Methodology → Accuracy Metrics

Eval Suite: 25 Golden-Set Cases

This page documents 25 curated internal fixtures covering different property types, inputs, geographies, and condition levels. They are regression checks for expected product behavior, not completed-project accuracy measurements.

Internal fixture coverage, not contractor-bid accuracy

Published Fixture Inventory

Curated internal fixtures

25

Synthetic and hand-authored cases used to catch scope, flag, confidence, and range regressions

Represented states

4

Texas, Arizona, Colorado, Illinois

Named markets / localities

7

Houston, Plano (Dallas-Fort Worth), Austin, Conroe, Scottsdale, Aurora (Denver metro), Chicago

Range sanity rule

≤ 2.5×

low < mid < high, with high no more than 2.5× low

Geographic Coverage

StateMarkets coveredCases
TexasHouston, Austin, Plano (Dallas-Fort Worth), Conroe22
ArizonaScottsdale1
ColoradoAurora (Denver metro)1
IllinoisChicago1

Evidence is deepest for Houston (RSMeans CCI 0.97). National coverage uses RSMeans city cost indexes. Local contractor pricing may vary ±20–30% — always verify with a local contractor quote.

Full Golden-Set Case List

#CaseConfidence
01Old Houston house — foundation + pre-war systemsMedium
02HVAC end-of-life — single dominant systemHigh
03Roof leak + water stainingMedium-High
04Outdated electrical panel — code upgradeHigh
05Plumbing and sewer issueMedium
06Cosmetic-only rehab — no major systemsVery High
07Vague description — limited input qualityLow
08Lender-grade request — full detailMedium-High
09No ARV supplied — internal analysisMedium-High
10Uploaded inspection PDF findingsMedium
11DFW suburban ranch — HVAC + roofMedium-High
12Austin TX urban infill — full gutLow-Medium
13Severe mold + moisture intrusionLow
14Full gut rehab — all systems + finishesLow
15Condo — HOA owns exterior + roofHigh
16Phoenix AZ — HVAC-critical desert climateMedium-High
17Duplex — shared systems, two unitsMedium
18Partial fire damage — kitchen to atticLow
19Rural property — well and septicMedium
20Denver CO — freeze-thaw foundation + hailMedium
21Turnkey ready — fresh flip, minimal scopeVery High
22Chicago IL pre-war two-flat — knob-and-tubeLow-Medium
23Renegotiation — option period activeMedium
24Lender-grade detail — draw schedule formatMedium-High
25Minimal vague input — address + year onlyVery Low

What This Validates — and What It Doesn't

What these evals validate

  • Scope recall: known finding types produce the expected scope categories
  • Flag accuracy: risk flags appear when triggered, don't appear when not triggered
  • Confidence calibration: stated confidence matches actual accuracy band
  • Cost range sanity: ranges are internally consistent and proportionate to scope
  • Geographic routing: different markets use the correct cost index
  • Edge case handling: vague inputs, no ARV, condos, rural, fire damage all produce reasonable output

What these evals don't cover

  • Real contractor bid comparison — no live deal data is included in the eval suite
  • Post-renovation actual cost variance — actuals data is collected through the in-product actuals tracking feature
  • Properties outside the training data geography — accuracy may degrade in markets with no RSMeans data
  • Hidden defects not mentioned in the input — Scopebase can only scope what it can see in the description or PDF
  • Hyperlocal micro-market pricing — CCI indexes are metro-level, not neighborhood-level

Help improve calibration

Scopebase is a pre-inspection screening tool, not a final contractor bid. If you complete a renovation, you can submit final invoices through Actuals tracking. Those comparisons help identify where future calibration work is needed.

Submit actuals from your estimate detail page, or contact support@scopebase.org with your estimate ID and final contractor invoices.

Accuracy Metrics | Scopebase