly hard cost + land; it excludes some soft costs that the engine includes in its total project cost output. The MAPE we report is the gap between the engine's hard-cost estimate and HRERA's declared figure — not the gap between the engine's total cost output and final project spend. That comparison (engine vs. actual final cost) requires developer co-operation or access to completed-project accounts, which we don't have publicly.